Home

EXPLORATORYSTUDY - The Stanford University InfoLab

1. d 4 3 oo 4 og n i ve a amp Vr os e i Wee i e Ww p e e a D 1 4 t D s s Bo a a ee ee E E E en we 5 e t o ve y qu erty mae mee ca pam me poet P POT i a a e x D 5 gt gt v 4 7 4 4 4 4 1 i EA ae P e o roo o b e gt A ce i Q r 1 e om 1 1 s x o 1 qucm 2 72 5 D Q p bene ow ttr t pas lt a rame e menea o m gt 7 a e gt ome i un x ee n Tw bono x Vu eso A 55 toe amp e 26 D oe soe 5 LI E LI L ki gt seer gt fe E M
2. 0 25 degrees Likewise the screw can wobble about the tip of the driver ATgs rot x a erot y B aM where 5 degrees a 5 degrees 5 degrees s B s 5 degrees We are interested in producing a parameterized estimate for AT the relation between the center of the hole and the tip of the screw In this case the system finds only one acyclic path of relations linking the hole and tip Thr latip ls AandsT Tp sbox where Location of hole with respect to box Thd s Location of driver with respect to hand Tas Location of screw with respect to driver Location of tip with respect to screw box Location of box in work station hand Location of hand in work station In this case the nominal values for these quantities are given by IV 63 trans nilrotn vector 3 85 3 20 4 90 Distances in cm Tha niltrans trans nilrotn vector 0 0 25 4 trans nilrotn vector 0 0 3 18 box trans nilrotn vector 45 7 101 6 0 hand 2 trans rot y 180 deg vector 49 6 104 8 30 3 errors other than those described above are assumed to be negligible Using this information application of the algorithm gives us a parameterized form for AT ATgt where Apa Yevector 3 20 3 85 0 0 28 6 0 Pysvector 28 6 0 0
3. LI 2 a d 1 z Ms i o LI ow ist Uu 5 n are 5 e wt n o a 9 as PPM gt e EJ 6 P v s B i Figure 3 Handle and Spindle of Pencil Sharpener V 5 The whole assembly task was broken up in broad terms as follows a get and position handle crank b get and position base body get and insert spindle shaft through hole in base d screw spindle into hole in crank e get and position shell of sharpener against base placing it over spindle f seat shell and turn it 45 degrees into place g bring arm back to initial position FIXTURES A fixture is a holding device which supports the workpiece in a fixed orientation with respect to the tool in this case the manipulator hand Each fixture has locators to position the workpiece and clamps to hold it rigidly A free rigid body has three degrees of freedom of rotation and three degress of freedom of translation Locators restrict these six degrees of freedom in order to give points of reference As shown in Figure 4 the workpiece would lose three degrees of freedom when placed and maintained on the locators lettered A locators lettered B restrict another two degree
4. time p 0 960 sec 45 491 sec 90 468 sec 135 582 sec 225 1 21 sec 2709 1 54 3159 1 67 sec 90 degrees will therefore be chosen The exact values of dtry and standoff are less important The principal constraint is that they be large enough to guarantee that location errors in the hole or pin will not cause the motion to stop prematurely or to knock the pin into the object while approaching the approach point Currently arbitrary values dtr df 1 inch standoff 1 inch are used Thus for this case our destination approach and target locations will be respectively IV 37 pin holestrans rot zhat 90 deg vector 0 0 2 54 pin Aolestrans rot zhat 90sdeg vector 0 0 4 25 When values for dtry and standoff have been picked they are combined with the grasping strategy to form an embryo pin in hole strategy The expected time to execute it 1 Just the time expected for the pickup operation plus the time for moving to the hole 4 5 Accuracy Refinements In the absence of errors the strategies derived in the previous section would suffice to accomplish the task Unfortunately the world 15 not so kind and we must consider the effects of errors For each strategy we apply the machinery given in my dissertation 28 and illustrated in the appendix to estimate the error between the pin and hole at the approach point as a function of free variables APhp 2
5. Stanford Artificial Intelligence Laboratory August 1976 Meme AIM 285 Computer Science Department Report No STAN CS 76 568 STANFORD ARTIFICIAL INTELLIGENCE LABORATORY 3 Computer Science Department E Stanford University Stanford California 94305 PROGRESS REPORT 3 Covering The Period December 1 1975 to July 31 1976 EXPLORATORY STUDY OF COMPUTER INTEGRATED ASSEMBLY SYSTEMS by T O Binford D D Grossman C R Liu R C Bolles R A Finkel M S Mu jtaba M D Roderick B E Shimano R H Taylor R H Goldman J P Jarvis V D Scheinman T A Gafford Prepared for NATIONAL SCIENCE FOUNDATION WASHINGTON D C 20550 cO ABSTRACT The Computer Integrated Assembly Systems project is concerned with developing the software technology of programmable assembly devices including computer controlled manipulators and vision systems A complete hardware system has been implemented that includes manipulators with tactile sensors and TV cameras tools fixtures and auxiliary devices a dedicated minicomputer and a time shared large computer equipped with graphic display terminals An advanced software system called AL has been developed that can be used to program assembly applications Research currently uhderway includes refinement of AL development of improved languages and interactive programming techniques for assembly and vision extension of computer vision to areas which are currently infeasible
6. d distance of nearest hole or edge in s from xy comment 400 if x y is outside of s if ddbest then b e g i n dbestedjxbestexsybesteyjend end if dbest xa then done end The tapping place is then computed from xbest and ybest as pin holestrans nilrotn R t svector xbest ybest O p A where Tsh strans R H Psh position of hole with respect to s The results of a typical application of this method is shown below Here we are looking for a tapping place near one of the corner holes of our box located at 3 85 8 20 with respect the top surface of the box In this case we assume that the box location is known precisely so that the only xy error comes from the hand Thus max 0 3 inches Ax Ay s 762 243 226 cm 762 On the first iteration through our outer Y loop IV 41 450 762 cm 1 21 Going through our inner loop produces x Y 0 5 06 320 619 5 4 83 391 397 4 22 435 558 3 47 435 559 2 87 39 111 2 64 3 20 577 2 87 249 115 3 48 2 05 229 Thus xbest 2 64 e 3 20 and dbest 57 on this iteration Since this value of dbest is considerably larger than our possible confusion radius 243 cm we have found an acceptable tapping place and can stop looking The corresponding tapping point 15 trans nilrotn vector 21 00 2 0 Once such point has been found then At 15 re evaluated taking account of th
7. 1 15 FPU 4 85 02 127 ESTIMATED TIME 6 14 PICKUP 72827 W 315 dog Grasp Anglo 180 deg Grasp Distance s 3 54 OPEN BHRNO TO 2 88 MOVE BHANIP TO PINISTRRNS ROTN VECTOR 924 383 001 188 DEG VECTOR 6 0 3 542 VIR PINIsTRANS ROTN VECTOR 924 383 001 180 0 VECTOR 6 0 8 62 CENTER ON OPENING lt 185 00 RBORT GRRSP FAILED END AFFIX PIN BMANIP MOVEPINI TO PINI amp TRANS NILROTN VECTOR 0 0 6 78 FIRST ATTEMPT MOVE PIN TO BHieTRANS ROTN 2 90 8 amp 0EG VECTOR 0 0 4 25 VIA BH1eTRANS ROTN 2 90 G DEG VECTOR 8 8 2 54 ON FORCE ORIENT PIN1 2HAT gt 8 02 00 STOP ON ARRIVAL 00 GRORT CEXPECTED amp FORCE HERE DISTANCE_OFF ZHAT INV BHI amp TRRNS ROTN ZHAT 98 8050 VECTOR 8 9 1 71 sDISPL PINI IF RBS DISTRNCE OFF gt 168 THEN BEGIN PINIMISSEO BOOLEAN FLAG SEARCH LOOP RERL R OW W X YsFLAG FALSE 572 0 7540ROK 1 21 The program has been edited very slightly to improve readability by removing excess blanks and by rounding all numbers to three significant digits For instance the computer output had 0 00 106 instead of 0 00 1 IV 50 WHILE NOT FLAG R 1 72 00 BEGIN 8 C 872 R WHILE NOT FLAG AND H 259 0EG 00 BEGIN RBS XeRsCOS W lt 1 58 RBSCY RsSIN GD 1 15 THEN BEGIN FRAME SETPNT SETPNTe BHI
8. 760 649 800 lt 4 888 XHAT PV YHRTeR amp 5 854 VECTOR 768 649 000 5 888 YHRT PV XHRTaR amp 5 854 VECTOR 768 649 888 4 880 XHRT PV XHRTsRe 5 85 VECTOR 760 649 000 4 688 XHRT PV 888 Z2HAT PV WHERE R NILROTNeROTN ZHRT 2 ESTIMATE LIST _ amp amp 2 888 TO 888 379 381 et 801 TO 881 Ws 92 632eDEG 92 683 DEG cos 49 046 SIN HO 999 COS OW 1 888 R 888 2 Screw on Driver This example illustrates use of the differential approximation methods to estimate runtime errors The task 1s insertion of a screw into a hole of our favorite box The box is assumed to sit on the table with possible displacement errors in the xy plane and rotation error about the z axis Abox uy erot z Y QUI rS vm w here 0 inches lt s 0 3 inches 0 2 inches sys 0 2 inches 5 degrees Y s 5 degrees n F IV 62 The screw is held on the end of a driver as shown in Figure A 2 and the driver is held in the hand We assume that errors in the driver s position with respect to the hand are negligible However the hand s position will only be assumed accurate to within 0 05 inch in displacement and 0 25 degree in orientation Akand transi vector 6 5 5 erot x erot y by erot tb where 0 05 inches s by 8 0 05 inches 0 25 degrees
9. outline looks something like this begin pin in hole Declarations and initial affixments Grasp the pin Extract amp transport over hole Insert Let go of the pin end 2 3 2 Declarations and Affixments The declarations include frame variables for the pin hole and other points of interest In addition we must write affixment statements describing how the various frames are linked Strategy decisions are embodied in these declarations For instance we need to declare a frame in grasp for use in the grasping operation It seems natural to affix this frame to pin But where If there is any chance that the pin can bind in the rack or box hole then 6 As present capabilities are extended we will probably want to include other facilities like collision avoidance which are too expensive to be done at runtime and so require pre planning IV 6 it will probably be a good 14 to twist the pin during extraction and or insertion operations To do this effectively we must grasp the pin so that its axis lines up with the wrist axis Alternatively grasping the pin at an angle may be better for reasons of collision avoidance or may allow us to produce a more efficient program by reducing arm motion times In this case we ve decided to twist the pin so that the on grasping position must be used Usually one doesn t sit down and write all the declarations before writing any code This
10. For example if the current mode is PCODE then CONVERT SCODE SHOW PCOOE PADDRESS 132024 is equivalent to SHOW SCODE PADORESS 132024 ALAID will respond with the source language representation of the code at location 132024 This example also demonstrates the use of embedding which allows one ALAID query to be embedded in another one First the innermost query 48 serviced and the result of that query is treated as the argument to the next The conversion facility is used by the PDP translate modes that it does not understand 2 SUPERVISION COMPILE lt logical name gt lt source gt LOAD lt logical name gt START logical gt DUMP logical gt 11 37 GET logical name gt INITIALIZE These commands are intended to place ALAID as the sole supervisory program over the entire AL system Each program module can be given a dogical name by the user for example SAMPLE LNAME mode applies The source might be a file on the PDP IO FILE mode or a literal statement SCODE mode The result of compilation is a file with name SA this file can be loaded by the command LOAD LNAME SAMPLE Before the loading can take place the AL runtime environment must be available on the PDP 11 If it is not then the INITIALIZE command will provide that environment This command also can be used to flush any old A L programs that might still be cluttering the execution system The LOAD command w
11. NILROTN VECTOR 0 0 6 79 MUST TAP HOVE PIN1 BHIATRANS NILROTN VECTOR 1 21 882 5 68 VI A BHIeTRANS NILROTN VECTOR 1 21 802 08 ON FORCE ORIENT PINI e2HRT gt 8 02 DO STOP ON ARRIVRL DO ABORTC EXPECTEO A FORCE CORR e 2HRT INV BHle TRANS NI LROTN VECTOR 1 21 802 sDISPL PINI FIRST RTTEMPT MOVE PIN1 TO BHJeTRANS ROTN ZHAT 80 Q4DEG VECTOR 9 0 6 25 VIR BHieTRANS ROTN ZHAT 98 2 0 VECTOR 8 8 2 542 DN FORCE 1 2 gt 802 00 STOP RRRIVRL DO RBORT EXPECTED A FORCE HERE DISTRNCE OFF ZHAT INV ROTN ZHRT 80 8 DEG VECTOR 8 8 1 71 sDISPL PIN1 CORR RBS DISTANCE OFF gt 239 THEN SSED UNEXPECTEDLY LET GO OPEN BHAND TO 2 98 UNFIX PINA FROM BHANIP NOU GET HAND CLEAR HOVE BMANIP TO BMANIPsTRANS NILROTN VECTOR 6 8 5 68 In this case the error footprint in the plane of the hole is small enough so that no search is needed On the other hand the uncertainty along the hole axis is quite large Consequently the system has chosen a tapping place at trans nilrotn vector 2 002 0 with respect to the hole which is then used to locate the top surface of the box more precisely This is accomplished by moving the pin along a path starting two inches above the nominal height of the surface and ending two inches be
12. Reiser 751 which is used for debugging SAIL VanLehn73 programs is capable of displaying the exact text that is being executed by maintaining cross references into the source file and takes commands that are a 5 subset of standard SAIL commands especially procedure invocations In this case the disassembly process is made possible by keeping pointers to the source code not by examination of the object code BAIL contains a primitive compiler in that it can parse some constructs of SAIL for the purpose of patching code Other source language debuggers also exist COPILOT Swinehart 74 includes a sophisticated example None of these can display macros in their expanded form Style of debugging varies from person to person common technique 15 to proceed the program until a fatal error occurs like a memory reference trap By examining the failing instruction it is often possible to deduce what data are wrong these data are then examined If they are indeed wrong an attempt is made to localize the bug by installing a breakpoint at some place where it is expected that the data are still right The program is then restarted or backed up to a safe place and allowed to proceed When the breakpoint is encountered the suspect data are examined If they already are in error an earlier breakpoint 15 installed and the process is repeated If they are still good single stepping is employed to see where they go wrong One
13. SHOW VALUE SNAME brrm SHOW VALUE PNAME 32 EVALUATE expression The expression which is given in the source language SCODE mode is evaluated in the current context and the value is returned With this command ALAID has the full power of the source language to investigate data structures 3 MODIFICATION SET VALUE variable name gt expression The variable name is given as for SHOW VALUE The expression can be an expression variable in SNAME or PNAME modes or a source language expression in SCODE mode The facility for executing source language statements to be discussed in detail below Can also De USEC to set val ues EXECUTE SCODE variable lt expression gt Section 3 Control Structures 1 EXAMINATION SHOW CODE lt address gt The address can be in SADDRESS or PADDRESS format the code that is displayed will be in SCODE or in PCODE depending on the current typeout mode Thus the SCODE corresponding to a PADDRESS can be displayed If the PCODE is in the middle of a single SCODE statement the SCODE displayed will be annotated in progress 11 34 2 MODIFICATION SET CODE address code The given code in SCODE or PCODE mode is placed at the given address in SADDRESS or PADDRESS mode There is no space problem if both the address and the code are in P modes other combinations cause difficulties SADDRESS and PCODE is usually foolish it replaces the entire code for the statement with a si
14. movements which tended to cause unpredictable results For instance the WAIT after dropping the handle ensured that the drop was not affected by the hand closing before the handle had a chance to drop in place The CENTER and CLOSE operations are equivalent to the GRASP of the human but take much longer to do OPEN for the manipulator is equivalent to RELEASE performed by the human operator except that it is not quite as gentle and the hand usually opens quite suddenly The POSITION obtained generally by GO in the case of the mechanical arm is almost the same as GOTO MOVE and could have been considered the same and tabulated together The GOTO MOVE was movement to an approach point while the POSITION mainly GO was a one part directed move to the location of the grasping position a smooth move tended to cause a collision with the object being grasped even when the direction of approach was well defined since the arm did not successfully null out errors in all the six joints at the end of the allotted motion time a one part downward directed move required only the movement of three joints joint 2 to lower the arm joint 3 to extend the boom so that the hand would move down vertically and joint 5 to keep the hand approach vector vertical 15 ASSEMBLYDATA NUMERICAL BREAKDOWN OF ASSEMBLY ELEMENTS FOR ASSEMBLY BY HUMAN GIRASB P ROVE POSI RELE TURN TION ASE APPLY PRESSURE 2 PUT HANDLE PU
15. 1s necessary to tap the object surface to get a better determination of the pin hole relation before trying the insertion 4 What search pattern if any to use in error recovery 1 25 J The overall approach 1 fairly direct First a number of preliminary calculations are performed based on the task specification and initial planning model to obtain initial position and accuracy estimates and to determine basic tolerances Then the system generates possible ways to grasp the pin subject to geometric feasibility constraints For each distinct grasping strategy a best approach symmetry for the pin relative to the hole is computed using expected motion time as an objective function The grasp approach pairs are sorted by goodness and then are reconsidered in best first order to see what additional refinements are required based on the estimated pin to hole determination If the error along the hole axis is too large then a tapping place is found as near the hole as is safely possible Similarly if the errors inthe plane of the hole are too great then a decision to search is made The expected time required for tapping and search are calculated and added to the cost The process is continued until an optimal strategy can be chosen Once the decisions have been made it 1s fairly straightforward to generate the corresponding AL code sequences which are quite stylized Subsequent sections will describe each o
16. T 20 which indicates that 9 is equal to 180 degrees or T lt 0 in which case equation 17 is indeterminate and the following equation can be derived after simplification of the polynomial used to develop eq uation 17 0 tan 18 Once the tangent of the half angle is determined the sine and cosine of 9 can be computed from the following trigonometric identities 1 tan d x 19 tan z l tan Next re writing quations 15 and 16 we obtain the following expressions for the sine and cosine of 38 c0 c T 5 5389 j Substituting expressions 17 19 for the sine and cosine of 0 we obtain 13 8 50 250 c eT 20 Since we are limited to working with the extension of the prismatic joint S4 between 6 and 85 inches the ratio of above equations will be determinate and independent of S4 Hence we can co mpute 0 using an arc tangent function and quations 20 The extension of joint three can now be found by evaluating the following equation which was derived by simultaneously solving equations 14 and 15 for S4 11 4 S 5 21 5 Since mechanical stop limits prevent 0 from being either 0 180 degrees the equation for S is never indeterminate Having found values for 6 0 and we can now solve the remaining nine transform element equations for values
17. The link that has been described in such detail is fundamental to making information available to the various parts of ALAID In particular symbol tables residing on the PDP 10 provide correspondence between symbolic entities the source code and physical entities in the object code A symbol table can be pictured as a memory of bindings kept so that the decisions made in binding can be quickly simulated without being explicitly recomputed In this way all the information that goes into the decisions can be discarded the result itself 1s distilled for future reference Three examples of symbol tables will be discussed here variables processes and code Each of these symbol tables must work in both directions the PDP 11 representation must be resolved into PDP 10 representation and vice versa 1 VARIABLES The runtime representation of variables is based on their lexical level within the source program and the order of declaration within that level new lexical level 15 started for each COBECIN and PROCEDURE although procedures are not fully implemented a slightly different design might have counted each BEGIN the lexical level count Thus each variable is identified although not uniquely by level and offset Typical values for the level 0 1 and 2 and typical offsets are even octal numbers from 10 to 100 In the runtime system these two quantities are combined into one 16 bit word with the level in the left 8 bits and
18. all our candidate motion strategies have been generated we set about refining them in best first order To do this we generate the error terms and compare them against the requirements established at the very beginning For the rtrategy just shown we get AZ 180 1 50 cm 1 15 cm 60 degrees The value of Azis thus small enough so that we are sure not to be confused about whether the pin will make it into the hole Thus we don t have to tap On the other hand the error footprint is bigger than Arg so we will have to search The estimated extra time for this is 1 8 seconds giving us a total estimated cost of 7 27 seconds The refinement of strategies continues until we reach PHL SPEC 134823 PRELIMS NULL RECORD PICKUP PICKUP SPEC 163875 PRELIMINARIES NULL RECORD RPPRORCH OPENING 2 98 APPROACH BMRNIPaPINI amp TRRNS ROTN VECTOR 688 608 510 126 DEG VECTOR 5 0 4 42 GRASP OPENING 185 GRASP BMRNIPePINI amp TRANS ROTN VECTOR 608 698 618 126 DEG VECTOR 9 9 3 54 GRASP OETERM DEPARTURE POINT PINIsPINIsTRANS NILROTN VECTOR 0 0 6 79 GOODNESS 4 88 REMARKS 68 8 dogGrasp Rng lo 188 drg Grasp Oi ttancr 3 54 APPROACH PINIsBHIsTRRNS ROTN ZHRT 90 s0EG VECTOR 0 0 2 54 DESTINATION PINIsBHIeTRRNS ZHRT 90 G amp DEG VECTOR 6 0 1 71 TARGET PIN1sBHIsTRRNS ROTN ZNAT 90 8 0 VECTOR 8 0 4 28 XPORT TIME
19. amp TRRNS NILROTN ROT ZHRT 1 85 sVECTOR X Y 690 MOVE PI N1 TO SETPNT amp TRANS ROTN 2 98 0xDEG VECTOR 8 9 4 25 VIR SETPNT TRRNS ROTN ZHRT 90 00 6 VECTOR 8 8 2 54 ON FORCE ORIENT PIN1 amp 2HRT gt 802 00 STOP ON ARRIVAL 00 RBORT EXPECTED FORCE HERE DISTRNCE OFFeZHRT INV SETPNTa TRANS ROTN 99 8 0EG VECTOR 9 0 1 71 xDISPL PIN1 IF RBS DISTRNCE OFF lt 169 THEN FLRG TRUE END END RoR 572 END IF NOT FLAG THEN RBORT PINIMISSED END LET 60 OPEN BHANO TO 2 98 UNFI X 1 FROM NOU GET HAND CLEAR HOVE BMRNIP TO BHANIPsTRANS NILROTN VECTOR 8 8 S 88 The search loop used here works by generating x y offsets ever widening circles about the origin Each point generated is tested to see if it 1s within the footprint limits x Ax Ay sys If so then a displacement vector in the coordinate system of the hole is computed by 0 and used to produce an offest candidate location setpnt for the hole location If the insertion attempt for this point succeeds then flag 15 set to indicate success and the loop is terminated If the attempt fails or if x y was outside the error footprint then the next point is tried The loop continues to be executed until either the entire expected error range has been exhausted or the insertion succee
20. breakpoints and then allow the computation to continue High level languages like FORTRAN and ALGOL no longer maintain a unity of program and data and debugging techniques either use post mortem dumps and traces of procedure calls and variable assignments or they force the user to debug the generated machine code using a DDT like debugger Some student oriented languages SNOBOL and ALGOL W come to mind provide the facility of optional post mortem dumps that wind back through the stack of procedure calls and print out the values of all variables for each procedural level These languages also allow the tracing of procedures so that an examination of the program output will reveal the sequence in which control entered and exited procedures In short the debugging facilities allow examination of the path of control and the values of variables See Satterthwaite 75 for a discussion of the debugging facilities of ALGOL W and a very good summary of the history of debugging Interactive high level languages like LISP and SAIL increase the user s control over his program He can interrupt it at will and restart it This alone 15 a fantastic advantage over batch systems Many are the times that there 1s no apparent failure of the program but the 11 4 programmer suddenly remembers a mistake and he can stop the program to fix 1t without wasting unnecessary computer time to complete a worthless computation The idea of a program trace
21. cO cO cO 3 5 cO 50 cO 0 58 4 80 c0 30 5 S c0 s0 c0 8 0 9 0 0 50 5 5 Ty 501 0 80500 cO cO 506 060 50 6 9 s0 s050 0 0 0 50 0 0 0 50 0 0 50 7 To 50050 0 c s0 50 0 50 8 Ty 30 0 c0 S 39 5 0 50 50 5 c0 c0 50 S S 9 59 39 c0 30 00 c0 50 c0 10 50 50 c0 50 0 0 0 9 50 0 11 50 50 50 0 9 12 39 50 5056 cO cO S 0 5 5 13 Of these 12 equations only six are independent the three equations representing position 5 9 13 and any three of the remaining nine equations which specify orientation Furthermore no unique solution exists for the above set of equations For the geometric configuration of the Stanford Arm there are always at least eight sets of joint variables which satisfy the equations but due to physical stop limits only two of these eight can ever be attained 2 Solution for 0 8 5 To solve for the joint variables we begin by taking advantage of the fact that the axes of the last three joints of the Stanford Arm intersect forming a spherical joint This simplified geometry allows us to reduce the problem from one of simultaneously solving for all six degrees of freedom to two separate three degree of freedom p
22. compliance to the motion close blue with force xhat s0 force yhat 0 An alternative is to use the center statement which makes the motion compliant to the touch sensors on the finger pads 2 3 4 Initial Program Once the pin is grasped a single motion statement can perform the extraction transport and insertion operations After the pin is in the hole we can let go of it and move the arm back out of the way again being sure not to hit the pin with the fingers while moving off These operations and the revised grasp code gives us the following program begin pin in hole Declarations and initial affixments Grasp the pin open bfingers to 0sinches move blue to Pin grasp via pin grasp approach cen ter bf ingers affix pin to blue Extract transport and insert move pin to in hole position via pin withdraw Aele approach Let go of the pin open bfingers to 0 unfix pin from blue move blue to bpark via pin grasp approach 7 For this reason Paul calls z the approach axis of the manipulator We will adopt this usage occasionally also 1V 8 The value of pin grasp approach in the final move statement will have been updated as a consequence of its indirect affixment to pin If we had chosen to affix pin grasp to pin holder rather than to pin this updating would not occur and the motion specified would be rather wild In this case we could always invent a new
23. considering what refinements would be necessary 4 7 Code Generation Once we have selected a strategy the actual synthesis of program text is accomplished by calling procedures that extract the appropriate values from the strategy record substitute them into the appropriate slots in code skeletons and print the results Pickup Strategies The procedure for writing pickup strategies ldoks something like this procedure write pickup pointer pickup strategy pkp begin print PICKUP phe remarks pkp erlf 9 BHAND approach opening pkp i y write_motion_sequence approach_point pk p grasp_point pk p null s print CENTER BMANIP crif 5 Carriage Return Line Feed IV 45 print ON OPENING lt grasp_opening pkp DO crlf print ABORT GRASP FAILED crlf print AFFIX Jocation variable object pk5 TO write motion sequence departure point pk null end Pin in Hole Strategies The write pin in hole procedure is slightly more elaborate than write pickup which it uses as a subroutine In addition to generating more output write pin in hole must make several decisions about what code to emit 1 Is tapping to be performed 2 Is a search to be made Actually these decisions have already been made and are reflected in the data structures Thus our code writer looks at the tapping place field of the strategy recor
24. for the torque directed in both directions and the corresponding force vector is given by 0 0 0 0 0 2T strain gage readings for the six independent force vectors have been taken the procedure discussed in the previous section can be used to compute the calibration matrix for the force sensing wrist 4 esolving Forces and M oments at an rbitrary Point It is often necessary to resolve the strain gage readings into forces and moments that act at a III 11 point other than the position used for the calibration For example we might wish to monitor the forces at the finger tips to enable us to stop contact or we might be interested in the interaction between a tool we are holding and a work piece In either of these cases the problem is to determine the applied force vector at a point located at a distance d d d from the calibration point For a simple linear translation the new force vector Fy nF yF gM y MyM is related to the force vector at the point of calibration FF FA My MM by the following matrix equation Fyi 8 8 8 Fy Eo 8 1 8 8 0 0 Fy 8 8 8 0 8 F z 7 My 0 07 dy 8 0 My 37 My 97 8 dy 8 1 8 My We now call the 6x6 matrix on the nght D Then in order to directly resolve the strain gage readings into an equivalent force and moment at a point located at in the calibration coordinate system we combine equations 37 and 33 to obtain F Dx C xe Fina
25. from blue in Aole positione pin Update our model move blue to bpatk via fin grasp approach end 2 3 9 Further Discussion Cost of Error Recovery An important consideration in writing error recovery code such as the loop above is that it is not always cheap The amount of programming involved can frequently rival that required for the main part as indeed is the case here If a useful purpose is served this cost is generally unimportant aside from Procrustean considerations More important is the extra time rquired in execution The extra computer time spent head scratching isn t likely to be an issue The time spent in manipulator motion is another matter For instance each iteration through the loop may take nearly as long as does the initial attempt In an assembly line this kind of delay can get very expensive although some provision for buffering between stations can help to smooth things somewhat Fortunately some forms of error recovery impose almost no additional manipulation cost The principal example is the use of previous measurements to correct future behavior For instance suppose we are screws into all the holes in the box As each screw 15 inserted its location can be noted and used to update the value of bex Since the remaining hole locations are updated implicitly the likelihood of having to search decreases with each screw Vision 1 especially important in this regard since
26. geometric modeling of objects and constraints assembly simulation control algorithms and adaptive methods of calibration ar A EE Ie TRU DR TABLE OF CONTENTS INTRODUCTION I Overview AL SYSTEM AND ASSEMBLY Il ALAID An Interactive Debugger for AL 11 Improvements in the AL Run Time System IV Generating AL Programs from High Level Task Descriptions v Study of Assembly of a Pencil Sharpener VISION AND MODELING VI Mathematical Tools for Verification Vision VII Discrete Control of the Arm POINTY User Manual IX Monte Carlo Simulation of Toierancing MEA E SA ne PE EE ea OVERVIEW Thomas 0 Binford Artificial Intelligence Laboratory Computer Science Department Stanford University The author is a Research Associate in the Computer Science Department and is a co principal investigator on the Computer Integrated Assembly Systems Project 1 1 INTRODUCTION This report is the third in a sequence of reports summarizing research progress in Computer Integrated Assembly Systems This project supported by the National Science Found ation is concerned with the software technology of programmable automation including computer controlled manipulators and vision systems The basic goal is the simplification of assembly and visual programming Prior to the period covered in this current report a complete ha
27. has been done here largely for convenience of exposition frame pin pin grasp pin grasp approach frame pin holder pin withdraw f tame hole in hole position hole approach frame box affix pin withdraw to pin holder at trans rot zhat 30sdeg vector 0 0 4 cm pin holder frame nilrotn vector 5 inches O inches 0 affix pin grasp to pin at trans rot xhat J80s deg vector 0 0 2 cm affix pin grasp approach to pin grasp at trans nilrotn vector 0 0 3scm pin pin holder affix h hole position to hole rigidly at trans nilrotn vector 0 0 Jecm affix hole approach to hole at trans nilrotn vector 0 0 I cm affix hole to box at trans nilrotn vector 5scm 4 amp cm 2 cm y box initial box position There may be several choices of what affixments to make as well as where to make them For instance we have affixed pin grasp to pin One consequence is that if pin should be rotated the position of the hand with respect to the tool rack when the pin is grasped will similarly be rotated The rotation won t make much difference in this case since pin is assigned an explicit value and since the arm configuration won t be much changed by rotations of pin anyhow other circumstances arm solution or collision avoidance considerations may make it desirable to affix pin grasp to pin holder instead 2 3 8 Grasping the Pin To grasp the it is necessary to open the fingers an appropriate amount move the hand
28. has been strengthened to allow the user to do his examination whenever an interesting place is reached the features developed in DDT for machine language programs are incorporated into LISP debuggers LISP completes the circle it is a high level language that unifies program and data so that a program can be self aware it has become natural to write LISP debuggers in LISP itself Once it has become clear why a program is failing it 1s often useful to patch that error and let the program proceed In this way the next failure can be found and the patch can be tested An interactive debugger allows not only examination but also modification of the contents of the program In language like LISP where there is no distinction between programs and data structures the facility to modify data structures 15 extremely powerful because it implies a power to modify the program itself This power 15 reflected in the debugging packages found in most LISP implementations See for example Teitelman 74 Flow of control is also a plaything in the hands of interactive debuggers If some bad code is to be avoided the program can be made to jump over it Special code to be executed for restorative purposes initialization routines can be executed directly from the debugger and then control returned to the ailing program most algebraic languages the input syntax is compiled into a machine language representation which is then execu
29. history of programming system development is the progressive transfer to the computer of coding responsibilities The nature of coding is largely clerical One keeps track of particular facts and applies them in a stylized manner As elements of programming practice become better understood or at least better formalized this process has been extended into areas of increasing abstraction Thus symbolic assemblers feature the ability to keep track of addresses maintain a literal table etc providing a substantial improvement over octal or push the switches programming Similarly algebraic compilers perform many functions of an assembly language coder They keep track of information like assignment of variables to registers where temporary results are stored etc and follow highly stylized though sometimes extensive rules to generate programs that are equivalent their input specifications There are several important points concerning such automatic coding systems First they use their understanding of the formal semantics of the source language and of their own decisions e g to keep maxel in register 1 to keep track of those facts that are appropriate to its task as an assembly language coder Second optimizing compilers produce more efficient output code than non optimizing compilers can but they must keep track of more information to do it Third there are limits imposed by what can be slated explicitly in the sour
30. programs independently of debugging strategies Another suggestion that leads still farther into the LISP like realm is to make the debugger homoiconic with the language that is let all debugging commands be available as statements in the language Then extend the debugger so that any statement in the language can be given to it to execute The same argument holds at the level of the pseudo code not only should statements of the source language be directly executable from the debugger but pseudo operations should also be accessible These operations are useful for performing only a part of a full fledged statement For example to cause a new variable to exist it is most convenient to execute the pseudo code that creates variables The problem of unusual data types 1s related to the level of detail problem Not only does AL have algebraic types scalar vector transform it also has control types events expressions condition monitors force feedback variables motion tables each of which has a peculiar representation of its own How is a motion table to be displayed so that the user can see its destination initial point and what clauses have been associated to it This problem requires a disassembler with some rather special knowledge of not only the code but also the data structures involved in the runtime implementation The strange format used for such things as force feedback variables condition monitors and motion tables can a
31. programs look very much like ALGOL programs The language is block oriented and variants of the usual ALGOL structures are used for program control Since the programs must be executed in a real time environment where several things can be happening at once additional control structures for concurrency and synchronization are required The necessary capabilities are supplied by the well known cobegin coend and event signal and writ primitives Data Types One of the key attributes of a formal language for manipulator control is the use of named variables to describe positions forces and other relevant data Using only th data types of ALGOL would make programs hard to read and would increase the chance of bugs AL avoids these difficulties by providing data types and arithmetic operations for the physical and geometric entities rquired for describing manipulation The most important of these special types are frames which are used to represent coordinate systems and transes which tell how frames are related AL programs use frames to describe hand positions and object locations the set of frame variables and their associated values thus constitute a major part of a program s execution time model of the world Affixment In manipulation tasks it is common to have several frames associated with the same ob ject with each frame playing an important role When the object is moved the frames all assume new values AL provides two di
32. real objects 1s subject to design choices on the part of the programmer may only discover during debugging that his model is incomplete in a crucial way that some important feature of an object has been omitted from consideration Design of appropriate error recovery routines depends greatly on what errors are encountered It is a waste of effort to design the program to carefully detect and remedy an error like dropping a workpiece if in fact the arm never or rarely commits this error in practice It 15 even more frustrating to fail to foresee the possibility that a screw hole is mispositioned if that error turns out to be frequent Experience and sharpened intuition can slowly train a programmer in what sorts of errors to expect but the learning process is full of necessary and clumsy iterations through the debugging loop Another feature some would no doubt consider it a bug of the real world is that many actions are irreversible In algebraic languages most actions have inverse actions and if it 1s Ur2 important to be able to back up care can taken to preserve either the state of the computation or a history of the actions that have been taken so that the code can be retried But the moving finger dips and having smashed moves on Nor all our history lists nor glue shall lure it back to fix but half the wreck nor all our work shall make it look like new Even non destructive actions like putting a pin in a hole ca
33. shows that it is necessary to identify processes in order to properly access variables Each process is given a unique name by the compiler An obvious extension is to allow the user to assign process names himself this will allow greater ease in setting the context during a debugging session A simple pairing of user given names and compiler generated names suffices The actual implementation might use hash coding although the total number of processes 15 likely to be small enough so that even linear tables are adequately efficient 3 CODE The organization of code on the two machines is quite different The basic unit in the source language 15 the statement which 1s compiled into stream of pseudo operations There four different naming techniques that can be partially interconverted source code source labels pseudo code and addresses on the runtime machine Let us restrict our attention to the problem of determining the current statement the source language given the runtime address If the entire source program is kept in the PDP IO memory this is the case for the current implementation then markers can be emitted along with the code that refer back to statement pames that the compiler understands In this way the pseudo code can be made self descriptive and to find the source code from an address on the PDP 11 one need only go back in the pseudo code until a marker is encountered The price for this meth
34. slots to fill in are imit al opening minimum opening object_grasp and object_pickup point As we will see in Chapter 4 these may be computed from the situational and object modelling information 3 7 Concluding Remarks This chapter has discussed the role of planning information in programming and has described the particular kinds of information that are needed for manipulator programming A key point here is that the burden of maintaining this information cannot be escaped If manipulator programs are to be generated automatically then the planning information must represented in a form that the computer can use to make reasonable decisions A full discussion of the representation methods used by the automatic coding procedures described in Chapter 4 may be found in my dissertation 28 To make this report self contained a short summary of the most important technical results is given below The AL Planning Model The AL compiler itself performs a number of coding functions such as planning trajectories and rewriting motion statements not ordinarily found in algorithmic languages These functions require that the compiler keep a better model of situational information especially the expected value of frame variables and affixments than might otherwise be the case The compiler associates a data base of assertional forms with each control point in the program graph and uses simple simulation rules to propagate facts Th
35. that causes failure then reproducing the failure at will until the source of the error has been tracked down In summary what we might call classical interactive debugging has these interlocking features _ 1 Symbols are used extensively 2 Both examination and modification are possible 3 These constructs are available for examination and modification program labels program code flow of control values of variables data structures IL6 4 Backing up or restarting and patching typical operations The realm of examination includes such abilities as searching for code or data having particular form displaying instructions and data setting breakpoints or traces on code or variable reference so that the flow of control and history of variables may be traced and single stepping to execute only one instruction or procedure call at a time Modification involves depositing replacement code or data zeroing large blocks of data interrupting execution restarting execution at arbitrary places inserting new labels or moving old ones although no debugger is yet capable of changing all old references to a label to correspond to the new meaning with the possible exception of SNOBOL and directly executing instructions from the debugger especially procedure invocations 11 7 CHAPTER 3 INTERACTIVE ARM CODE DEBUGGING The fact that AL 15 a real time language for control of real world devices in an environment of
36. the source language The unification of the various stages of AL compilation and execution also provides a groundwork on which to base saving and restoring contexts from one stage of debugging to another and by the same token allows backing up to a previous state in any production run Real world problems that mitigate against reversability are not solved by but internal structures can be reset and then queried so that the user has some assistance in repairing the state of the world to what is expected 11 40 BIBLIOGRAPHY Beeler 76 M D Beeler XNET Cross net Debugger for TENEX User s M anual Updated by R S Tomlinson BBN report to be published May 1976 Binford 75 T 0 Binford D D Grossman Miyamoto Finkel B E Shimano H Taylor R C Bolles M D Roderick M S Mujtaba T A Cafford Exploratory Study of Computer Integrated Assembly Systems Prepared for the National Science Foundation Stanford Artificial Intelligence Laboratory Progress Report covering September 1974 to November 1975 Mader 74 Eric Mader N etwork D ebugging Protocol Request For Comments 643 Bolt Beranek and Newman division 52 50 Moulton St Cambridge Mass 02138 July 1974 Reiser 75 John Reiser BAIL A Debugger for SAIL Stanford Artificial Intelligence Project Memo 270 Stanford Computer Science Report STAN CS 75 523 October 1975 Satterthwaite 75 Edwin Hallowell Satterthwaite
37. the computations can be done in the background in parallel with necessary motions For example suppose there is some chance that the pin may be misaligned in the fingers If a picture is taken when the pin is removed from the rack the actual pin fingers relation can be computed during the time that the pin is being transported to the hole This correction can then be used to get the insertion right the first time 13 Tf the program won t fit into the runtime space available to it then it is necessary to decide what to cut out In many cases the answer may be to get a larger machine Computers are already cheap compared to other components in a manipulator system and are getting cheaper by a factor of ten every five years This suggests that manipulator systems should be designed for easy expansion since the marginal cost of going to a whole new system is considerably greater than expanding pre existing one 14 An exception is if something really hairy is contemplated For instance several systems to do problem solving to figure out how to correct errors have been proposed See 12 Sproull 25 has investigated the question of when runtime planning is cost effective 15 Bolles 7 is currently investigating techniques for accomplishing exactly this kind of task Although his system isn t quite up to the real time requirements described here his results indicate that the task could be performed with essentially the present hardwar
38. the current context and a set of modes with which various data are printed The context 15 a thread of execution possibly containing other threads within it as subprocesses Contexts are used to disambiguate the meaning of variable names and to select processes for interaction Typeout modes dictate the format in which variables and code are displayed Commands that affect execution like halting and jumping commands influence all active processes in the current context therefore one should be careful to distinguish contexts and threads This section will discuss the way modes are set the appropriate typeout modes will be discussed in the sections in which they arise Once a mode has been set it is permanent until reset Most versions of DDT have both permanent and temporary modes RAID associates a mode for every one of the twenty or so variables that can be concurrently displayed Every time a query is answered in some mode the name of the mode prefaces the result The purpose of this is to make all output self identifying so that the result of one query can be used as the input to the next For example the result of a query for the current locus of control might be PADORESS 132024 or it might be SCODE HOVE BARM TO BPARK VIA BPI 1 TYPEOVT MODES Contexts can be displayed by the code that starts up the thread of execution mode That code can be named by location or by contents Locations in t
39. the potting material was that it chipped very easily and each time the fixture was used a little bit of it wore out or broke off resulting in more uncertainty of the locations This disadvantage could be overcome by using other potting compounds instead Another possible improvement in fixtures would be to design them to allow functional inspection By putting sensors in the fixture at the appropriate places it would be possible to tell if the object has fallen at the right position An interesting but unanswered question is whether or not a universal fixture can be designed ASSEMBLY OF THE SHARPENER The handle was located with respect to the edge at the hole end It was placed in the fixture and pushed until it touched the side of the fixture The base fixture had parts removed to ensure that the bottom or the back end did not bind against the base when the base was lifted In fact without those parts removed the fixture came up with the base when the latter was lifted The shell fixture did not need to have any parts removed since it held the shell securely at the bottom Problems were encountered with the original main fixture which the assembly was done When the handle was being positioned in the fixture either the hole end or the roller end touched the fixture and then the orientation of the handle was lost As a result the handle ended up at unpredictable places making correct positioning impossible To correct thi
40. the same as those given in Section 4 5 1 In addition we assume that the rotation errors are as previously stated and that the object in which the hole is drilled is sub ject to small displacement errors in x and y This gives us the following expression for pin hole displacement errors A pA s vevector 3 20 3 85 0 nyevector 2 5 2 5 0 nyavector 2 5 2 5 0 n evector 0 0 0 5 evectar 707 707 0 707 707 0 szhat xhat eysyhat where ny ty and 9 represent rotation errors in the hand and 6 represent displacement errors in the hand represents rotation error in the object containing the hole _ our familiar box and and 6 represent object displacement errors The corresponding constraint equations are 1 00 000 000 969 883 998 32 fu 1 00 000 900 888 888 000 1 00 000 000 000 1 V GI 889 1 88 88 8998 900 gt 0080 900 U 1000 009 883 808 V1 26 127 127 888 989 1 08 809 4 368 008 J VI 900 000 9988 1 88 888 000 VIS 4360 2 t 009 888 008 1 8 900 098 2 V1 2 4360 2 t 869 888 I 000 008 1 08 9 888 V lt 4360 2 008 888 e000 000 1 00 000 7 Vi2 6369 2 888 000 000 1 000 9 000 1 00 Vi 4360 2 008 888 00
41. the six independent force vectors equation 35 must apply so we can write for 1 1 2 6 1 2 8 These 48 equations can be re written as the following matrix equations amp 111 f12 fis 114 5 fig eni 2j 21 122 123 124 fas fog cnj2 33 fa 132 133 134 135 136 573 36 4j fal 142 143 144 145 fag end 151 152 153 54 55 fog 5 6 761 52 163 fea fes fee cn ig This formula represents a system of eight matrix equations each of which relates the 111 9 response of an individual force sensing element to a series of force vectors For each of the matrix equations we have read the six values of the specified strain gage and so long as the six force vectors are independent the 6x6 matrix in equation 36 will be non singular Therefore by applying a standard routine that solves sets of linear equations we can solve equation 36 for the values of the elements of one row of the inverse calibration matrix cnjk 1 to 6 By repeating this procedure for each of the eight matrix equations all of the elements of the pseudo inverse calibration matrix can be determined Equation 34 can then be used to compute the forward calibration matrix It should be noted that this basic method of calculating the pseudo inverse calibration matrix would still be applicable if one wanted to utilize the information from more than six force vectors If there are n samples taken n 6 the matrices in equation 3
42. these special processes have unusual states that the debugger should be able to influence For example a condition monitor has these states inactive active but waiting for the next checking time busy checking executing the conclusion and uncreated its block is not being executed The debugger should be able to put on the monitor so that it does not leave its present state and then be able to force it into some other state in this way it 1s possible to test out the conclusion of a condition monitor without having to actually create the condition that normally triggers it 11 16 CHAPTER 4 DESIGN OF A DEBUGGER W hat s in a name That which we call a rose By any other name would smell as sweet Shakespeare Romeo and Juliet II ii 43 This chapter describes some of the design of a debugger for AL Following the pattern of RAID AID and BAIL this debugger is termed AL AID although other suggestions Cerf personal communication include TRYAL ADDALD DEBAL and even ALDEBERAN AL DEBugging Execution Arm Environment ALAID is an attempt to meet the special needs of an arm code debugger Its driving principles are these 1 The link between the two computers allows a partitioning of planning and runtime information 2 Debugging should proceed equally well from either machine they should be symmetric as far as possible 3 Debugging should be possible without the link insofar as the necessary information is avai
43. to pin grasp and close the fingers The corresponding AL code is open bfingers to 0sinches The 1 0sINCHES is sort of arbitrary move blue to pin grasp close bf ingers affix pin to blue The pin will move if the hand does The most serious difficulty with this code is that the manipulator may collide with something on the way to pin grasp Since the AL compiler does not do collision avoidance 7 we must tend to this detail for ourselves by specifying enough intermediate points so that we stay out of trouble What points are required depends on where the manipulator is before starting the motion which we haven t specified and on what other objects are in the workspace For the moment we assume that the manipulator is clear of any extraneous obstructions and consider only the possibility that the fingers might collide with the pin while moving to pin grasp This may be avoided by moving through an intermediate point pin grasp approach affixed to pin grasp in such a way that the final part of the motion will take place along the wrist axis of the hand Note that this affixment structure guarantees that the fingers will stay out of the way of the pin even if we change the relation of pin grasp to pin Another difficulty is that the execution time value for pin holder may be inaccurate If the rack is bolted to the table the CLOSE statement may overstrain the manipulator This problem can be avoided by adding some
44. to processes that have a special scheduling priority Even less explicit 1s the use of processes to implement joint servos and force feedback variables Parallelism in AL is therefore only partially a result of the explicit nature of the language any implications such simultaneity may have for the debugger are equally valid for any language supporting concurrency like SAIL or concurrent PASCAL Other parallelism derives from the real time aspect of AL condition monitors are intended specifically for rapid response to real time feedback Still other parallelism is due to implementation decisions taken in the coding of AL Both servoing and force monitoring make use of the process structure available in the runtime environment Each joint is separately treated by a software servo to read forces on the arm it is necessary occasionally to recompute some configuration dependent information Force calculation is readily scheduled as an infrequent concurrent process while servos are frequently executed i During a debugging session one would like to attack problems in one thread of execution without interfering with other threads This consideration 1s especially important if one of the threads 1 causing an arm to move and a bug has been discovered in a different and independent piece of code It 1s not a good idea to throttle the machine by debugging at high machine priority while waiting for interactive input unlike 1 I DDT which assumes con
45. tool or carried parts and other parts to a state of contact Le device touches something makes contact RELEASE Device causes tool to release its grasp on part or other tool RETURN Device returns tool to storage area ACCOMMODATION Device allows the forces between parts to modify the motion of parts according to one of the following subprimitives V 13 COMPLEX ACCOMMODATE Accommodation executed during a complex motion having no convenient name to describe motion INSERT Push shaft into hole DEPRESS Deflect a part in its compliant direction WAVE MOTION PRIMITIVES The following 15 a list of WAVE motion primitives used in programming the arm PARK Generates a trajectory to the PARK at rest position GOTO Generates a three part trajectory consisting of the departure center and approach segments One part direct move without liftoff setdown MOVE Same as GOTO except that a smooth trajectory is fitted through the three segments CHANGE Generates a trajectory for differential motion PLACE This causes the hand to move down until it meets some resistance OPEN Opens the hand CLOSE Closes the hand CENTER Closes the hand centrally over the object to be grasped WAVE ASSEMBLY PRIMITIVES The following primitives similar to motion primitives were used to describe the assembly using the mechanical arm CENTER ana CLOSE Results in the hand grasping the object GOTO MOVE Thr
46. variable and affix it to hole Alternatively we could compute the withdrawal point directly as in move blue to bpark bluestrans nilrotn vector 0 0 acm This works because the values of all points in the destination list are computed before the motion is begun If we have a number of such motions it may be convenient to invent a frame and affix it to the manipulator frame withdraw 3 affix withdraw 3 to blue at trans nilrotn vector 0 0 3scm move blue to bpark via withdraw_3 2 3 5 Critique of Initial Program The program we have just written is complete in the sense that it describes a sequence of operations that should transfer the pin to the box hole Whether it will work reliably enough is another question Certainly any easy things that we can do to make the code more robust ought to be given careful consideration We have already built one important form of flexibility into the program by using variables rather than constants to describe locations This has several advantages The code is easier to understand since an identifier like pin holder is generally more informative than an expression like frame nilrotn vector I5sinches JOsinches 0 Modification of programs to accommodate changes in part locations is much easier since the values only appear explicitly once These advantages could also have been derived from the use of compile time variables or macros for symbolic definition of c
47. wrist automatically 1 without the intervention of manipulator programmers For our measurements we will define HW to be the weight of the hand DCM to be the distance from the center of mass of the hand to the center of the force wrist and DH to be the distance from the center of the hand fingers to the center of the force wrist We will resolve all forces and moments at the geometric center of the force wrist 1 The first force vector to be applied is 0 02 0 0 0 To obtain the corresponding strain gages readings the arm is first moved to a position with the hand pointing directly down and a series of readings are taken Then the arm 15 re positioned so that the hand 15 pointing directly up and 111 10 second set of readings are taken The difference between the two sets of readings are saved as our lj 2 Next the wrist of the arm is rotated until the hand is horizontal and the X axis of the force balance is vertical The readings taken in this position and with the hand rotated 180 degrees about its central axis correspond to the force vector 2H W 0 0 0 2H W DCM 0 3 The third set of readings are to be taken in exactly the same manner as the second set except that the hand 15 first rotated about its central axis 90 degrees to align the Y axis of the force sensor with the vertical The force vector associated with this set of readings is 0 2HW 0 2HW DCM 0 0 4 In order to obtain two more independent rea
48. 0 000 000 1 88 V1 2 4360 2 1 88 000 600 V23 762 t 1 88 808 990 1 V2 2 762 809 1 00 008 1 V2 68 t 008 1 88 0888 V2 2 508 088 099 1 88 5 873 1 e800 009 1908 0 V2 2H 8730 5 where IV 43 V1 5 Ny Ny n J Computing amp for six values of gives us ki 0 1 05 cm 30 1 43 cm 60 1 50 cm 90 1 24 120 1 16 cm 150 1 15 cm Consequently Ax 1 50 cm 1 15 cm and f 60 degrees If Ar is less than Arg then we won t have to worry about searching since the pin will always be within the allowable error radius of the hole If not then a search will have to be planned The search loop used is shown in Section 4 8 If a search is requited the cost of the strategy must be adjusted to account for the time spent doing it This is difficult since we don t know anything about the distributions of the errors A worst case estimate can of course be obtain amp by multiplying the time to make one try by the total number of points in the search pattern However this seems too pessimistic Therefore we only count those points within Ax 2 and Ay 2 of the hole 4 6 Selecting a Strategy We wish to select the strategy with the smallest execution time The most direct way to do this is to plan all strategies out fully evaluate them and then take the cheapest T
49. 0 0 0 a vector 0 3 18 0 Bavector 3 18 0 0 Ax ARpt 21 YM 4M Sub ject to constraints 1 16 806 888 1 lt 762 1 18 000 000 V1 2 762 C 688 1 88 8000 1 Vis 508 200 1 88 000 1 V12 508 aee 008 1 88 1 V1 lt 873e 1 198 888 1 00 Vi2 873e 1 1160 eee 00008 888 880 5 127 6 0600 000 gt 888 008 888 V2 gt 127 lig 14 88 000 000 000 800 J V2 lt 127 008 1 88 000 000 888 080 J V22 127 0089 e00 1 00 e00 800 088 1 5 127 299 000 1 00 808 000 688 V2 2 127 690 000 888 1 88 000 888 l V2 lt 436 2 098 000 00 888 008 V2 2 436e 2 l age 808 0080 1888 1 00 000 V2 lt 436e 2 009 688 000 000 1 88 000 V2 gt 436 2 209 000 000 000 1 00 3 V2 436e 2 000 000 000 000 000 1 680 V22 4360 2 1 00 888 7 V3 5 873e 1 l 1 08 000 e V3 2 8730 1 0 00 1 00 7 S 873 1 000 1 88 e V3 2 873 1 where _ IV 64 Vie Y V2 t 1 81 We interested in finding the maximum displacement errors in the plane of the hole Ax and Ay and along t
50. 11 DDT but the debugger is fairly useless for debugging manipulator programs especially if the person using AL 15 not an expert on the implementation The problem 15 mostly one of level H DDT is a low level debugger and AL is a high level language The intent of this report which is taken from the third chapter of my thesis is to examine the problems of preparing correct manipulator code and to suggest the design of a user interface that assists the programmer in fulfilling his function This interface will have some of the flavor of a debugger and some of the flavor of an operating system Although it is based specifically on the implementation of AL the design of the interface is of interest for more general reasons it provides added insight into control structures for operating mechanical devices under programmed computer control and it proposes a uniform debugging and preparation technique that might find use in any large and complex programming environment For this reason this report may be read independently of the remainder of my thesis The discussion has two distinct flavors On the one hand philosophical issues dealing with debugging in general and arm code debugging in particular are treated in a rather abstract fashion On the other hand details of implementation are mentioned in an attempt to demonstrate how needed facilities can be obtained This second type of discussion deals both with a preliminary test Implementation cur
51. 2 56 GOOONESS 6 64 TAP NULLJ ECORO FINE TIME 609 PH 02 008 PH FP 890 PH FP DY 888 PH FP nor ee8 This strategy will take at least 6 64 seconds to execute and all the rest will take even longer However at this point the best completely refined strategy is 1 49 SPEC 138523 PREL IMS NULL RECORD PICKUP PICKUP SPEC 72827 PRELIMINARIES NULL RECORD RPPRORCH OPENING 2 98 RPPRORCH BMANIPsPINIe TRANS ROTN VECTOR 924 383 081 188 DEG VECTOR 8 8 8 62 GRASP OPENING 185 GRASP BMANIP2PIN1s TRANS ROTN VECTOR 924 383 881 189 sDEG VECTOR 6 8 3 54 GRRSP DETERM NILTRANS DEPARTURE POINT PINisPINI TRANS NILROTN VECTOR 0 0 6 79 GOODNESS 4 16 REMARKS 315 dog Grasp Anglo s 188 deg Grasp Distance 3 54 RPPRORCH PIN1sBHIsTRRNS ROTN 2 90 9 0 6 VECTOR 0 8 2 54 DESTINATION PINI sBH1 TRRNS ROTN eHRT 98 888 DEG VECTOR 888 808 1 71 TRRGET PINIsBHISTRRNS ROTN ZHRT 98 8206 VECTOR 8 8 4 25 XPORT TIME 1 38 GOOONESS 6 14 NULL RECORD FINE TIME 608 02 127 FP OX 1 58 PH FP OY 1 15 PH FP ROT 1 85 Since we already have a refined strategy better than any of the remaining unrefined strategies we can stop looking and write the AL code for our current best strategy In this case the computer generated the following program text PIN IN HOLE STRATEGY 138523 DROK w 762 FRPX 1 58
52. 6 can be replaced by n x m matrices and an approximate solution for the rows of the inverse calibration matrix can be found by the method of least squares 3 Calibration Procedure for the Stanford Arm As we are no longer restricted to applying pure forces and moments we will be able to calibrate the force wrist while it is mounted on the Stanford Arm In fact we can utilize the positional and rotational degrees of freedom of the manipulator together with the weight of its hand to aid 1n the calibration procedure Since the force wrist is mounted between the last active rotary joint and the hand of the manipulator the known weight of the hand will be used as a standard of reference against which all other weights will be compared While the weight of the hand is not the best of all possible references due to its light weight compared to the maximum load that the force wrist can measure it does have the advantage of being constant and ever present Also since the weight of the hand must be subtracted from whatever readings we take with the force wrist using it as a reference will reduce the absolute error when small forces are to be measured Furthermore since calibration measurements only require the reading of static forces many readings can be taken for each force vector and digital filtering can be applied to increase their precision We now present the outline of a simple program which can be used to calibrate the force sensing
53. 600 computer in which a set of peripheral processors each has access to the main memory and can cause interrupts in the main processor although there is little communication in the reverse direction and the several peripheral processors cannot pass information among themselves The concept we wish to develop in the context of debugging 1s slightly different Two machines with overlapping but not identical capabilities cooperate to solve problems where the problems may originate on either machine and the solution may be found on 11 13 either machine This concept is a generalization of that found in the XNET debugger Beeler 76 which allows programs in a PDP I 1 to be debugged across a network XNET works with a skeletal debugger in the PDP 11 capable of handling a small set of examination and deposit operations not in a symbolic fashion and a sophisticated symbolic debugger in a larger remote machine Thi two machines are linked by a communications protocol described in Mader 74 The coricept mentioned above generalizes XNET type interactions by allowing the debugging to proceed under direction from either machine with symmetric question asking potentials in the two computers Linking two machines has many delightful properties beyond the ability to compile and recompile Firstly it allows the debugging to take place from either machine Information can be distributed in such a way that the information necessary for the response to s
54. 8 SECONDS Complex accommodate V 18 DISCUSSION COMPARISON OF MOTION PRIMITIVES A comparison of different motion primitives reveals several interesting features More motions are required by the mechanical arm than the human arm 64 vs 36 or 78 more since the mechanical arm cannot perform complex motions as easily as the human arm motions have to be broken up in order to prevent the hand from hitting something when it tries to null out errors and mechanical arm motions take longer to complete especially when nulling out errors Consider GRASP vs CENTER and CLOSE 42 jiffies 0 7 second are required by the human operator for IO GRASPs during the whole assembly while the mechanical arm requires 805 jiffies 13 4 seconds for 1 I CENTER and CLOSE Average motion times were 4 vs 64 jiffies 0 07 vs 1 07 seconds or a factor of 16 times Grasping is thus performed a lot more quickly by the human arm than by the mechanical arm The human operator can move his fingers according to what he sees while the mechanical arm in the CENTER operation closes the hand until one touch sensor is triggered and then moves the arm until both sensors are triggered By moving only small inertias the human operator is able to accomplish the GRASP much more quickly than the mechanical arm While the difference in total number of GRASP and CENTER and CLOSE may be small their distribution between the different tasks of the assembly 15 different T
55. 9 888 lt 65 019 YHAT PV XHRTsRe 5 85 VECTOR 768 649 600 lt 4 610 XHAT PV YHAT Re 5 85 VECTOR 760 649 888 lt 5 00 YHRT PV XHRTeR 5 855 VECTOR 760 649 888 lt 4 018 PV YHRTeRe 5 85 VECTOR 760 649 000 lt 5 610 YHRT PV XHRTsR amp 5 85 VECTOR 768 649 800 lt 1 8 18 XHAT PV 9 000 ZHAT WHERE N LROTN ROTN 2HRT U PV X Y 2 Applying the algorithm gives two possible orientations ESTIMATE LIST ITEM 16 Xs 204 TO 204 vis NE 555 0 555 ei 881 T O 881 H 87 36840 6 O 92 632 0 COS H8 000 SIN WO 1 888 COS OW 999 046 IV 59 gt Co RA EN Figure Box in Fixture Figure A 2 Screw on Driver IV 60 ITEM 17 X1 204 264 555 TO 555 2 901 TO 881 92 632 DEG 87 368 0E6 COSMO 000 1100 1 888 0 04 999 R 046 These results also iilustrate the replacement of equality constraints with a pair of inequalities here Z goes from 0 001 to 0 001 This approximation 1s not strictly necessary However it proved useful in some other cases where overdetermination was a problem Second Problem We now assert that one of the corner edges of the box is in contact with a side of the fixture contacts bxbtm 61 55 inside of con tacts be9 bj1 52 ex ten t_irrelevan t This gives XHATe
56. ATION In an earlier report 6 we discussed the design criteria and operational specifications for a six degree of freedom force and moment sensing wrist that was designed and built for the 111 7 Stanford Arm At that time we noted the importance of calibrating the force wrist since we found that unpredictable mechanical coupling had caused our theoretical estimates of the response of the wrist to be in error by as much as 5 percent Furthermore while the thermal drift and hysteresis were fairly small it nevertheless seemed a wise precaution to re calibrate the wrist from time to time For our initial tests we calculated the calibration matrix using a method that required that we apply three orthogonally oriented forces and three orthogonally oriented moments to the geometric center of the wrist To this end we set up an elaborate system of pulleys and weights While this type of procedure is acceptable for testing purposes it will prove to be too time consuming to employ when the wrist 1 in daily use Indeed once the force wrist is mounted on the manipulator applying pure forces and moments to the force sensing wrist may beimpossible without either detaching the hand or attaching special collars more acceptable method than either of these two alternatives is to use a method of calibration that can deal with coupled combinations of forces and moments Also in order to minimize the set up time we wanted to employ a method of calibratio
57. AUTOMATIC CODING OF PROCRAM ELEMENTS In Section 2 3 we described the process of writing AL code for a common subtask in assembly operations insertion of a pin into a hole We saw that writing the program required a number of decisions based on our expectation of where the objects will be and how accurately their positions will be determined at runtime The discussion of Chapter 2 focused on the modeling requirements for automatic coding The key point was that automation of coding decisions requires that the necessary information be represented in a form usable by the computer This chapter describes the use of the computer s planning model to make these decisions automatically The program outline followed 1 essentially that derived in Section 2 3 1 Grasp the pin 2 Extract it from the pin rack and transport it to the hole via a point just above the hole 3 Attempt insertion by moving the pin along the axis of the hole until a resisting force is encountered Use the distance travelled to determine whether or not the pin insertion is successful 4 If the insertion is unsuccessful then use a local search to attempt to correct the error The decisions that must be made include Where to grasp the pin 2 How to approach the hole Although we have decided on a co axial approach we still must decide the relative rotation of the pin and hole frames 3 What threshold values to use on our success test Also whether or not it
58. Artificial Intelligence Laboratory Memo AIM 177 Stanford Computer Science Report STAN CS 72 311 November 1972 2 Berthold P Horn Hirochika Inoue Kinematics of the MIT I VICARM M anipulator Massachusetts Institute of Technology Working Paper 69 May 1974 3 Donald L Pieper The Kinematics of M anipulators Under Computer Control Stanford Computer Science Report STAN CS 68 116 October 1968 143 Richard A Lewis Autonomous Manipulation on a Robot Summary of M anipulator Software Functions Jet Propulsion Laboratory Technical Memorandum 33 679 March 1974 5 P C Watson S H Drake Pedestal and Wrist Force Sensors for Automatic Assembly Proceeding of the 5th International Symposium on Industrial Robots September 1975 pp 501 511 6 T 0 Binford Paul J A Feldman Finkel Bolles R Taylor B E Shimano K K Pingle T A Gafford Exploratory Study of Computer integrated ssembly Systems Prepared for the National Science Foundation Stanford Artificial Intelligence Laboratory Progress Report covering March 1974 to September 1974 AL PROGRAMS FROM HIGH LEVEL TASK DESCRIPTIONS Russell H Taylor Artificial Intelligence Laboratory Computer Science Department Stanford University author is currently a Research Staff Member in the Computer Science Department IBM T J Watson Research Center P 0 Box 218 Yorktown Heights N Y 10598 At the tim
59. DP 11 and _ packages them into a single load module for future reference In this way programs that are constructed piecemeal can be combined together into larger modules The source code for the various modules currently resident on the PDP 10 might be in part available on the PDP 10 a similar process to unloading creates a source file that combines the various modules together into one program Together these facilities allow programming by experimentation Routines are written and tested until they seem to work and then they are embedded in larger drivers A legal statement in the source language would be MODULE lt name gt that refers to a previously compiled module The compiler could either read the source code back in and compile it again at some cost of duplication of effort or recover the decorated parse trees from the compiled file The modules currently loaded in the 11 are therefore a dynamic set new ones can be added patched in and old ones can be removed A simple symbol table keeps track of where each module begins in core where it ends and where it is referenced which should only be in one place since procedures are not yet available To remove a module its 11 29 physical space is reclaimed and the place where it 1s referenced is patched to give an error should it ever be cafled While programs are being written in this experimental mode it 15 useful to be able to manually move the arm re
60. Eqn 3 5 1 ARpp 1 subject to constraints 169 b j on the free variables We are principally interested in three things Axis misalignment 0 between the pin and hole 2 Displacement error At along the axis of the hole 3 Displacement errors Ax Ay in the plane of the hole Each of these entities 1s discussed below 4 5 1 Axis Misalignment For suitably small values 0 may be approximated by 3 y Apt Thus we can use Eqn 3 5 1 to compute the maximum expected misalignment max 140 1 where A vector cos sin 0 1 38 At present we consider six values of ranging from 0 to 315 degrees For example suppose that we are considering the in hole position pin holestrans nilrotn vector 0 0 1 71 hand pinstrans rot zhat 315sdeg rot xhat I80sdeg vector 0 0 3 54 corresponding to grasping parameters W 315 degrees 7 180 degrees and d 3 45 cm and pin hole rotation angle 90 degrees We assume that the hand holds the pin with essentially no error but the hand may be subject to orientation errors of up to 0 25 degrees about the hand x y and z axes and the hole orientation may be subject to rotation errors of 5 degrees about the z axis These values give us an estimate of the pin hole rotation error ARpp 21 ROT G228 M nyMy 9 2 228 VM where My and Mz are related to infinitesimal rota
61. IL Process mechanisms The most primitive routines on each processor are those capable of receiving and sending notes The implementation could make use of the interrupts that each machine can generate on the other but at present this 1s not done the receiver on the PDP 11 sleeps for a short time between checks of the notebox and the receiver on the PDP 10 is a SAIL process that is explicitly called when communication is expected When the receiver gets a note it makes a copy and zeroes the first word of the notebox as a sign that the note has been seen When the sender 15 asked to transmit a note it waits until the proper notebox is free its first word is zero and then dumps the note in place putting the first word in last These routines are under the control of a process known as the server which perpetually loops calling the receiver to get a note and then deciding how to treat that note on the basis of its type For example the server on the PDP 11 must interpret RELBUF GETBUF and USEBUF The first involves only the free storage allocator the second also must call the sender to inform the PDP 10 that a buffer has been allocated The third note USEBUF implies that a message has arrived in which case the routine treatmessage is called to handle It Treating a message involves different actions for different kinds of messages If the message 1s a request a new process is started to handle it and eventua
62. Jr Source Language Debugging Tools PhD Thesis Computer Science Department Stanford University May 1975 Swinehart 74 Daniel C Swinehart COPILOT A Multiple Process Approach to Interactive Programming Systems Stanford Artificial Intelligence Project Memo 230 Stanford Computer Science Report STAN CS 74 412 PhD Thesis Computer Science Department Stanford University August 1974 Teitelman 74 Warren Teitelman Interlisp Reference M anual Xerox Palo Alto Research Center Palo Alto California 1974 revised December 1975 VanLehn 73 Kurt A VanLehn ed Sail User Manual Stanford Artificial Intelligence Project Memo 204 Stanford Computer Science Report STAN G 73 373 July 1973 updated by James R Low Hanan J Samet Robert F Sproull Daniel C Swinehart Russet H Taylor Kurt A VanLehn March 1974 IMPROVEMENTS IN THE AL RUN TIME SYSTEM Bruce E Shimano Artificial Intelligence Laboratory Computer Science Department Stanford University The author 15 graduate student in the Mechanical Engineering Department 111 1 KINEMATIC SOLUTION PROGRAMS To date several methods have been described for computing the joint angles necessary to position the Stanford Scheinman Arm at a given point with a specified orientation Pieper 3 presented a method of solution for the general case of any manipulator with three intersecting axes presented a geometric solution which has been used at the Arti
63. Re 1 8 VECTOR 797 797 888 lt 707 YHRTsRe 5 8Se 760 649 000 5 888 YHRT PV XHAT Re 5 85e VECTORD 788 649 888 lt 4 888 XHRT PV YHAT Re 5 85e VECTOR 760 649 000 5 868 PV XHAT Re 5 85 VECTOR 768 649 888 lt 4 888 XHAT PV YHAT Re 5 8Se VECTOR 760 649 800 lt 5 888 YHRT PV XHAT Re 5 854 VECTOR 760 649 888 lt 4 888 XHRT PV 5 852 VECTOR 769 649 888 55 808 YHRT PV XHRTeRe 5 85a VECTOR 760 648 888 5 4 888 XHAT PV YHRTeRe 5 856 VECTOR 760 649 888 lt 5 888 YHAT PV XHRTeRe 5 85s VECTOR 760 649 888 lt 4 888 XHAT PV YHRTsRe S 85 VECTOR 760 649 9000 5 5 989 YHRT PV XHRTsRe 5 85s VECTOR 768 649 000 lt 4 888 XHAT PV YHATeRe 5 85s VECTORD 768 649 800 lt 5 888 PV XHRTeRe 5 85s VECTOR 760 649 880 lt 4 888 XHRT PV YHATeRe 5 85 VECTOR 760 648 000 lt 5 888 YHRT PV XHRTeRe 5 85 VECTOR 760 649 898 lt 4 008 XHRT PV XHRTeRe S 85 VECTOR 760 649 000 4 008 XHRT PV 8 099 2HRT PV WHERE R NILROTNeROTN 2HRT 2 ESTIMRTE LIST ITEIM26 Xs 000 TO 288 Y MIS 0 10 950 001 COS H8 623 31168 1 888 0 1 880 023 Notice that we have now rid ourselves of the ambiguity in the gross orientation of the box Fina
64. T BASE PUT SPINDLE PUT SHELL HOVE BACK TIMES REQUIRED FOR THE ELEMENTS JIFFIES 60 jiffies 1 second TURN APPLY PRESSURE PUT HANDLE PUT BASE 40 PUT SPINDLE 62 PUT SHELL 51 ESTIMATED TOTAL ASSEMBLY TIME 830 JIFFIES 13 9 SECONDS 16 BREAKDOWN OF ASSEMBLY ELEMENTS FOR YELLOW ARM USING WAVE CENTER GOTO POSI TURN HOVE TION APPLY CLOSE PRESSURE 2 3 3 2 2 PUT HANDLE PUT BASE PUT SPINDLE GET SPINDLE PLACE SPINDLE TURN SPINDLE ASSEMBLE SHELL PARK ARM TIMES FOR ASSEMBLY BY YELLOW ARM USING WAVE CENTER GOTO POSI TURN MOVE TION APPLY CLOSE PRESSURE PUT HANDLE PUT BASE PUT SPINDLE GET SPINDLE PLACE SP I NDLE TURN SPI NDLE ASSEMBLE SHELL PARK ARM TOTAL 2835 1095 355 1330 ESTIMATED ASSEMBLY TIME 108 SECONDS V 17 BREAKDOWN OF ASSEMBLY TASKS BY YELLOW ARM USINGDRAPERANALYSIS POSI PRE TION CISE POS PUT HANDLE PUT BASE PUT SPINDLE GET SPINDLE PLACE SP I NDLE TURN SPINDLE ASSEMBLE SHELL PARK ARM ASSEMBLY TIMES FOR YELLOW ARM BY DRAPER ANALYSIS HFFIES EL ROT INT RET CCO INS CPX EASE ATE ERF JRN 1 00 RES ERT ACE ATE PUT HANDLE 50 PUT BASE 140 PUT SP NDLE GET SPINDLE PLACE SPI NDLE TURN SPINDLE 1170 ASSEMBLE SHELL PARK ARM chido uid sers v w v mim pe nl ESTIMATED ASSEMBLY TIME 10
65. ability of successful assembly The assembly of discrete parts 1 strongly influenced by imprecise components imperfect fixtures and tools and inexact measurements Production engineers must choose among alternative ways to select individual tolerances in order to achieve minimum cost while preserving product integrity Grossman describes comprehensive Monte Carlo method for systematically analyzing the stochastic implications of tolerances and related forms of imprecision The method is explained in the section M onte Carlo Simulation of Tolerancing This work is one example in which technology developed initially for programmable assembly is proving applicable in a much wider domain particularly manual assembly ALAID AN INTERATIVE DEBUGGER AL Raphael A Finkel Artificial Intelligence Laboratory Computer Science Department Stanford University The author is currently an Assistant Professor in the Computer Science Department University of Wisconsin Madison Wisconsin 53706 At the time this research was p amp formed he was a graduate student in the Computer Science Department at Stanford University IL 1 CHAPTER 1 BACKGROUND The road to constructing working code in any programming language can be long and tedious Several of the important milestones are these 1 understanding the problem conceptualization 2 creating an algorithm to solve it design 3 writing that algor
66. ad the position with ALAID and use the frame value as a constant in the program A simple facility that allows the result of a previous query to be embedded in a new query will allow the arm position to be embedded in the program under construction Each snippet of program that the user constructs is remembered as a module and together the modules can be assembled into a working program then stored for future reference 11 30 CHAPTERS COMMANDSFOR DEBUGGING This chapter demonstrates ten tative subset of the commands to be available in a full version of ALAID Some of these commands have been implemented in the first preliminary version others are proposed In his work on COPILOT Swinehart 74 Daniel Swinehart gave great attention to the use of East video displays for showing the state of a multi process job In addition to standard debugging and control commands he includes a set of display oriented commands to distribute the limited screen among the various data that could be shown and to point to objects of interest by moving cursors The display orientation of COPILOT could form a useful base for ALAID and the commands listed in this chapter would be enriched by the addition of display features It is likely that such facilities would be available only on the ALAID member that resides on the larger machine since space is at a premium on the small machine Therefore rapid redisplay of changing status may not b
67. anslated into new values to be deposited in variables in the running program This interfacing of high leuel or computationally expensive feedback is easily obtained by using a program to emit the commands that the human usually would feed into the debugger That same program would be in charge of controlling the television camera or other feedback device If the debugging commands are the same as the source language statements the debugger is also an ideal hook on which to hang strategist programs that reduce abstract task descriptions into AL programs 11 14 Section 5 Side Effects AL is designed with two independent types of programmable side effect condition monitors and affixment structures The ability to associate side effects with situations that are suspected of leading to bugs 18 very useful in debugging but the fact that hidden responses are taking place can be very confusing How can condition monitors be used to assist debugging If it is suspected that some code is failing because the arm is never feeling a desired force a condition monitor can be designed to stop the arm after some period of motion as if the force had been felt and the program can be continued In this case a structure 15 used to simulate a desired result so that other errors can be found The condition monitor can also be used to scan for the presence of a bad condition for example suppose that an erroneous signal is being emitted on an event somewhe
68. ar multiplied by a trajectory changes the overall timing of the motion This concept of increasing flexibility by explicit naming also arises in the context of limitations to the language The fact that the same suggestion arises naturally in the context of debugging supports the hypothesis that debugging and programming are very similar activities that are carried out in the same domain and require identical structures EM Section 4 Multiprocessor Environment The best way to gain the full power of the compiler without actually writing one for the execution system is to use the one that already exists we have seen the trajectory calculation problem is hard enough to warrant using larger computer for at least that stage of the compilation This desire leads to the need for a linkage between the target machine and the compiler that will allew parts of programs to be recompiled and reloaded during the debugging phase If the input formalism to the debugger is to be the same as the statement formalism of AL then for every debugging request it 1s necessary to compile the code implied in the request and make the resulting pseudo code available to the debugging process The idea of a two machine link in which one machine can monitor and control the other 15 found in some recent computers The best example is the DEC KL 10 which uses PDP 11 40 to monitor and control the PDP 10 main machine Another example is found in the CDC 6
69. aried research efforts that constitute the Computer Integrated Assembly Systems project The progress report is divided into two main sections work directly related to AL and assembly and work on computer vision and modeling It 15 through a comprehensive project of this sort that prototype systems will eventually be developed for practical programmable assembly AL SYSTEM AND ASSEMBLY Extensive experience has shown that the debugging process is vastly more time consuming than people are willing to admit In AL this problem is accentuated because the system Interacts with the real world in real time and because there are two computers and a language hierarchy In response to these problems Raphael Finkel has developed ALAID 1 2 an interactive debugger that allows AL program execution to be monitored and that provides a means of patching AL programs to avoid the otherwise lengthy debugging loop A LAID resides on both computers and partial recompilations are done on the planning machine which maintains symbol table information A by product of the ALAID design 15 that it can be used to interface complex feedback routines on the planning machine to the runtime execution of AL programs Finkel s work is described in the section ALAID Interactive Debugger for AL In industry as microprocessors spread multi machine hierarchies will be usual and debugging systems like ALAID will have wide potential application Within th
70. ategies ranging in cost from 4 08 seconds to 8 58 seconds These are then elaborated into unrefined motion strategies with time estimates of 5 47 seconds to 12 7 seconds A computer generated summary of the best of these strategies is shown below PHL SPEC 132757 PREL IMS NULL RECORD PICKUP PICKUP SPEC 162775 PRELIMINARIES NULL RECORD APPROACH OREMINA 2 88 a APPROACH BMRNIPsPINI amp TRRNS ROTN VECTOR 678 679 281 149 DEG VECTOR 3 59 7 13 GRASP OPENING 185 GRASP BHANIP PIN1 TRANS ROTN VECTOR 679 679 281 1494DEG VECTOR 8 0 3 54 GRRSP DETERM NILTRANS DEPARTURE POINT PINIsPINI amp TRRNS NILROTN VECTOR 8 0 6 79 GOODNESS 4 13 REMARKS W 90 6 dog Grasp Angie 135 dog Grasp Dirtancr 3 54 APPROACH PINI amp BHIeTRANS ROTN ZHAT 90 DEG VECTOR 0 0 2 54 DESTINATION PINIaBHI amp TRRNS ROTN 2 88 DEG VECTOR 8 0 1 71 TARGET PINI amp BHITRANS ROTN 2 98 eDEG VECTOR 0 8 4 25 XPORT TIME 1 34 GOODNESS 5 47 TAP NULL RECORD fields below aren t filled yet FINE TIME 000 PH 02 800 PH FP DX 808 The output has been edited slightly to improve readability IV 48 PH FP 0Y 888 PH FP ROT 888 In terms of the parameters described in earlier sections this strategy corresponds to 90 degrees Y 135 degrees grasp distance 5 54 cm dtry 4 25 cm standoff 2 54 cm 90 degrees Once
71. ategy is where to grasp the pin The present hand consists of a pair of opposed fingers which open and close through a range of about 4 5 inches On each finger is a circular rubber pad and in the middle of each pad is a microswitch touch sensor The AL center command assumes that the object being grasped will trigger the touch sensors whenever it is in contact with one of the fingers Since we intend to use center the finger pads must be centered on the pin shaft The important parameters remaining are thus Y The grasp angle between the pin axis and the approach vector 2 of the hand d The grasp distance along the pin axis 0 The orientation of the approach vector of the hand about the pin axis Following the convention that the long axis of pins is the z axis this means that the grasping position will be given by blue s pinsgrasp xf pinstrans rot zhat W srot xhat 7 vector 0 0 d Geometric Considerations In selecting values for these parameters it is important to guarantee that the hand not get in the way of accomplishing the task In general this might require much better geometric modelling capabilities than the system described here currently possesses Therefore we must assume a relatively uncluttered environment The following considerations are however enforced by the present implementation 1 The hand cannot intersect the body in which the hole is drilled As approxim
72. ation we enforce this constraint with two sub constraints for both the initial and target holes 6 When sensitive force sensors are added to the fingers center will presumably be modified to respond to forces on the fingers rather than triggering of a microswitch This would allow greater freedom in picking finger positions and would relax the accuracy req uiremen ts IV 31 The fingers may not pass below the plane of the hole b The hand s approach direction must be at least 90 degrees to the outward facing normal to the hole 2 The palm of the hand cannot intersect the shaft of the pin Method It is necessary to verify that arm solutions exist for the approach grasp and liftoff points However the computation required by the arm solution procedure 1 non trivial Thus we proceed by pretending that the hand is moved by levitation Arm solutions will only be attempted for those grasping positions that do not try to do something bad with the hand If we assume that conditions a and b above are sufficient to guarantee that the hand stays clear of any objects then we can ignore W in selecting grasping positions to consider Our overall selection method looks like this 1 Use the position of thepin in the initial and target holes to determine legal limits on Y and d 2 Use the limits established in step 1 to generate significantly distinct values for Y and d For each such Y d pair determine values of w for whic
73. avoid this problem would be to attempt small hand motions after Insertion and check for resistance For instance move pin to pinsrot x hat Osdeg on torque xhat gt J sozsinches do begin stop in hole checketrue end on arrival do begin If the motion goes all the way we lost in hole checkef alse end 10 Two objections to this check that the extra motion statements take time and that the box may be moved inadvertently Always Stop on Force Another possibility is to alter the insertion statement so that the successful insertion as well as a near miss will trigger a force monitor that stops the motion Success and failure can then be distinguished by looking at how far the motion actually went move to in ole position vector 0 0 2vinches Via pin withdraw hole approach on force pinszhat 02 do Stop distance off that inv in_hole_position pinsvector 0 0 0 if distance off lt 2xinches then missed box flag true else if distance off gt 2sinches then hit top flag true else in hole flag true An additional advantage of this plan to hit something strategy is that it is much less vulnerable to small errors in the vertical position of the hole If a fixed destination point had been used and the hole were slightly higher than the runtime value said it was then the forces produced as the arm tried to servo to the nominal position could become quite larg
74. bly resulted in the collision of the fingers with the part even though a vertical approach vector had been specified since although feedback was used to correct errors at the end of the allotted motion time there were errors in some of the joints which had to be corrected While the fingers did eventually get to the position desired they did so with a lot of pressure and hard pushing against the part since the approach vector may tend to be tilted slightly from the direction of force application To overcome this problem the arm was asked to go to a point vertically above the part to be picked up and then told to swoop down upon the part in a vertical motion so that there was no danger of lateral movements of the fingers hitting the part since joint I the motor at the shoulder did not move The arm performed differential motion precisely when the movement involved only one joint and the change was of the nature of the joint movement e g angular motion could be performed precisely by the rotary joints as long as the joint axis was parallel to the axis of the desired rotation This was illustrated particularly when the hand performed rotation precisely around the z axis when the wrist was vertical but tended to change the wrist orientation when told to make a differential vertical motion FORCE CONTROL OF THE MANIPULATOR has shown that arm can exert forces with a typical tolerance of 10 oz Depending on the motor used the tolera
75. cannot be big enough to cause confusion e lt where 7 is a suitable fudge factor 0 75 designed to keep us well within the safe region If the maximum value of At falls within this limit then no further refinement is needed If not then tapping is considered as a means of getting the necessary accuracy To use this strategy the system must select a place to tap The principal considerations in making this choice are 1 The point should be as close to thehole as practical to minimize the effects of rotation errors in the hole surface and to minimize the time wasted in moving to a tapping place 14 Actually this consideration is too strong The right thing to do is to compute the expected misorientation and then use that result to compute the allowable distance from the hole IV 40 2 The point should be far enough from any confusing features like holes so that we are sure to hit the surface we expect to hit The method used is roughly as follows 5 surface into which the hole is drilled x44 location of hole in coordinate system of surface radius of hole radius of pin tip Arq 0 3 inches maximum distance of any point on s from the hole dbest 0 for r step Ar q until rmax do begin real amp for amp 0 step until 27 do begin real x 5 dj Xx xy racosk y
76. cant it would be fairly easy to provide a try harder mode where all possibilities are retained 4 3 3 Approach and Departure Positions The purpose n approach point for the grasping operation is to prevent the arm from trying to run its fingers through the m Currently the only approach direction considered is one along the approach vector of the hand as shown in Figure 4 2 One plausible alternative would be to move to a point above the pin and then move down along the pin axis to the grasping position If it should prove desirable to consider such alternatives we could do so by planning each route and then selecting the via point which gives the shortest time Similarly a departure point 1s needed to get the pin clear of its initial hole before trying to move it away We presently only use a standard takeoff point two inches above the hole move pin to pinstrans nilrotn vector 0 0 2 inches d where 4 is the distance the pin is inserted into its starting hole If this fixed choice should ever become troublesome it would be fairly easy to generate a set of alternative departure points and then pick the one giving the shortest motion time IV 35 Figure 4 2 Approaching the Pin 4 3 4 Hand Openings The present decision for hand opening is similarly arbitrary On approach the hand is opened by 1 inch plus the diameter of the pin at the grasp point The closure threshold 15 set to the pin diameter minus 0 1 inc
77. cation path between these machiries so that each machine can appear to be both a debugging user and a source of information from the point of view of the other machine This link has been implemented as described below 1 PROTOCOL FOR THE LINK The actual link implementation uses the hardware interface that connects our PDP 10 with the PDP 11 That interface allows either processor to generate interrupts on the other and the PDP 10 can read and write the PDP 11 unibus In this way information in the memory df the PDP 11 15 available to both processors In the following discussion it is assumed that the interface 1 a foolproof channel all communications in both directions reach their destinations Much work has been done in computer networking to provide for noisy channels most methods developed in such environments involve positive acknowledgement retransmission and sequence numbers These techniques could be employed in the ALAID context as well if the channel were unreliable since the present hardware 1 no less reliable than the processors themselves problems of inaccurate transmission have been ignored The current link is also less powerful 11 18 than a full fledged network based communcations protocol the two machines must be physically connected by the PDP 11 unibus These restrictions are not fundamental to the i idea of the two machine link In order to avoid conflict in the allocation of this mem
78. ce language In general it is much more difficult to infer the intent of a particular piece of code than to write code to achieve a particular purpose The computer has no understanding that our loop is intended to compute the maximum element of a It could not for instance decide because of some earlier code that is sorted and compile the loop as though it were maxel alik On the other hand if the user s program were expressed in terms of concepts like sort array and select the maximum element then the computer might in fact be able to write the appropriate code Recently there has been a great deal of interest in very high level languages in which programs are expressed in exactly this fashion e g 2 13 15 26 3 Occasionally a user map wish to share coding responsibility with the programming system Fdr instance he may wish to hand code the inner loops of an ALGOL program in the belief however deluded that he can do a better Job This creates certain difficulties for the compiler which generally only really understands code that it has written itself and there has been a tendency among language designers especially those wishing to enforce particular programming methodologies to outlaw such tampering An alternative would be to provide constructs that allow the user to tell the system about relevant assumptions or effects for aparticular piece of code such registers used to contain t
79. chine link can be extended to treat many machines The two issues raised in the previous section directionality and circularity pose the major problems to extending ALA ID to more members The circularity issue becomes worse with many processors since one question can generate several subordinate questions each of which can be directed to a different machine The directionality issue is the more severe Each member could have a list associating types of queries and members that are capable of responding to them If a query is not found on that list then it might be sent to each of the members in turn under the hope that the current recognition list is out of date If a member returns failure from a query it must also indicate what other members have been asked to look at the same query otherwise there could be wasted effort involved in sending the same query back to members that have already seen it A spanning tree that contains all the members could be used to direct queries it would be 11 24 illegal to pass a query along any link not in the tree or back across the link by which it arrived In this way it is also possible to reduce the interconnectedness of the entire graph of members It may be unreasonable that each processor should be able to recognize all the possible questions In this case a set of clearinghouse processors can keep track of who will answer a given query So the first action on receiving a query that cannot be
80. d on one machine and is being executed on another The XNET debugger Beeler 76 also operates in a multi machine environment The compilation phase not only transforms the input code into a form acceptable to the target machine but also develops a planning model of the values of variables throughout the program Furthermore it creates a symbol table associating the printing names of variables with level offset pairs One might expect the compilation phase to emit some of this state information into the output code so that the debugger can reside entirely on the target machine However since the compilation machine would be needed anyway if the programmer decides to rewrite a section of code it seems reasonable that the compiler or at least some of its tables remain available during a debugging run in order to assist associating names and planning values to variables and to help in making patches In this way the runtime environment does not become cluttered with too much information that is not directly needed during the execution of a program such information can be found in the other machine In order to make use of several machines it is necessary to have a form of communication between them In this case of AL the compiler process resides on the PDP 10 running under a timesharing multiuser system and the target machine resides on the PDP 11 running under the kernel The purpose of the link is to provide an efficient communi
81. d responding to requests For the sake of completeness one can imagine a portal connected only to symbol tables such a portal can be queried but will never generate a request Given the multiplicity of ways to direct queries that cannot be directly answered how can ALAID know that a query cannot be answered at all A simple approach for two member ALAID 15 to try only the other member and if he cannot help then the request 15 unanswerable If the impossible part 1s only a subset of the whole request it can be useful to report back that this particular part was the bottleneck Otherwise a simple failure return suffices complicated situations ensue if there is more than one portal to which the request can be directed or if there are several ways to subdivide the query into easier questions Then the process of finding an answer has the appearance of a depth first tree search every node represents subquery and has a set of alternative strategies each of which leads to a collection of other nodes that must be successfully treated for that strategy to work In order to prevent the search from becoming circular each query must contain some history that tells which members of ALAID have tried to answer the question and have failed The initial history list indicates the originator of the query If a member ever sees a query that he has seen before he immediately returns failure In fact queries should never be sent to any me
82. d summary of the actual internal structures You have been warned so don t vet confused 26 tapping place Point on the object to be tapped to reduce the error along the hole axis if necessary Ax Parameters summarizing the error footprint of the pin tip with respect to the hole Define a rectangular parallelepiped with sides Ax A2 rotated by about the hole axis See Figure 4 1 A6 Maximum expected tilt error for the pin axis with respect to the hole axis ttime Expected time spent in grasping the pin and transporting it to the hole finetime Time expected to be spent in fine ad justment motions Currently time for tapping motion search time goodness Estimated cost of this strategy Here ttime finetime goodness pickup Grasp Strategy object The object to be grasped preliminaries As before a list of preliminary actions that must be performed before the object here a pin can be grasped typical element would be code to put down a tool grasp point Point where object is to be grasped The structure used to specify such destination points is discussed below approach point Via point on the way to the grasp point approach opening Required opening for the fingers by the time the hand gets to the approach point grasp opening Minimum expected opening for the hand to hold the object grasp determ An estimate of the accuracy w
83. d to decide question 1 If the record is null it does nothing if a point is specified it emits the appropriate code An example may be found at the end of Section 4 8 Similarly in deciding whether to emit code for a search it lookt to see if Ax is greater than Arg If so the search is produced otherwise a perfunctory check ABS DISTANCE_OFF gt To d pds THEN ABORT Din MISSED hole UNEXPECTEDLY is written instead The program text produced for a typical strategy together with further discussion of the particular constructs used to implement search loops may be found in Section 4 8 Motion Sequences Both write_pickup and write_pin_in_hole use a procedure writ amp motion sequence to generate motion statements This procedure works roughly as follows 17 bmanip is an alternative name used in the current AL implementation for the blue arm and bhand is the name for the blue hand 5 Recall that in Section 4 5 3 we selected so that AxzAy IV 46 procedure write motion sequence amp t destinationssstring qualifiers begin integer i j k while j length destinations do begin iejej l controllable what destinations i while j lt length destinations and what destinations j41 scontrollable do jj corn men t Now destinations i destinations j is a subsequence with the same controllable frame print MOVE Jocation variable controllable TO location variable base de
84. de such process structure during the debugging phase what is clear 15 that each process that the user may wish to examine must be accessible either by an explicit name or implicitly in relation to other processes It is conceptually easiest for the debugger always to be pointed at one particular subject process known as the current process commands to examine or modify data or control structures will apply only to that process If another process should encounter a breakpoint it will send the debugger a message and wait but will not be automatically connected as the current process This concept can be fruitfully generalized to the idea of current context which is a related set of processes all of which are under scrutiny If the context includes only one process and it splits into several new ones then each of the offspring are also part of the same context no explicit switch 1s necessary to examine them Contexts are related to each other according to the lexical pattern of the program it makes sense to switch to the next more general context the parent or to one of several more specific contexts the offspring Parallelism makes the problem of backing up especially difficult One seldom discovers a bug until it has caused incorrect actions the debugger must assist the user in restoring the world to the state it had before the bug struck in order to track it down and try repaired code The 11 10 essential ability is t
85. different grasping orientations as well One of the key ideas of the earlier work was planning by progressive refinement Within this paradigm a program outline is prepared then elaborated into a more detailed one and the process 15 iterated until a finished product 18 produced The advantages of this approach are that planning for individual operations can proceed within the context of other parts of the program and that effort 15 not wasted on contradictory or irrelevant strategies Before these advantages can be obtained for real manipulator programs we need a better understanding of how individual coding decisions affect each other Although 3 Sacerdoti 22 28 successfully applied similar ideas to a different domain 54 the modelling techniques developed in this dissertation particularly those for representing object relations and for relating planned actions to accuracy information can perhaps provide a basis for such understanding much very hard work needs to be done The development of a good constraint formalism for position requirements discussed earlier would be especially helpful 5 1 Acknowledgements I wish to thank my thesis advisory committee Jerome Feldman Vinton Cerf and Thomas Binford for their continued interest and encouragement and for many suggestions which have helped give this work some measure of coherence Also I must thank all the many people with whom I have discussed various aspects
86. dings that combine forces and moments along the X and Y axes the manipulator is now directed to locate and pick up any convenient object in the work area After the object has been grasped all that we need to know 15 the position of center of mass of the object relative to the center of the force wrist For this purpose it 1s convenient to work with a fairly symmetric object that can be grasped such that its center of mass coincides with the geometric center of the finger tips li this is true then the weight of the object can be determined by repeating step 1 with the object in hand The new readings can be scaled against the old and the weight of the object can be determined in terms of the known weight of the hand We will call the weight of the object WT We can now repeat step 2 with the object in hand The force vector corresponding to these readings will be 2HW 2WT 0 0 0 2HWsDCM 2WT DH 0 5 We now duplicate step 3 with the weight in hand to obtain a new set of readings The force vector produced bythe combined weight of the hand and object will be 02HW 2WTT 0 2HW DCM 2W T DH 0 0 6 For the final force vector the manipulator must grasp an object fixed in place This can be a vise another manipulator or even a willing and strong human volunteer After ensuring that no net forces or moments exist along any of the axes the motor of the last rotary joint of the arm 1s driven with a constant and known torque T Readings are taken
87. ds Variation The example above required a search but no tapping since the error along the z axis of the hole was much smaller than the expected penetratiion of the pin into the hole if we Some people have commented on thecomputational inefficiency of generating possibly many values of x y which will be thrown away For any reasonable error limits however this cost can be ignored since the time required for moving the manipulator far exceeds that rquired to compute a target point IV 51 increase the uncertainty along this axis then a tap or some other measurement must be used before the insertion can be tried This is illustrated by the code below which was written for the same assumptions as those used earlier except that the box position is assumed to be subject to no rotation or in plane displacement errors but may have an error of up to 0 75 inches up or down PIN IN HOLE STRATEGY 1343561 DROK 762 FPX 243 FPY 226 FPU o 624 02 188 ESTIMATED TIRE 6 67 PICKUP 147157 90 0 Grasp Angle 135 dog Grasp Distance 3 64 OPEN BHAND TO 2 98 MOVE TO PINISTRRNS ROTN VECTOR 679 679 281 J49eDEG VECTOR 9 0 3 54 VI PINISTRRNS ROTN VECTOR 679 679 281 149 0 6 VECTOR 3 59 0 7 13 CENTER BMRNIP ON OPENING lt 185 DO BEGIN RBORT GRRSP FAILED END RFFIX PI N1 TO BMANIP _ MOVE 1
88. e provided that someone wanted to do the necessary programming on the runtime machine 16 If we could first know where we are and whither we tending we could then better judge what to do and how to do 1t Abraham Lincoln Speech before the Illinois Republican Convention June 16 1858 Chapter 3 PLANNING MODELS 4 Introductory Remarks This chapter explores the relation of planning information to programming in general and to manipulator programming in particular Programming 1 a form of planning the essential quality of a computer program is that it is a prior specification of how the general capabilities of the machine are to be applied to a specific problem Every program embodies some assumptions about the special circumstances in which it will be executed Thus an inherent part of the programming process is the maintenance of information about the predicted execution time environment and the use of such information as a basis for programming decisions Indeed the intellectual burden of maintaining such a planning modd is one of the major factors in determining the effectiveness of a particular programming formalism when applied to a task domain This burden cannot be escaped if we wish to help the programmer by taking over some of the coding effort then the computer must keep track to the information relevant to the coding decisions it 1 asked to make 5 2 Planning Information in Algorithmic Languag
89. e If the hole were slightly below nominal then no real damage would be done for this particular task since the pin would most likely drop into place when released However other tasks are not so forgiving If we were inserting a screw for instance the initial insertion must bring the screw threads into contact with the threads in the hole In such cases it is much better to get a positive contact than to rely on brute force accuracy Tapping An important requirement for using distance travelled along the hole axis as a success criterion is that the plane of the hole and the expected penetration distance be known well enough so that the various cases can be distinguished In this case there is no problem since the pin goes in a considerable distance and the box sits firmly on the table However we may not always be so lucky For example the box might have been placed in a vise Instead of aligning pins we could be inserting screws that go in only a short distance before the threads engage In such cases it is sometimes possible to win by deliberately missing the hole on the first attempt and then using the result to tell where the box surface is This might be done as follows Actually this is a slight oversimplification since we would probably push down while driving the screw IV 11 move pin to spot on surface vector 0 0 1 Osinches via pin withdraw spot on surface vector 0 0 1 Osinches on force pinszhat g
90. e additional measurement If the potential error has now been sufficiently limited then the tapping place is entered into the strategy record and the estimated cost 1s updated to include the time of the extra motion In this case the reduced error is Az 180 cm which is much smaller than the required accuracy of 1 71 cm and the estimated extra time is 1 2 seconds If no tapping place can be found then the system currently must give up on the strategy and hope that one of the other grasping positions will produce more accuracy along the hole axis Unfortunately this hope 15 frequently a forlorn one Eventually we would like to consider other measurement tricks to try if tapping doesn t work These alternative tricks could be weighted according to their expected cost and best combination picked 4 5 3 Errors in the Plane of the Hole These errors cause the pin to miss the hole and are overcome by searching To estimate in plane errors we compute Ey max cos p sin 0 A pn for 30k degrees Osks5 Then we take 18 Negative values mean outside surface on top of a hole 42 amp 6 This produces an error footprint rectangle with sides 24x and 2Qy rotated by with respect to the hole We set max Ax Ay A typical instance of this calculation 15 illustrated below Here the nominal pin and hole positions are
91. e AL run time system the speed of routines that transform between Cartesian space and joint angle space is of considerable importance Bruce Shimano has derived faster procedures for computing these transformations Additionally he has found simpler procedures for calibrating force sensors needed in those industrial applications that involve force controlled compliant motions Compliance using sensing is an essential means of coordinating two or more devices Shimano explains his contributions in the section Improvements in the AL Run Time System Russell Taylor has been studying ways to generate AL motion programs automatically from higher level task descriptions A paradigm of progressive refinement is used to expand single statement like put peg in hole into a succession of detailed steps necessary to get the job done The task is non trivial if attention 15 given to making the generated code rugged with respect to positioning errors The work therefore requires an ability to maintain extensive planning information concerning the description of the semantics of the manipulator language the definition of the task the objects being manipulated and the execution time environment Taylor discusses his approach in the section Generating AL Programs from High Level Task D escriptions Shahid Mu jtaba has analyzed the automatic assembly of a pencil sharpener and compared the motion times for the Stanford Arm with those obtain
92. e calibration matrix which was described by Watson 5 The calibration matrix C and its pseudo inverse Cn are related by the following formula Cn x Cn yl x 34 The pseudo inverse matrix has terms called 1 to 8 1 to 6 The pseudo inverse is analogous to the normal inverse matrix but it is defined for non square as well as square matrices Cri matrix can be used for computing the response of a single strain gage to the application of a specified force vector This relationship can be written as follows CnxF 35 Once we have computed Cn we can use equation 34 to compute the elements of the forward calibration matrix C 2 Computing the Pseudo Inverse Calibration M atrix To compute the elements of the pseudo inverse matrix six independent known force vectors must be applied to the wrist These force vectors need not be orthogonal nor do they have to pure forces or moments however we will require that their values be known at a single point whose position is known relative to the center of the force wrist For each of these force systems all eight strain gage readings are to be recorded We will define the values of the force vector and the readings as follows the reading of the jth strain gage due to the application of the ith force vector i 1 to 6 1 to 8 fik kth component of the ith independent force vector i l to 6 k 1 to 6 For each of
93. e possible in the ALAID environment but even occasional redisplay would be useful The commands are divided into functional groups by the entities they deal with internal state of ALAID data structures control structures control flow and advanced commands Each group has three sections First the set of relevant typeout modes is introduced These modes dictate which of several equivalent forms output is to take Next commands for examination are listed each with a brief description Last commands for modification are listed again with descriptions MODES Many of commands require specification of variables or code For example in order to ask for the current value of a variable one needs to name that variable In general there are several alternate formalisms Some can be immediately recognized in the PDP 1 1 and others require the symbol tables that reside in the PDP 10 Whenever alternatives exist the user should preface his typein with an identifying word that indicates what type he 1 using This is a simple way of restricting the input syntax to avoid complex type determination It is not necessary to the ideas behind the debugger In the following discussion typein modes will be introduced as needed As example to refer to the variable creativity by its source language name one would type SNAHE creativity 1 31 Section 1 Internal State of ALAID The internal state of ALAID consists of
94. e same mechanism a multiple world assertional data base is used by the automatic coding procedures discussed in Chapter 4 to keep track of situational information Object Information Object modeling is done by attribute graphs in which shape information is represented in the nodes structural information by links and location information by properties of the links The most interesting point 1s that coding decisions can generally be based on local properties of the object IV 23 Situational Information In order for the computer to make reasonable coding decisions it must often have numerical estimates of object locations and of how accurately those locations will be known at execution time Techniques were developed for expressing semantic relations between object features in terms of mathematical constraints on scalar degrees of freedom and for applying linear programming techniques to predict limits in inter object relationships Differential approximation methods were also developed and used to predict errors The appendix to this paper gives examples of both techniques Action Informrtion In the system described in this report action information is not represented explicitly Instead it is embedded implicitly the procedures that make the coding decisions and generate the output programs as we will see in the next chapter IV 24 J Watch me pull a rabbit out of my hat Bullwinkle Moose Chapter 4
95. e this research was performed he was a graduate student in the Computer Science Department at Stanford University EE 1 Chapter 1 INTRODUCTION This report summarizes recent work on the automatic generation of AL programs from high level task descriptions It is divided into three major chapters First the AL language is reviewed briefly and an extended programming example is used to illustrate the problems that arise when people write manipulator programs Next the modeling req uiremcnts for automatic coding are analyzed since the automation of coding decisions requires that the necessary information be represented in a form usable by the computer Finally the extended programming example is revisited this time with the computer using its planning model to generate the AL code automatically The material contained in this report is a condensed version of part of my dissertation 28 The main omissions are a discussion of the AL planning model object models and the representation of location and accuracy information although the appendix in this report summarizes some of the latter material IV 2 Chapter 2 THE AL PROGRAMMING PROCESS 2 1 Introductory Remarks This chapter provides a brief overview of the AL language illustrating its characteristics by an extended programming example which allows us to examine in some detail the AL programming process 2 2 Overview of the AL Language Superficially AL
96. ebugger the same power as the source language there is no reason why a changer could not be patched into the graph structure for a variable or a condition monitor spawned by the debugging process to perform the desired tracing In this way the power of the side effects can be brought under the control of the debugger Side effects although easily brought into the house as pets are not so easily domesticated Those instances wherechangers calculators and condition monitors influence the values of important variables or arm motions can be difficult to perceive Mistakes in affixing frames 11 15 can cause particularly opaque results All that can be observed is that the arm goes to the wrong place But how the destination frame managed to get such a value may be a complete mystery Condition monitors cause a new flow of control to temporarily exist and during that time commands that move arms may be encountered In fact it is possible to program changers so that a side effect of assignment into a particular variable is to park an arm The debugger must help the programmer figure out the causes for alt observable behavior If the arm starts to move unexpectedly the user must be able to halt it and find the source code that is is causing the motion That means that the processes that implement condition monitors and changers must be subject to the same scrutiny as all other garden variety threads of control In particular
97. ed by the technique of Methods Time Measurement for a person doing the same assembly The task was also analyzed using assembly primitives developed at Draper Lab For this task the mechanical arm was considerably slower than the human Identifying the sources of this lethargy should make considerable speed improvements possible in the future This work is discussed in the section Case Study of Assembly of a Pencil Sharpener In recent years the manipulator hardware has stabilized and the need for hardware construction has declined considerably Nevertheless certain hardware necessary for AL s operation is being developed These additions which are not described in this report include hardware that provides increased PDP 11 memory for the run time system a second arm interface to facilitate testing AL s novel software capability of controlling two arms in simultaneous coordinated motion a force sensing wrist to provide greater accuracy in determining forces and torques than is possible by monitoring servo errors and an improved gripper that should allow greater versatility in grasping small objects 1 3 COMPUTER VISIONANDMODELING Robert Bolles has shown that the execution time and memory size of his experimental program for visual inspection are almost practical and that programming is simple His recent work is concerned with establishing a firm mathematical basis for making verification decisions A least squares tec
98. ee part move that uses GOTO primitive to make a gross motion POSITION Motion which allows some form of mating of one object with another or adjusts the position of the hand so that the next motion can be easily executed OPEN Has the same effect as releasing the object WAIT A short pause between movements TURN amp Turning about an axis while applying force along the APPLY PRESSURE _ axis usen a e Aran ali gt 5 pe ae vri BR V 143 COMMENTS It should be noted that the comparisons are made between assembly primitives rather than motion primitives except where it is of interest to show correspondence between them As there was no one to one correspondence with MTM and WAVE primitives because similar motions could be specified in several ways in WAVE it was decided to use primitives similar to those in MTM for the mechanical arm Different fonts are used when referring to different primitives to enable easier recognition of what the primitives are The fonts are summarized below MTM ASSEMBLY DRAPER ASSEMBLY WAVE MOTION WAVE ASSEMBLY No distinction is made between MOVE and REACH in the case of the mechanical arm since the parts are so small and light compared with the arm that it does not make a difference in the movement whether the arm is carrying anything or not WAIT are tabulated since these were explicitly inserted for the purpose of preventing the overlapping of consecutive
99. efined names Along with a current process currently under scrutiny by the debugger is a symbol table that dictates how that process names its variables The naming problem takes a different form with regard to names of processes themselves Somehow it must be possible to point to a process and sic sic the debugger on it Naming problems are exacerbated by the fact that processes come and go during the execution of the program Itis necessary to be able to name a process that does not yet exist in such a way that when it does exist the debugger will use its symbols There may be some use to structuring the set of processes that have been examined so that one may return to a previous one stack of processes that have been under examination 15 one such technique but it suffers from the fact that the order in which one examines processes during debugging 15 often independent of the actual structure of processes in the program Therefore stacking the processes violates the dictum of naturalness which implies that the debugging structures should be the same as the programming structures A more attractive alternative to stacking 1 to name processes as parents siblings and offspring of other processes If the current process splits into three for a COBEGIN nest then each of the three can be considered an offspring of the current process and siblings to each other It gt is however unclear to what extent it is necessary to provi
100. en discussing ways for the program to discover that it has lost Once a 12 An optimist would say discover that it has won but this is unjustified There is bound to be at least one failure mode for which a program check has been left out Even if it 112 failure has been detected we must do something about it The simplest course is to give up if not in hole flag then abort Pin is not in hole A somewhat more graceful termination might include some cleaning up to get ready for the next iteration if not in Aole flag then begin Put your toys away move pin to pin Aolder via pin widthdraw We really should do some checking here too open bfingers to Osinches unfix blue from pin move blue to bpark via pin grasp approach is not in end In many cases this is all that can be done On the other hand it would be nice if some degree of error recovery could be built into the program Searches Even if the first attempt to find the hole misses it 1 plausible to assume that it is somewhere near where the runtime model says it 1s This suggests that we try searching the vicinity of our first attempt The original AL design included a very complicated search construct for doing this This construct has since dropped from sight the desired effect can still be had by means of a loop however were possible to anticipate and test for all failures it would not necessar
101. equence of motions of the manipulator hand The software system used for programming hand motions is of considerable importance in determining the ease with which a manipulator can be used and the path along which it moves For analyzing the hand motion times however software is much less important than hardware since the hardware characteristics of the manipulator largely determine the speed and types of motions which may be made It is believed therefore that whether the pencil sharpener is assembled using WAVE or AL will not significantly affect the types of motions used or the speed of execution We plan to check this conjecture by assembling the pencil sharpener both in WAVE and in AL The results reported here use WAVE The main results which emerged from this study are these a Special purpose fixtures were desirable for holding the parts in place so that the manipulator could work on them Plaster of Paris fixtures were easy to design and produce relatively cheap and adequate for the purpose b Positioning of parts could often be accomplished more readily by dropping the parts and tapping them into place than by trying to position them accurately c Analysis of the movements made by the manipulator showed them to be similar to human movements as defined by Methods Time Measurement MTM 2 but since the mechanical arm was larger clumsier and less versatile and had to avoid objects in the path of movement had to check the
102. er allows the program to continue The breakpoint influences only that process that hits it all others that are active will continue The message that the affected process transmits to the user includes context specification that uniquely defines which process it 15 in this way the user can set 11 35 the context appropriately before issuing investigatory commands SINGLESTEP Once a breakpoint has been encountered the user often wishes to execute a small piece of code and observe its effect The single step command allows a halted process to continue a short distance and then once again pause The process that this command affects is the one pointed to by the current context if that context includes several active processes then the command applies to all of them This exemplifies the convention that ALAID uses with regard to contexts all commands given affect all processes in the current context It is always possibk to restrict the context to contain only one process If the single step command is given to a process that is not in a halted state then the process will be halted The amount that an affected process will execute when singly stepping depends on the current code typeout mode if the mode is SADDRESS or SCODE then one statement of the source language is executed if the mode is PADDRESS or PCODE then one statement of the pseudo code is executed After the single step command has been executed each affected process
103. erdoti The N onlinear N ature of Plans Stanford Research Institute Artificial Intelligence Center Technical Note 101 January 1975 Earl D Sacerdoti A Structure for Plans and Behavior Stanford Research Institute Artificial Intelligence Center Technical Note 109 August 1975 Hanan Samet Automatically Proving the Correctness of Translations Involving Optimized C ode Ph D Dissertation Stanford Artificial Intelligence Laboratory Memo AIM 259 Stanford Computer Science Report STAN CS 75 498 May 1975 Robert F Sproull Title Unknown Ph D Dissertation Stanford Computer Science Department Summer 1976 J T Schwartz Automatic D ata Structure Choice in a Language of Very High Level Courant Institute NYU 1974 IV 57 27 Norihisa Suzuki Automatic Verification of Programs with Complex D ata Structures Ph D Dissertation Stanford Artificial Intelligence Laboratory Memo AIM 279 Stanford Computer Science Report STAN CS 76 552 February 1976 28 Russell H Taylor The Synthesis of Manipulator Control Programs from Task Level Specifications Ph D Dissertation July 1976 29 Richard Waldinger Achieving Several Coals Simultaneously Stanford Research Institute Artificial Intelligence Center Technical Note 107 July 1975 1V 58 Appendix A EXAMPLES OF LOCATION AND ACCURACY CALCULATIONS Box in a Fixture This sequence of problems illustrates the translation of symbolic relations into constraints and s
104. es 3 2 1 An ALGOL Fragment Consider the ALGOL fragment below which is intended to select the largest element from an unsorted array 8 IV 17 integer array 1 100 integer n maxel maxel e 237 largest negative number in machine Assume we want the maximum of the first n elements of a for ie step J until n do if maxel a i then maxelea i When we write the statement in the loop body we know that variable i will contain a value between 1 and n that maxe sa j 1514 that maxelsa j for at least one j in that range and that by the time the loop has exited we will have examined all values for i from 1 to n Further we assume ns100 A process of great interest to researchers intent on proving the correctness of programs has been the formalization of these assertions and the use of well formulated language semantics to prove the assumptions correct I 1 27 1 similarly one of the strongest claims of structured programming advocates 1s that one should proceed from such assertions to a correct program 9 There has been a great deal of interest in applying theorem proving methods to automate the generation of programs from assertions e g 161 My own impression is that one does not usually write programs in such a step by step fashion Rather than working out from first principles how to synthesize this loop to compute a maximum element most programmers would reach into a grab bag of tric
105. f these phases in greater detail 4 Data Structures Internally strategies are tepresented by SAIL record structures summarizing the decisions that have been made This section describes the more important parameters kept for pin in hole and pickup strategies Pin in Hole Strategy prelim in aries A list of preliminary actions that must be performed before the code for the actual pin in hole code is begun Typically this involves cleanup actions left over from the previous task and 1 set up by the initial processing pickup pickup strategy to get the pin affixed to the hand and free of any obstructions For picking up an object by grasping it in the fingers this field would point to a uin strategy defined below dtty The distance into the hole that we will try to poke the pin standoff The distance above the hole that we will place an approach point The relative rotation of the pin to the hole upon insertion The combination of grasping method approach symmetry and expected penetration distance into the hole constitutes sufficient information to write a first order program that ignores errors such as was produced in Section 2 34 However as we saw earlier the job isn t yet half done The structures shown here are slightly different from those actually kept The changes have been made for ease of explanation the information content is the same Section 4 8 includes a computer generate
106. face coplanar with the hole or else miss the object altogether and so on 20 These assumptions do not get built into programs by accident Information about objects is used extensively in both the problem solving and coding functions involved in manipulator programming In mechanical assembly programs the task is largely defined by the design of the object being put together In addition to specifying what is to be done the design also dictates many aspects of how to do it such as in what order the various parts must be assembled how the parts can be grasped by the hand or put in a fixture what motions are required while mating parts and so forth For manipulator programming the most important aspects of object descriptions derive from the shape of the objects being manipulated Unfortunately good shape representations for computer use have yet to be developed Many decisions that are intuitively obvious to a human programmer require a laborious computation by the computer On the other hand it is possible to identify many local properties that play an important role in coding decisions For instance in coding the pin in hole example of Section 2 3 we used object information in a number of ways Filling in parameters The most obvious example is the location of the hole with respect to its parent object affix hole to box at trans nilrotn vector 2 8 cm 3 2 cm 4 9 cm y Other uses include setting the minimum grasp t
107. factor of 2 in speed and a factor of 3 number of movements In each of the assembly subtasks the mechanical arm does three V 19 times as many POSITIONS as the human arm since it has to get itself vertically over the part before grasping and vertically above the main fixture before releasing the part and then actually position the part in place The human arm does 9 RELEASE in 38 jiffies 63 second compared to 12 in 355 jiffies 5 92 seconds by the mechanical arm i e 4 vs 30 jiffies 07 vs 5 second per motion which means a factor of 7 in speed and a factor of 1 3 in number of movements The human does twice as many RELEASE in the turning of the spindle as the mechanical arm just as it did twice as many GRASP but in the placement of the parts the mechanical arm does twice as many OPEN since the mechanical arm must first open the right amount to grasp the part then do a second OPEN to release the part The mechanical arm requires 3 WAITs after opening the hand to ensure that the subsequent motion does not overlap with the opening of the hand Waits are not tabulated for the human arm since the human operator does not consciously have to take any discernable pauses between overlapping motions All the motions seem to be fundamental and necessary in the mechanical assembly except for the 3 WAITs which took 150 jiffies 2 5 seconds and trying to locate the position of the hole which took 4 MOVEs and 2 POSITION and 1140
108. ficial Intelligence Project since 1972 More recently Lewis 4 described a method using vector cross products to obtain expressions for the last three joint angles and 2 presented a method of solution using Euler angles for the MIT Scheinman Arm The following is yet another method of solution which has the advantage of extreme speed The equations presented below were derived by Lou Paul and Bruce Shimano using two different methods one using vectors and the other algebraic Only the algebraic solution is presen ted here 1 Transformation Equations Given 4x4 matrix 1 representing the transformation from a coordinate system fixed in the hand of the manipulator to the base coordinate system we wish to find one set of link variables 0 0 5 9 0 0 for the Stanford Arm which will produce an equivalent transformation Ty Tis Ti 1 y l The transformations for the individual links of the Stanford Arm were derived by Paul 1 If we multiply these six matrices together symbolically and equate the components of the total manipulator transformation to the required matrix elements 1 we get the following 12 equations in the six joint variables where 5 denotes the sine function and denotes the cosine function T s0 s0 s0 0 0 0 50 50 cO s0 cO cO 0 0 50 2 Tyo 501050 0 0 0 50 cO 50 5050 0 50 0 50
109. form expressions for 0 we will now make use of the remaining two columns of the transform These first two columns give the orientation of the hand about its Z axis We will find it convenient to immediately combine equations 2 3 6 amp 7 and 10 amp 11 to eliminate some of the variables T 4 50 T c0 50 s0 cO cO cO 29 T5150 0 59 60 s0 cO cO 30 50 50 0 31 From these three equations we can obtain explicit formulas for the sine and cosine of 8 However rather than use these equations we can obtain much simpler expressions if we combine the three equations above with 22 24 and 25 to obtain the following formulas ILE id MIS 6 50 50 5 50 I his completes the solution for the joint angles of the Stanford Arm from a desired tranformation 11 6 4 Solution Execution Time A new arm solution program has been written which employs the equations presented in this paper The execution time for this routine is approximately 2 2 milliseconds on a PDP 1 1 45 using hardware floating point arithmetic This is roughly half of the time that was formerly required to compute the joint angles given a tranformation 4 R everse Solution Program To compute the hand transformation from the joint angles Horn 2 demonstrated that for the MIT Scheinman Arm it was very efficient to symbolically expand the matrix products of
110. h 4 4 Moving to the Hole Once the pin has been grasped and lifted clear of its initial hole the next step 15 to try inserting it into the target hole For the sake of simplicity we assume that the origin of the pin coordinate system is at the tip being inserted into the hole Thus our motion statement Will ook something like move pin to Aolestrans rot zhat vector 0 0 dtry via Aolestrans rot zhat vector 0 0 standoff on force pinszhat 8eoz do 2 This latter figure comes from the observed behavior of the center primitive relevant factors Include of the fingers and compression of the Anger pads IV 36 rotation angle of pin with respect to hole dtry distance try to push pin into hole standoff distance of approach point from the plane of the hole Of these parameters the most important is The considerations in choosing a good value are essentially the same as for selection of the grasping orientation Q9 The method followed is also the same except that a single value of 1s picked to minimize the expected motion time and the destination location is used instead of the initial pin location Thus the expected final position of the hand vill be given by hand destination pin destinationsgrasp xf holestrans rot zhat vector 0 0 d Q sgrasp xfi For our example situation Section 4 8 and grasping parameters Y 135 degrees 90 degrees d 3 54
111. h there is an arm solution Each Y 4 9 will then specify a possible grasping position for the pin 3 Once a grasping position has been generated the remaining parameters to the grasping strategy may be filled 1n and the cost of the strategy assessed This process may result in some of the proposed grasping positions being rejected due to inability to find a suitable approach or departure position or because of accuracy considerations These steps are discussed in somewhat greater detail below Determining values for and d To simplify the discussion let us assume that the pin initially has its z axis parallel with its starting hole and that the origin of the pin s coordinate system 1 at the pin tip inserted into the hole The first step in determining Y and d is to determine the distances di and dg that the pin goes into the initial and final holes There are two subcases The same end of the pin is inserted in both holes If additional feasibility tests are to be made this would be good place to include them For instance if good enough shape models e g those produced by CEOMED 4 available then a check can be made to see if the hand or arm do in fact interfere with ob jects in the environment Two problems with this check are 1 the difficulty of distinguishing intersections caused by approximations and those caused by actual collisions and 2 the difficulty of modelling sets of possible posit
112. he axis of the hole Az These quantities are given by the objective functions AX 3 28 800 28 6 000 000 3 18 1 08 00 900 008 V Ay C 3 96 28 8 008 888 3 18 808 000 1 88 080 1 88 0801 V Az 888 008 998 000 000 000 000 000 000 080 1 00 V where Vel 7 6 Solving these linear programming problems the system gets 1 57 lt Ax lt 1 57 1 57 0 62 inches 1 37 1 37 1 37 0 54 inches 127 s At 5 127 127 cm 0 05 inches Also we need to know the maximum direction error between the screw and hole axes This quantity will be given by where 0 008 1 88 000 008 1 88 000 V AQ 808 000 1 00 008 000 1 00 I and s 7 Solving gives us 0916 s A0 s 0916 0 0916 radians 0 525 degrees 4916 s A0 lt 0916 V CASE STUDY OF ASSEMBLY A PENCIL SHARPENER M Shahid Mujtaba Artificial Intelligence Laboratory Computer Science Department Stanford University The author is a graduate student in the Industrial Engineering Department V 1 INTRODUCTION A mechanical pencil sharpener was assembled using the Stanford Arm to gain insight into analyzing the mechanical assembly process The process can be considered as a sequence of motions of the component parts these motions in turn dictate the need for a s
113. he control store can be referred to either by octal location in the PDP 11 PADDRESS mode or by labels and offsets in the source code SADDRESS mode The contents of the control store can be shown either as pseudo instructions PCODE mode or by the source code that generated them SCODE mode Each process has a compiler generated identifier The identifier associated with the top level thread of a context can be used to identify the context CONTEXT BY IDENTIFIER mode 2 EXAMINATION SHOW CONTEXT The current context 1s displayed For example CONTEXT BY CODE SAOORESS LAB3 1 32 SHOW MODES Each of the current modes in effect is listed The only typeout mode in which this list can be printed is LIST mode For example LIST CONTEXT BY IDENTIFIER SADDRESS 3 MODIFICATION SET CONTEXT lt thread name gt The thread can be currently in execution or not f not then no information will be available for variables local to that thread The named thread can include many subthreads only those variables in active subthreads may be accessed The name of a thread can be given by the same modes used for showing the context MOVE CONTEXT lt list of codes gt The context is to be changed from the current thread One legal code is UP where is a positive integer This code moves the context to the surrounding thread n levels more global Another code is ACROSS where is any integer The context is to be
114. he human hand cannot rotate through 360 degrees and 2 GRASP 2 RELEASE need to be done for one done by the mechanical arm rotating through 360 degrees However the mechanical hand has to release the object and close the fingers before it can tap the part in place unlike the human who can position the object precisely while holding it all the time The human operator performs IO REACHes or MOVE in 341 jiffies 5 68 seconds compared to the 18 GOTO MOVE by the mechanical arm in 2835 jiffies 47 25 seconds which amounts to 34 vs 160 jiffies 0 57 vs 2 67 seconds per movement The human arm performs faster than the mechanical arm by a factor of 5 white the mechanical arm does 2 times as many movements as the human arm when it 15 trying to position the spindle in place The reason 15 that the human operator does not need to worry about nulling errors and utilizing visual feedback does not have to spend time trying to locate the relative positions between the spindle and the hole precisely something which the mechanical arm requires 4 MOVE and 2 POSITION to accomplish and in addition the use of movements that are too rapid result in overloading the motors with a demand torque that 15 too high The human operator performs 4 POSITION in 157 jiffies 2 62 seconds compared to 13 POSITIONS in1095 jiffies 18 25 seconds performed by the mechanical arm i e 39 vs 85 jiffies 65 1 42 seconds per movement which means a
115. he might want to track through the graph structure assuming that he understands how it is put together If he 1s trying to implement a new pseudo operation he may want to work at the level of machine instructions In general the debugger has the responsibility to make accessible all information and power that the user will need in a form that he can understand This implies that there should be commands for examining graph structure even if it 15 not subject to scrutiny in the source language The instructions that are being executed should be visible either in source formalism pseudo code or machine code at the desire of the user Part of this problem would disappear if the source formalism were directly interpreted in the target machine This 1 not an absurd idea although the implementation chosen did not follow that direction Debuggers for LISP See Teitelman 74 for example take full advantage of the fagt that all code in LISP is representable within the data structures native to the language Therefore there are no hidden structures except for compiled routines which are usually not used unless they are assumed to be bug free Short of implementing AL in AL some steps could be taken to add features to the language that allow examination of structures One such feature would allow the program to discover if one frame 15 attached to another It may happen that such investigatory functions would be useful in their own right as parts of
116. he results of machine language statements Another possibility investigated by Samet 24 for LISP progams to write both high level and hand coded versions of the same program The system can then verify that both programs are indeed equivalent even though it isn t necessarily clever enough to figure out the hand coded versi n on its own M IV 19 This sharing of coding responsibility is especially important early in the evolution of an automatic coding system when many things cannot yet be handled by the computer In Chapter 4 we describe procedures for making the AL coding decisions of Section 2 3 automatically Incorporation of this facility into a manipulator programming system requires either that enough primitives be available so that manipulator level coding decisions can be made by the system or that coding be shared perhaps by having the computer generate program text for subsequent modelling by the user Again some assertional mechanism is almost certainly necessary to help the system understand code written for it by the user 3 3 Planning Information in Manipulator Programming Many of the book keeping requirements of manipulator programming are essentially the same as those for algebraic programming One must keep track of what variables mean what things are initialized what control structures do etc In addition to these general requirements the domain requires the maintena
117. his approach has the drawback that we may spend considerable time refining strategies whose basic motions are so inefficient as to rule them out Therefore we first decide on the basic motions for each distinct grasp point All candidate strategies are sorted according to gross motion time and then considered in best first order If we reach a point where the next best unrefined strategy is more expensive than a fully planned strategy then we can stop searching strategies null for each g such that g is a grasping strategy do begin Decide best way to get pin to hole using g if there is a way then Create a pin in hole strategy amp insert it in strategies ranked by expected time e n amp IV 44 best strategy Incantation shortest time 1077 seconds very long time minimum refinement cost e lower bound on fine motion time for each 5 such that 2 strategies do begin if cost s eminimum refinement costzshortest time then done best strategy is the best strategy we ve found Refine sto account for accuracy considerations Revise the cost estimate for if cost s lt shortest_time then begin shortest time cost s best strategy 5j end end Here we have used minimum refinement cost to tighten our cutoff somewhat It may be computed by assuming that there is no error in the arm or grasp so that all error between pin and hole comes from errors in the hole location and then
118. hnique is used to combine available information and derive estimates for location and location accuracy of objects Bayesian probability is used to determine necessary confidences within a sequential pattern recognition scheme These well known techniques are combined to answer various questions raised within a verification vision system The work is described in the section M athematical Tools for Verification Vision Michael Roderick has been investigating the possibility of reducing the sampling rate used by the run time computer to control the Stanford arm His analysis is based on the use of z transforms since Laplace transforms are not applicable for low sampling rates Specific recommendations are derived for the lowest possible sampling rates at which the Stanford Arm might be controlled Roderick s approach is described in the section Discrete Control of the A rm A previous progress report described POINTY an interactive program for generating object models by manual positioning of the manipulator During the course of writing applications programs Shahid Mujtaba has found this technique to be an aid in reducing the labor of coding AL models He has prepared a guide to the system entitled POINTY User M anual useful for both training and reference purposes David Grossman has used a geometric modeling program to simulate discrete parts tolerancing showing how manufacturing errors can propagate until they affect the prob
119. hole for the insertion to succeed Thus and Arg constitute a measure of the effectiveness of accommodation during insertion The direction must be supplied by the user as part of the task description 4 In principle df and d may be computed by looking at the profiles of the pin and hole At one time this was done However the computation turned out to be extremely tedious and ignored some important factors such as friction Therefore these numbers were determined by experimenting with the actual objects and included in the object models as were Arg and This approach does not seem unreasonable since pins and holes may be standardized Presumably a data base could be built containing the relevant parameters for each tip hole combination encountered in a class of assemblies 4 3 Grasping the Pin Once the initial computations have been done we proceed to generate alternative strategies for picking up the pin For each such strategy we create a grasp strategy record as described in Section 4 1 Although we confine ourselves to grasping the pin directly between the fingers it 1s interesting to note that alternative methods such as loading a screw onto the end of a screwdriver could be handled similarly The rest of the pin in hole code except for the part about letting 20 makes no assumptions about what the hand actually holds onto The important data used by the rest of the planning arc l The relative pos
120. hows the output estimates that result from application of the iterative method described in my dissertation 28 Here we have placed our box into open topped fixture as illustrated in Figure A 1 In the first problem the box is allowed to rattle around loosely inside the confines of the fixture In subsequent subproblems we push the cornei edges up against sides of the fixture thus further restricting the box First Problem The box has been placed in the fixture with the bottom surface of the box in contact with the bottom inside surface of the box This 15 reflected in our data base by the assertion contacts bxbtm bjl sb inside of where bxbtm is the bottom of the box and 811 56 is the bottom of the fixture This produces the constraint set YHRTeR amp 5 854 VECTOR 760 649 886 lt 5 8 10 YHAT PV XHRTaR amp 5 85 VECTQIR 760 649 800 lt 4 818 PV YHRTsRs 5 854 VECTOIR 760 649 000 4 5 018 PV XHRTsRe 5 85 VECTOR 768 649 888 5 4810 XHAT PV YHRTsR amp 5 85 VECTOIR 768 669 808 4 5 018 YHAT PV XHRTaRe 5 85 VECTOR 760 649 000 lt 4 818 XHRT PV YHRTaRe 5 854 VECTOR 760 649 800 lt 5 010 YHAT PV 5 85 VECTOR 768 649 888 lt 4 810 XHAT PV VHATeRe 5 85 VECTOIR 768 649 868 lt 5818 YHAT PV XHRTsRe 5 85 VECTOR 760 649 000 0 lt 4 018 PV YHATeRe 5 85 VECTQRC 768 64
121. hreshold the expected penetration of the pin into the hole and selection of a grasp point that kept the fingers out of the way 2 Estimating the accuracy required to guarantee that the pin will seat properly in the hole The allowable error is determined by such factors as the point on the pin chamfering around the hole clearance between the pin shaft and hole bore etc It 15 Important in deciding whether the insertion method used here will work and in setting the step size for the search loop The object representations used in this work are described in my dissertation 28 It is important to note here that these uses predominantly involve local properties of features e g the chamfering around a hole or the placement of holes in a surface that are in principle computable from a uniform shape representation but which may also be represented directly in several different forms to serve different purposes 3 5 Situational Information Manipulation programs transform their environment by moving objects around This means that the principal fluent properties that must be considered are 1 Where objects are in the work station 2 What objects are attached to each other By fluent properties we mean any factors relevant to the task which may not remain constant during execution of the task E IV 21 3 How accurately relevant locations are known by the manipulation system The use of this information was ill
122. ill load the named module after whatever modules are already loaded so that several modules can be linked together Part of the LOAD command is to make the PDP 10 environment aware of the necessary symbol tabks The pair GET DUMP are used to save an entire state of the world DUMP creates the file SAMPLE ALD which contains the entire core image of the PDP 11 the information necessary to continue it and pointers to the necessary symbol tables in the PDP 10 GET reverses this operation In this way safe points can be created during the debugging of the program After a GET command has restored the state it 1s wise to Issue the command EXECUTE SCODE MOVE BARM TO BARH WITH DURATION 5 which will have the effect of slowly moving the arm from whatever position it happens to have back to the place that it occupied when the DUMP was performed 3 HISTORY Messages pass through portals in both directions Those portals that connect to humans contain the most important information from the user point of view therefore it is natural to keep track of that information so that it is easy to recover If ALAID has responded to a query it is likely that the user will wish to use that response as part of his next query Therefore two history lists are kept for each portal that leads to a user queries he has issued and the responses that have been engendered by them A legal field in any query is a reference to a previous query o
123. ily be economical to do so IV 138 if not in Aole flag then begin vector 428 scalar n dp vector 0 J inches 0 0 for n J step J until 6 do begin rot zhat 60edeg sdp _ Try to put pin in perturbed hole move pin to in Aole position dp vector 0 0O sinches via Aole approach dp vector 0 O Isinches on force pinszhat gt 8 oz do stop Check distance travelled etc if in Aole flag then ne7 This terminates the search end if not in hole flag then abort The hole doesn t seem to be there end There are many variations possible on this theme depending on how large an area 15 to be searched what pattern is to be used etc If vision 1s available we may want to use it to compute a correction for the next trial 2 3 8 Refined Program Combining a search loop with the other refinements and adding a check to be sure that the pin is successfully grasped we get the following program begin pin in hole f tame pin pin grasp pin grasp approach frame pin holder pin withdraw f tame hole in Aole position hole ap proach frame box affix pin withdraw to pin holder at teans rot zhat 30edeg vector 0 0 4sem pin holder frame nilrotn vector J5 inches JOsinches 0 affix pin grasp to pin at trans rot xhat J80sdeg vector 0 0 2 cm affix pin grasp approach to pin grasp at trans nilrotn vector 0 0 3ecm pin pin_holder IV 14 affix in Aole positio
124. ime values This approach produces well behaved motions so long as the required modifications are not too great However it also creates a number of problems for the compiler which must maintain the planning model Eventually it is hoped that trajectory planning can be done completely at run time However this will not eliminate the need for a planning model which is also used for affixments and for other purposes 2 3 Sample AL Program Manipulator programming is a non trivial intellectual activity even for simple tasks This section illustrates the use of AL to accomplish a simple assembly operation the insertion of an aligning pin into a hole which is a typical subtask for many assembly programs The discussion provides some insight into the process of writing AL programs First an outline for the program will be developed simple first cut program implements this task outline We then examine the flexibility and toughness of this program Methods for error detection and recovery are discussed and a new more elaborate program is produced 2 3 1 The Task Initially the pin sits in a tool rack and a metal box with holes in it sits on the table in some known position Our mission is to get the aligning pin into one of the holes The way to 5 is to grasp the pin between the manipulator s fingers extract it from the rack hole transport it to a point over the hole and insert it into the hole Thus our program in
125. immediately answered 15 to query the clearinghouse for the name of the members to which to direct the original query Thus a legal datatype of the response language includes member names What does it mean to have many processors running on one algorithm In the case of AL one is an execution processor the other is a planning processor The most general case is to have many execution and many planning processors In AL the two processors have very restricted normal communication The general case may include many channels of communication among the execution machines This communication could be handled exclusively through ALAID on each machine through the program portal In this way the communication link that is already present would be put to good use How to connect the various members together is a well studied problem ALAID takes some advantage of the interface between PDP 10 and PDP 1 1 The requirement for many way connection is that any machine must be capable of getting messages to any other machine One common memory with one non reentrant process for parceling out memory will work but short of that especially for more than 10 or so processors at which point such linking may get too expensive a new kind of message could be used the transit message It is not intended for the recipient but should be handed on At this point we are getting into networking not debugging issues As long as any processor can get a message to an
126. in a few ways Changers can be applied to a global variable in such a way that after control leaves the block in which the changer was applied the global variable still has the changer associated with it this anomaly should perhaps be considered more of an implementation bug than a part of the design of the language During the course of debugging it often happens that control must be forced to exit from a block or transferred into the middle of another block This facility 1s most easily implemented by associating information on variables that must be created and destroyed with each block entrance and exit A premature block exit is then simulated by jumping to the end of the block where the exit code is to be found If programmer wishes to make a wild jump from one part of the program to another it is only necessary to carefully close all the blocks that lead out to the first common ancestor of the current locus of execution and the desired one and then to enter all blocks that lead in to the specified place Once the appropriate block has been entered a direct jump to the indicated code should work Section 2 Parallelism The first problem that AL presents 1s its parallelism The source language itself allows the user to split control explicitly into several threads of execution These threads are treated as separate processes in the running program Condition monitors are implicitly also understood to refer
127. ing like this for w 0 step 45 deg until 315sdeg do begin trans Aand place grasp xfi grasp xf trans rot zhat srot xhat Y vector 0 0 d hand place e initial pin locationsgrasp xfi if solve arm Ahand place then begin cost e abs 45sdeg joint angle 5 insert into list of candidates ranked by cost gt gt end end For the example situation described in Section 4 8 and grasping parameters 10 Shimano is currently investigating the possibility of a closed form solution that will give the range of possible approach orientations for a given hand position Such a solution would be extremely useful both as a guide for selecting grasping positions and as a means for evaluatingthe robustness of a particular position under variations in object position Alternatives include examining the error hypercube at the fingers or just using the expected time to reach the grasping position The latter objective function will eventually be applied to any points that get through this filter see Section 4 4 34 7 135 degrees d 3 54 we get cost 07 421 459 9 98 90 31 49 135 56 0 180 56 0 225 560 270 455 315 56 0 At present only the best three values are retained so we will select W 45 90 and 0 This pruning introduces some risk that the program will fail to find an acceptable strategy in some cases where it might otherwise have won If this problem should become signifi
128. ions If the initial hole and pin axes are anti parallel the modifications required are obvious IV 32 Opposite ends of the pin are inserted in the initial and final holes Le we must turn over the pin while transporting it from hole to hole In the first case the lower bound on d will be given by de dmin max d d T fp where radius of finger tips small extra clearance factor currently 0 1 inch To compute the upper bound dmax we must consider the pin geometry if the pin has a painted tip then we must grasp further down the shaft d 5 dmax where length of pin taper length of point on pin tip If the interval dpray dmin is relatively short currently less than 2 5 inches then we just pick the midpoint d e d min dmaxJ 2 Otherwise a succession of values must be considered Currently three values are considered one near the top of the pin one near the bottom and one the middle d 0 6 inches 42 dmin 0 6 inches ds dmin 4max 2 For each value of d the system must generate values Similarly three approach directions considered 7 180 degrees 1 e anti parallel to the pin axis 7 135 degrees 7 100 degrees 1 approximately perpendicular to the axis With Y 180 degrees it is necessary to check that the pin doesn t poke up through the palm of the hand This is ea
129. ions The rude awakening came with the transition to real data The work reported in this dissertation has been largely concerned with representation of planning knowledge about real world situations and then using it to make rather more local coding decisions Although it is certainly possible to put up asystem which plans each task oriented operation independently of the others interactions must be considered if really efficient programs are to be produced In Chapter 4 we saw that when selecting a grasping point to pick up the pin we had to consider both the initial and final positions of the pin The estimated motion time included both the time for the hand to reach the pin and the time for the pin to move to the hole Also we discovered that some grasping strategies gave larger search patterns than others All these factors affected our final choice This sort of interaction 1s not confined to choices made within individual assembly operations For instance suppose we must place pins into two holes in our favorite box Then in selecting a grasping method for the first pin we should remember that our choice will also affect how much time will be required to pick up the second pin Other interactions may be more subtle Inserting the first pin gives us information about the box location Since this information can be used to reduce the search required for the second insertion we perhaps ought to consider the accuracy associated with
130. is way For example to set the value of X to 1 1 3 one would tell ALAID EXECUTE SCODE X VECTOR 1 1 3 _ 11 36 SIGNAL event variable WAIT event variable These commands are not strictly speaking either control or data modification commands but have some of the flavor of both Their intent is to allow processes to proceed from event waits by explicitly supplying the signal and to deactivate the ALAID portal until the program supplies a signal These facilities allow feedback routines residing on the PDP 10 to communicate with the program on the PDP I 1 When the AL program wishes feedback it signals a particular need feedback event and waits for the feedback ready event The cooperating routine waits for the need feedback event computes the desired quantities and feeds them into the program by means of ALAID commands and then signals the feedback ready event thereby allowing the program to proceed Section 5 Advanced Coin mands This section describes some miscellaneous powerful features of ALAID that do not easily fit into the pattern used to describe the other commands 1 CONVERSION CONVERT new mode string Direct conversion of typeout modes is possible by means of this command The string should be prefaced with the mode that it carries Not only does this facility allow direct symbol table lookup but it also allows temporary modes to override the permanent modes
131. ise or simple fixture Appendix 2 illustrates a typical error calculation for a screw on the end of a driver This is just pin in hole with a twist at the end IV 53 The important characteristics of these tasks are that they can be performed with relatively simple motion sequences and straightforward verification tests that the accuracy rquirements are fairly easy to state and that the coding decisions all rely on fairly locad properties Where these characteristics are not present automatic coding will be much harder Assembly tasks requiring clever uses of force working in cluttered environments and handling limp objects are typical difficult tasks It should be pointed out that humans don t know much about programming such operations either Since it is very difficult to automate coding decisions which cannot be clearly identified these tasks must be much better understood before much success can be expected Planning Coherent Strategies My early research on automatic manipulator programming was primarily concerned with the problem of how to write coherent programs which took account of interactions between individual coding decisions This work was done at a somewhat symbolic level typical decisions were selection of the order in which operations were to be performed selection of good workpiece positions etc It proved fairly easy to get a system to make these decisions inatoy world of symbolic assert
132. ith which the object will be held by the hand once the grasping operation 1 successfully completed departure point Via point through which pickup and move operation must pass goodness Measure of the cost of this pickup strategy Typically an estimate of the amount of time required Destination Points Destination points in AL motion statements really involve two components 1 frame valued expression specifying some location in the work station IV 27 Figure 4 1 Error Footprint 2 controllable frame variable whose value is to be made to coincide with the target value Thus move to b is in some sense a manipulatory equivalent to the assignment statement ae b For our present purposes it will be sufficient to restrict the right hand side component of all destination points to the form lt object or feature namerecconstant trans expression Thus the data associated with each destination point consists of whrt The ob ject or feature which is to furnish the controllable frame base Object or feature for target expression xf Constant trans for target expression IV 28 Section 4 7 describes how sequences of such destination points may be turned into motion statements 4 2 Initial Coinputations The most important initial computations are those responsible for calculating the expected initial positions and error determinations of the pin and hole Following the me
133. ithm in a suitable programming language formalization 4 submitting the program to the scrutiny of the computer compilation 5 running the program execution 6 getting the program to do what was intended debugging 7 making sure that the program behaves under diverse conditions testing and 8 production runs of the finished prdgram bliss These steps are not necessarily distinct it often happens that conceptualization design and formalization are performed simultaneously and the stages from formalization through debugging are often repeated several times The problem of successfully traversing this route 1 compounded in the particular case of arm code by several factors The first obstacle is that the real world 15 less tractable than the highly controlled world of the computer Any given strategy to accomplish a given task may fail because the actual real world result of the program may not be what the programmer desired An effort to insert a pin in a hole may result in a jammed pin or a jarred workpiece Not only is the world recalcitrant it also 1 complex A programmer in a purely algebraic language can attempt to keep in mind the various states of his program at different places State information like loop invariants provides enough environment so that reasonable code may be written Not so in the realm of mechanical manipulation Objects can be modelled but only partially and the extent to which models reflect the
134. ition of the pin to the hand 2 The accuracy with which the pin is held with respect to the hand So tong as this information 1 available the remaining decisions can proceed more or less in ignorance of the actual technique used 4 This may not be strictly necessary If the pin is to be part of a finished assembly then the direction and df may be obtained from the description of the object being assembled Alternatively it might be possible to tell which end to use by looking at the pin and hole diameters or to keep a data base telling what the standard direction for each pin type 5 Whitney 18 has done an exhaustive analysis of some of the factors required to compute tolerance requirements for insertion of a peg into a hole IV 30 4 3 1 Assumptions The grasping method described in this section assumes that the pin initially sits in a hole and that the hand is empty The basic strategy 1s to open the fingers move the hand to the grasping position and center the hand on the pin Thus we require that there be at least one grasping position reachable by the manipulator and that the pin s position be known with sufficient accuracy for the centering operation to succeed Once the pin is grasped it is extracted from the hole We assume that the pin will be free of obstructions once its tip has cleared the plane of the hole by some fixed amount currently 1 inch 4 3 2 Grasping Position The key element in our grasping str
135. jiffies 19 seconds of assembly time Eliminating these items would have resulted in a time saving of 21 5 seconds or roughly 20 Spiral searching for the hole was not considered since the arm performed this operation only some of the time VALIDITY OF MOTION PRIMITIVES FOR THE MECHANICAL ARM The analysis done has tried to model mechanical arm motion primitives in the light of motion primitives known for the human arm enabling a direct comparison of the two It 18 apparent that the process of programming the manipulator to do the assembly the way a human being does required special techniques in positioning and force and touch sensing which the human operator takes for granted The human operator makes use of vision which enables him not only to precisely locate the parts but also to avoid obstacles and perform smooth and precise motions Having more fingers and additional degrees of freedom over the mechanical manipulator enables the human operator to perform motions without having to go through the contortions the manipulator does For instance the mechanical arm has to turn through 90 degrees when the hand touches the top and sides of the main fixture to maintain the fingers parallel to the surface being touched since otberwise the spindle would tilt 1n the hand in the plane of the fingers and lose its orientation 4 V 20 COMPARISON OF MOTION PRIMITIVES WITH THOSE OF DRAPER ANALYSIS The similarities between the assembly primitive
136. ks pull out a skeleton program structure and then fill in the appropriate slots To some extent the program is thus composed of higher level chunks with the programmer acting in a dual role as a problem solver and coder translating between the conceptual units in which the program was composed and those made available by the programming system Planning information is used at both levels For instance the fact a 1 an unsorted array or that the loop sets maxel to the maximum element of afln would be typical high level facts useful primarily in performing the problem solving function Coding information includes the fact that is available for use as an index variable that is the name of the array to be searched etc Several people have commented that the loop should be written maxelea 1 for 2 step l until ndo Itis interesting to note that this form is equivalent only if nz1 In other words we can make a marginal improvement in program performance if we have an additional piece of planning information Program bugs happen when some precondition for using the trick is forgotten E g i might be in use for some other purpose It 1s not necessary to accept the psychological validity this paragraph in order to appreciate the main point that much coding can be done by adaptation of standard skeletons to fit particular situations 1V 18 3 2 2 Getting the Computer Involved A dominant theme in the
137. l Problem We now proceed to add two more edge to surface contacts IV 61 contacts bxbtm 61 56 inside of contacts be9 bj1 52 extent irrelevant contacts 610 811 53 extent irrelevant contacts 8611 811 54 extent irrelevant and wind up with the final estimate YHRTsRe 1 881 VECTOR 787 787 000 707 1 00 VECTOR 787 787 800 lt 707 XHRTeR amp 1 802 VECTOR 707 787 880 lt 707 YHRTsRe 5 85 VECTOR 760 649 808 lt 5 008 PV XHRTsRe 5 85 VECTOR 760 649 000 lt 4 888 XHRT PV YHRTsRs 5 850 VECTOR 760 649 800 5 988 YHRT PV XHAT Re 5 854 VECTOR 760 649 000 5 4 899 KHAT PV YHAT R 5 85 VECTOR 768 649 800 lt 5 888 YHRT PV XHRTsR amp 5 85 VECTOR 768 649 800 lt 4 888 XHAT PV YHRTsR 5 85 VECTOR 760 649 000 lt 5 800 YHRT PV XHRTeR amp 5 85 VECTOR 760 649 000 4 888 PV 5 854 VECTOR 760 649 808 lt 5 000 YHRT PV XHRTsRe 5 852 VECTOR 768 649 800 lt 4 888 XHRT PV YHRTsR amp 5 85 VECTOR 760 649 000 6 888 YHRT PV XHRTsRs 5 85 760 649 800 lt 4 888 XHRT PV 5 85 VECTORL 768 649 000 5 888 YHAT PV XHRTsR amp 5 854 VECTOR 760 649 000 S 4 889 XHRT PV YHRTeR amp 5 855 VECTOR 760 649 808 5 888 YHRT PV XHRTaR amp 5 854 VECTOR
138. lable so that all code and data structures can be examined as they appear in the source language and modifications can be made in the source language formalism Modules of acceptable code are joined together into larger modules and eventually a working program is prepared all under the unified control of the debugger In order to increase the investigatory power of the debugger as many data structures as possible are available to the scrutiny of the source language program and the debugger has access to all the constructs of the source language Structures necessary to the implementation of such a debugger include special purpose symbol tables to keep track of the correspondences between the source code and the object code multi machine links and debugging processes that act in parallel to the processes they are manipulating The entire programming system that uses ALAID to cement it together and to act as a supervisory program can be extended to include computationally expensive sensory feedback by direct communication between the feedback processes and the running programs through the links In addition this unified structure allows simple programs to be written in AL that mimic several different modes of manipulator programming from tape recorder mode in which positions are manually set and a program is written to repeat those positions to completely textual modes in which all positions and uses of feedback are specified in
139. lable on the machine used 4 Debugging should consist of symbolic examination and modification of data program and control flow 5 The debugger should be usable as a top level command structure for the system composed of the compiler the runtime and the debugging package 6 Insofar as possible all structures of AL code should be available for examination and modification under formalisms present in the source language The purpose of this chapter is to discuss the structures needed to implement such a debugger This treatment is heavily based on the nature of the current AL implementation at use at Stanford both its software and its hardware The ideas however are generally applicable to any AL implementation and in part to any programming language implementation The state of ALAID at the moment is fairly primitive it resides on the two machines and can start up the interpreter examine and set arithmetic variables signal wait events and cause the runtime system to enter 1 1 DDT These initial facilities alone make the current A LAID implementation quite useful for testing out new runtime routines AL programs and high level feedback that requires the PDP 10 Many of the concepts introduced here are therefore ideas for extensions to ALAID and design strategies for accomplishing their implementation IL17 Section 1 The Link Between Machines ALAID is intended for the interactive debugging of a program that has been compile
140. lized tools V 6 Figure 4 Placement of Locators 1 D wet 2 m 5 p b or A pooole i E B i 5 E 9 USE H p pah 4e m m 59s t x p P y sab E 6 y Ie 1 oe x A met os he LE ess 9 L SACS L tta wem CN a 2 m e r ote om v Sai A m rape o 9 e 9 ps D 5 m z D E an o kume gt v 955 99909 0 dip vl 2 659 e 9 eye GE 9 00 modb LJ at Y a e P s amp E 5 Nw be t a m vo n ee s shee m Pd gt shee 3 e D vme 9 d ve 2 Figure 5 Fixture for Pencil Sharpener V 8 Figure 6 Fixture for Pencil Sharpener Shell V 9 Figure Section Through Main Fixture V 10 Overmachining of the plaster resulting in too much iossage could be easily corrected by the application of more wet plaster which was later allowed to dry The biggest disadvantage of using plaster as
141. lly if we desire to rotate the coordinate system along which the forces and moments are resolved we can again pre multiply the calibration matrix by an appropriate 6 x 6 matrix Assuming that the rotation is represented by a 3 x 3 matrix R which defines the rotation from the calibration to the new coordinate system the total transformation from strain gage readings to the desired force vector will be given by Fe R xDxCxe where 5 Current State of Force Sensing System The calibration method that has been described in the preceding sections has been used to calibrate a force sensing wrist with an attached hand that was not as yet mounted on a manipulator From these initial tests it appears that the calibration method works quite well We were able to compute a calibration matrix that could accurately resolve subsequent forces and moments to within approximately 1 At present we are awaiting the mounting of our force sensing wrist on one of our Stanford Arms In anticipation of this event software has been written which can be used to calibrate 111 12 the wrist automatically In addition the software now exists to compute forces from the strain gage readings given the calibration matrix and to resolve these forces at a point separated from the calibration center by a linear transformation Bibliography 1 Richard Modelling Trajectory Calculation and Servoing of Computer Controlled Stanford
142. lly to return an answer If the message is an answer then it is made available to whatever process sent the corresponding request The correspondence between process and request message number 15 kept in a list 1X 21 that is modified whenever a request is sent or an answer is received across the link If the message is a tidbit then its contents are directed to a tidbit handler The usual response to a tidbit is to print it for the user to see process that has encountered a breakpoint makes its plight known by passing tidbits Just as there is only one server on each machine there is likewise only one requester Sending a rquest involves adding the name of the requesting process to the waiting list and sending notes across the link to agree on the transfer of the message The process that asked for the transfer is suspended until an answer arrives 7 FLOW OF INFORMATION To demonstrate the manner in which information is passed among the various pieces of ALAID consider the user request to place a breakpoint at a particular point in the code Assume that the user 15 communicating directly with the PDP IO and that he uses the name of a program label to identify the spot The request looks like this BREAK SADORESS L1 Now the PDP 10 cannot itself insert breakpoints so it passes the entire request to the PDP 11 The PDP 11 cannot interpret the label since it has no symbol table Therefore it must ask the PDP 10 to identify the loca
143. locations they represent then the program will not work correctly It is worthwhile to consider what can be done to reduce the accuracy required since extreme precision may be rather expensive and difficult to attain 2 3 6 Error Detection Missing the Hole Earlier we noted that a simple close statement could overstrain the arm if the pin rack were bolted down off center A similar difficulty can arise if the box is displaced from where its location variable says itis lf the error is big enough then the pin tip will hit the top surface of the box rather thah go into the hole Here we cannot just add a simple compliance clause to the motion statement and expect things to work We can however detect failure by monitoring force and stopping if amp collision is detected in hole flag true Assume it will work move pin to in Aole position via pin_withdraw hole_ap proach on force pinszhat gt 8 oz do begin stop Stop the motion in_hole_flagefalse We lost end The force threshold of eight ounces is rather arbitrary a certain amount of tuning may be required to get the best value Post insertion Checks This code assumes that successful pin insertion occurs if and only if the pin doesn t hit the top of the box However if the box is displaced far enough the pin may miss it entirely Since the force threshold isn t exceeded in that case the fingers will open dropping the pin on the floor One way to
144. low it When the pin hits the surface the motion is stopped and used to compute a correction com for use in the success test 1V 52 Chapter 5 CONCLUSIONS AND FUTURE WORK The goal of this research was the generation of AL manipulator control programs from high level task descriptions The full topic of automatic generation of AL code is extremely broad and many narrowing assumptions have been necessary in order for us to demonstrate basic feasibility while keeping the scope of effort within reasonable bounds This report has explained how AL programs have been generated automatically for a particular programming example the insertion of a pin into a hole which is a typical subtask of many assembly operations The example was first discussed from the point of view of a programmer coding directly in AL to show that the task is non trivial if attention is given to making the code rugged with respect to positioning errors Next the modeling requirements for automatic coding were analyzed since the automation of coding decisions requires that the necessary information be represented in a form usable by the computer Finally the programming example was revisited this time with the computer using its planning model to generate the AL code automatically Extensions Although the pin in hole task was used as an example throughout this work a conscious effort was made to avoid undue specialization The modelling requirements f
145. lso pose problems for modifying their information According to the philosophy of representing constructs in source formalism wherever possible the way for the user to enter a correction to these structures would be to restate his source code and let the debugger regenerate the proper forms In this sense the debugger should have the full power of the compiler 11 19 Yet another suggestion with regard to language design can be made in regard to these complex data structures In order to easily point the debugger at any object it is most convenient to somehow labelthat object Variables automatically have names that is why it is so easy to refer to them during debugging We have already seen that the fact that processes do not have explicit names Causes problems in pointing the debugger at a particular process The same consideration carries over to such cumbersome structures as motion tables not to mention statements A reasonable suggestion is to associate variables of the appropriate type with each of the control and data structures used by the AL language Not only would this association facilitate debugging but it would also render the concepts represented by the structures more flexible For example if motions are values that can be named by variables it becomes natural to consider composition and extraction operators that can act on this datatype Even some arithmetic operators may not be out of the question perhaps a scal
146. lution also lends itself to the problem of 11 28 finding the pseudo code location of a given source language statement Section 3 Control Over AL Various programs must be applied in order to achieve execution of an AL program One of the purposes of ALAID is to control the compilation loading and execution process so that a unified face is presented to the user The ideas in this section lay a groundwork for such a facility these concepts have not been implemented in the current version of ALAID The primary unit of compilation is the module It is one statement long and is self contained In general the statement 15 a substantial program but it can be very short as well To refer to variables that are not in that module a global declaration is given The planning values for all global variables starts as undefined so assertions are necessary before these variables can be used The output of a compilation is a load module that has symbol table linking and planning model information in the form of decorated parse trees A linking loader is invoked to take this load module and insert it into the current runtime system This loader is one of the resident programs on the PDP 10 symbol tables on the PDP IO are referenced and modified during the loading process Direct memory access on the PDP 11 is used to actually put the program in place One useful concept is unloading which takes the current set of modules on the P
147. ly investigate the information can be serviced from any of the sharing members An alternate technique is to allow any of the sharing members to make modifications but to force them to send tidbits whenever so doing The disadvantage of this strategy 15 that it is not possible to be sure of getting the most recent value of such information because one of the members may have changed it without yet informing the rest of the community In the first approach a good value can always be generated by asking the owner specifically Tidbits advising that multiply copied information must be updated could be broadcast to the entire community of members or each type of duplicated information could include pointers linking it to the various members who have copies 8 GENERALIZATIONS OF THE LINK The idea of connecting two processors together to share in the work of debugging has some obvious generalizations The link that has been described 15 designed to share information structures in a domain that naturally divides labor between two processors Cases in which more than two processors are in use are becoming more frequent the ARPA network of computers presents an environment in which cooperation among many machines may be used in the execution of a single algorithm Furthermore the decreasing cost of processors seems to be creating a trend to divide complex algorithms among several perhaps many machines The natural question 15 how well the two ma
148. mber already on the history list If a query generates a subquery that is somehow equivalent then circularity is still possible because the subquery does not share the history list of its parent but can perhaps spawn another subquery that has exactly the same form as the original one The only way to avoid this problem is to disallow equivalent subqueries or to give them the same history list as the original queries Some frequently used information might be duplicated in several members In this case any one that receives a query that requires that information can service it long as the information 15 static there 1s no danger that inconsistencies will arise between the several repositories If on the other hand the information is subject to fluctuation for example the 111 23 current context then some must be developed to keep the various versions as consistent as possible Tidbits can be sent to all interested parties to inform them of a change in their data base the issues reside in who should initiate these tidbits and to whom they should be sent If every piece of shared data has an owner then only that owner should be allowed to make a change to the data and when he does a tidbit should be sent to the other members to inform them of the change That means that a request whose effect 1s to change information that has duplicate copies must be directed to the member that owns that information queries that on
149. med and to relate this information to the physical reality being modelled It does little good to know that once we have closed the fingers on the pin it will move when the hand does unless that information is reflected in the program by a corresponding affix statement 3 6 Action Information Clearly it 1 necessary to understand the semantics of the manipulator language in order to write programs in it5 This is essential both for translating desired manipulator actions into the corresponding code and for keeping track of situational information Earlier we described the coding component of programming as adapting previously defined code skeletons to fit particular facts This occurs in manipulator programming to a surprising extent For instance the grasping sequence of our pin in hole example is readily adapted to pick up more or less arbitrary objects 5 Of course this knowledge does not have to be perfect There are those whose approach to programming is empirical to say the least Even where a certain amount of experimentation is attempted however one generally requires at least an approximate model of what a particular statement 1s supposed to mean 17 22 open blue to initial opening move blue to object grasp via object graspstrans nilrotn vector 0 0 4 inches center blue on opening lt minimum opening do begin stop abort It just isn t there end move object to object pickup point The
150. moved to sibling thread either forwards gt 0 or backwards lt 0 The last code is DOWN which moves down one level only to the nth daughter thread An abbreviation for DOWN n DOWN is DOWN n SET MODE lt mode specifier gt The typeout mode is set to the one given in the command Section 2 Data Structures 1 TYPEOVT MODES All arithmetic quantities are displayed according to their type which is built into the runtime data structure That is vectors will always be typed as three numbers However there is some flexibility in typing rotations and therefore frames and transforms which have rotation components One mode ROT mode reduces the rotation to one swivel about one axis and reports the rotation the same way the source language accepts them ROT vector angle The other mode EULER mode reduces the rotation to up to three rotations about cardinal axes This mode 1 far easier for the human to understand 11 53 Non arithmetic quantities include expressions and events Expressions are printed as code the relevant modes are PADDRESS PCODE SADDRESS and SCODE as described above Variables can be named as they are called in the source language SNAME mode or as they are translated for the pseudo code PNAME mode 2 EXAMINATION SHOW VALUE variable name The variable must be available in the current context The name can either be in SNAME or in PNAME modes
151. multiple processes residing on several machines presents some unique problems for the design and implementation of a debugger The purpose of this chapter is to discuss debugging issues raised by manipulator programming in general and by the AL language in particular Some conclusions will be reached not only for the design of a debugging package for AL but also for the implementation of AL so that it might be easier to debug These conclusions can be generalized beyond the context of debugging to the larger issue of preparation of workable code in general and arm code in particular Section 1 Block Structure AL 18 a block structured language In most ways it conforms to the scope rules standard in such languages As a block is entered all variables for that block are declared and room made for them in the environment of the current process These variables include special ones for condition monitors force feedback events and calculators As control exits from a block each of these variables is released and its space reclaimed An attempt is made to sever any connection between these variables and the state of the computation outside the block variables and expressions are unlinked from any graph structural relations condition monitors are awakened to tell them to disappear and events are returned to the kernel which awakens any process waiting on them with a failure indication Due to details of implementation block structure is violated
152. n by deriving quations for the sine and cosine of 9 in terms of 0 0 and S Combining quations 22 amp 23 and 22 amp 24 amp 25 we obtain s y T T 440 c0 T 4 50 40 93604 26 The sine quation reflects the fact that 0 can be arbitrarily selected to be positive 11 5 negative Since the last three joints have intersecting axes any two sets of Joint angles 9 0 0 and 8 180 0 0 180 are equivalent and in fact occupy the same volume in space If joints 4 or 6 have less than 360 degrees of rotational freedom the duplicate solutions can be used to minimize the loss in orientation capability Otherwise the choice among the solutions can be made on the basis of producing the minimum total change in joint angles If 0 is equal to 0 180 degrees then either 0 or 0 can be selected arbitrarily and the remaining angle of rotation must satisfy the transformation equations As the full range of 9 for the Stanford Arm is 110 110 we need only concern ourselves with the case of 0 equal to zero Equations 10 and 11 yield the desired equations for this degenerate case T sin 0 8 cos 0 9 z 27 If the configuration is not degenerate than we can use the following expressions for 0 which can be derived from equations 22 24 and 25 T 3 0 9380 T4389 458 s 4 s 50 25 In order to
153. n accomplished as evidenced by WO REL RENT a s i uen TR y T V 11 not encountering reaction at the end of the spindle the spindle was pushed down and twisted at the same time in order to seat it without binding It was then twisted into the handle hole by applying a downward force and turning the hand through 360 degree revolutions The shell was lifted vertically out of its fixture and positioned over the assembled spindle then lowered in place and released at a height of 0 2 inches from the mating surface regrasped and wobbled gently while being pressed down in order to ensure seating It was then turned 45 degrees to finish the final assembly stopping when a resisting moment was encountered COMPARISON OF ASSEMBLY BY MANIPULATOR VERSUS ASSEMBLY BY HUMAN TIME DATA The compiled program of about 20k bytes was able to assemble the pencil sharpener in 2 4 minutes Of this time about 1 8 minutes or 108 seconds was actual CPU time mainly for servoing The rest was overhead due to interprocessor communication and loss of the processor in time sharing mode A theoretical estimate of the time taken for the arm to perfotm the assembly also gave 108 seconds assuming continuous motion and disregarding lossage of the processor and overhead Detailed analysis of the assembly procedure by the manipulator is given Appendix 2 and it should be noted that much time 15 spent in opening and closing the hand ce
154. n that requires the minimum number of special purpose attachments to the arm With this in mind the following calibration procedure was devised 1 Forward and Reverse Calibration M atrices our force sensing wrist a total of eight pairs of strain gages must be sampled in order to resolve the three components of force and the three components of moment applied to the wrist If the assumptions of superposition and perfect elasticity are made any one of the components of force or moment would in theory be only a function of two or possibly four of the strain gage readings In fact we found that it was necessary to consider each component of force to be a function of all eight strain gages in order to achieve better than 1 accuracy If we let Fy FF My MyM represent our force vector and i 1 to 8 the eight strain gage readings calculating the force vector from the strain gage reading can be accomplished by the following matrix operation 33 F Fy Fy Fz My My om 1 612 613 614 615 616 617 618 621 622 623 624 625 626 627 628 C a c31 632 633 634 635 636 637 638 641 C42 C43 C44 45 C46 647 648 C61 652 653 654 655 656 657 658 962 63 64 665 66 67 668 The objective of calibrating the force sensing wrist is to compute the j matrix elements based upon experimental information To do this we will first compute the elements of the 11 81 pseudo invers
155. n to hole rigidly t trans nilrotn vector 0 0 affix Aole approach to hole at trans nilrotn vector 0 0 Iscm affix hole to box at trans nilrotn vector 5scm 4 cm 3scm initial box position Grasp the pin open bfingers to 0sinches move blue to pin grasp via pin grasp approach cen ter bf ingers on opening lt O Isinches do begin Stop abort Grasp failed to pick up pin end affix pin to blue Extract transport and insert move pin to in Aole position vector 0 0O 3 inches via pin withdrau hole approach on force pinszhat oz do stop distance off zhat inv in_hole_position pinsvector 0 0 0 if not 0 2xinches distanceoff gt J2sirlches then begin vector dp scalar n boolean in Aole flag dp vector 0 Jeinches 0 0 in_hole_flag false while 1 6 and not in hole flag do begin dp rot zhat 60sdeg dp Try to put pin in perturbed hole move pin to in_hole_position dp vector 0 0 1 sinches via Aole ap proach dp vector 0 0 I inches on force pinszhat gt 802 do stop t Check distance travelled etc distance off that inv in_hole_position spinsvector 0 0 0 if 0 2 inches gt distance off gt 0 2sinches then in_hole_flagetrue end if not n hole flag then abort The hole doesn t seem to be there end IV 15 Let go of the pin open bfingers to 0sinches unfix pin
156. nce could be worse This imprecision of the force application and measurement caused problems where these should not have occurred Firstly low contact forces of the order of 2 or 3 oz were dominated by the noise force Secondly the manipulator tended to apply more force than necessary or specified especially in the sideways direction when trying to locate the hole position by touching the sides of the fixture and at times caused a slight movement or tilt of the spindle in the hand that resulted in difficulty later when insertion of the spindle into the hole was attempted The magnitude of force applied in the downward direction did not matter so long as buckling did not occur or the spindle did not tilt since the reaction of the table prevented any movement in the vertical direction Draper 4 has shown that such forces are important to the extent that jamming occurs and this is discussed further below FURTHER ANALYSIS OF THE SPINDLE IN HOLE INSERTION PROCESS Insertion of the spindle into the hole was an example of the pin in hole problem studied 4 intensively at Draper with the parameters being as follows d shaft diameter 0 40 inches D d clearance 0 004 inches l insertion depth 1 06 inches at full insertion ECC CUSCO UE v Soit V 22 C clearance ratio 1 d D 0 01 20 j minimum wobble at full insertion 2C 0 430 deg j wobble from center line 0 2 15 deg To allow initial entry into
157. nce of information particular to the problems of manipulation This information may be divided roughly into the following categories l Descriptive information about the objects being manipulated 2 Situational information about the execution time environment 3 Action information defining the task and semantics of the manipulator language Subsequent sections discuss these issues in greater detail 3 4 Object Models Programs which specify explicitly what actions are to be performed by the manipulator generally need contain little explicit description of the objects being manipulated In the AL program developed in Section 2 3 for instance there is no information about the shape of the pin hole or anything else The principal language construct for describing objects is the affix statement which is used to specify how the location of an object is related to the location of its subparts or features For instance affix Aole to box at trans rot xhat 20sdeg vector 2 4 1 3 3 2 On the other hand a great many assumptions about the objects have been built into the program For instance the check used to verify that the pin has been grasped successfully relies on knowledge of the pin diameter the extraction grasping and insertion positions implicitly assert that the hand or pin will not crash into anything the insertion strategy assumes that the pin will accommodate to the hole somewhat that misses will cause the pin to hit a sur
158. ngle PCODE and a jump to the next SCODE entry PADDRESS and SCODE is interpreted to mean that SCODE at that PADDRESS 15 to be changed from the beginning even thought the PADDRESS may be in the middle SADDRESS and SCODE is hard because the new code might not fit in the old location The newly compiled code is therefore placed in a fresh location and appropriate jump instructions are inserted to patch it in Section 4 ControlFlow 1 TYPEOUT MODES Locations in the control store can be referred to either by octal location in the PDP 11 PADDRESS m d by labels and offsets the source code SADDRESS mode The contents of the control store can be shown either as pseudo instructions PCODE mode or by the source code that generated them SCODE mode An extension to this facility would be tb allow MC ODE and MADDRESS to refer to machine instructions If the location is a pseudo instruction in the middle of several that all accomplish the same statement then the SADDRESS and SCODE outputs will be annotated in progress 2 EXAMINATION SHOW EXECUTION interpreters in the given context are listed with the name of the statement currently in execution The statement is listed in pseudo code and its address is given 3 MODIFICATION BREAK address A breakpoint is inserted at the given address When execution encounters this point a message will be sent to the user and control will pause until the us
159. nnot be reversed automatically since there 15 no way to determine what forces ought to be applied during the backward motion they are not closely related to the forces applied during the initial insertion In many programming languages the debugging loop is exceptionally long AL is no exception In order to fix a known error it 1 necessary to modify the source code and then to resubmit the program to the compiler The compiler which is fairly slow produces an intermediate output which is finally loaded into the PDP I 1 And then the pieces must be reinitialized to their starting positions and the race must be run up to the point where the failure occurred in order to test the fix The failure may not be very reproducible so the new code may not be easily tested The effort that must be exerted to locate a bug in the first place can be immense Making some small change and resubmitting the program in the hope that the change will make the bug more traceable 15 very tedious due to the long turn around There are several debuggers for the PDP 1 1 machine one typical example is 1 1 DDT based on RUG another similar debugger and implemented by Jeff Rubin at Stanford Binford 75 DDT is a symbolic interactive debugger but it has no knowledge of AL its microscopic vision cannot see the forests of manipulator code built out of the trees of machine instructions The AL runtime environment has been implemented and debugged with the assistance of
160. ntering over the object and trying to verify the position of the hole Estimates of a human operator using one hand on the basis of MTM data to do the same assembly showed that it took 14 seconds verified by the author taking 15 seconds to do the Job a factor of 8 times faster than the manipulator It must be remembered that the manipulator did not utilize vision for help as a human operater does and was thus akin to a blindfolded one armed two fingered human operator doing the job A table showing the analysis of the MTM study for the human operator is given in Appendix 1 It should be noted that the assembly procedure does not represent the optimum sequence of movements or placement of the component parts initially The determination and elimination of inefficiencies would mean running the system at the limit of its capability which would result reduced assembly time ASSEMBLY AND MOTION PRIMITIVES MTM and Draper use assembly primitives in studying the sequence of tasks involved in putting an assembly together while WAVE uses motion primitives to specify arm motion The former primitives are descriptive in nature since they describe the actions performed V 12 but not the motion for the manipulator to achieve the action WAVE primitives are strategic since they specify where and how the arm has to move rather than what the assembly task is MTM ASSEMBLY PRIMITIVES 2221 MTM primitives generally consist of several
161. o have access to all the state of the machine that defines the world at any point in the execution and to be able to modify it First the point of execution must be defined in the presence of parallelism it refers to a point along each of the many threads that may be active That 1s the context in which execution 1s understood is the most general context of all the outermost block This definition is more restrictive than is strictly necessary since backing up may only be needed with respect to one or several threads that 1s within a more specific context However there is no guarantee that a given context will contain any active processes unless that context 1s the most general one Next a debugger that deals with parallel processes must be able to observe and manipulate all synchronization control between them If one thread has produced a signal that another thread will eventually await and a bug strikes it must be possible to back up past the signalling of the event This consideration forces the debugger to keep track of what events are signalled and awaited If the event has been successfully awaited by another process then to back up the first requires that the second one be backed up at least to the point at which the event was awaited This consideration forces the debugger to hold a context large enough to include all processes that use any important events An alternative that has shown some success when signals and waits are pai
162. od is the space occupied the pseudo code for the marks which might amount to about 25 percent of the total code not counting trajectory files and constants If these are also taken into account then the expetise of the marks 1s only about five percent A related technique is for the compiler to emit a separate symbol table that coordinates pseudo code addresses with source language statements Such a table would be searched by a binary chop method About the same space requirements would be necessary in this case the advantage is that the symbol table is separated from the objects it is describing and it can therefore be moved to the other machine The trouble with this method is that during compilation no information is maintained about the location where the code will be placed and furthermore it is hoped that AL will soon become capable of compiling relocatable modules The implication of this observation 1s that the symbol table must be manipulated by the loader to convert relocatable addresses into actual addresses The best solution is to combine these two approaches The relocatable output should contain marks that relate the code to source statements and the loader should remove these marks constructing a symbol table in the process In this way the binding of the symbol table takes place at the time that all necessary information 1s available and at that point extraneous information the marks can be discarded This so
163. of 0 0 3 Solution for 0 0 8 There are primarily two forma of the following equations that can be used to solve for the last three joint angles The two sets of quations differ in how the degenerate condition for the last joints must be treated The arm configuration is called degenerate whenever is equal to 0 or 180 degrees At these times the axes of rotation for joints 4 and 6 are collinear and only the sum of 0 and 0 can be treated as an independent variable One form of the solution equations for the last joints has the advantage of producing valid results whether or not the arm is in a degenerate position However these equations are slower to evaluate than the equations which require that degenerate configurations be treated as a special case For this reason only the latter form of the equations will be presented We will find it convenient to work with combinations of the equations for the third column of the transform matrix in deriving expressions for 6 and 0 Since this column indicates the direction of the Z axis of the hand all of its terms are independent of 0 From equations 4 8 12 we obtain the following expressions T1389 Toc 0 50 22 T9380 0 0 5010 0 50 0 50 23 T4369 T4450 50 50150 50 c0 c0 0 50 24 369 T44c0 50 50 cO cO 50 0 50 50 25 In order to distinguish between degenerate and non degenerate configurations we will begi
164. of this work I owe a special debt to Dave Grossman who had the patience to read these chapters several times and who made many valuable editorial improvements 55 Chapter 6 BIBLIOCRAPHY 1 Proceedings of an ACM Conference on Proving Assertions A bout Programs SIGPLA Notices January 1972 2 Association for Computing Machinery Proceedings of a Symposium on V ery High Levd Languages SICLAN Notices April 1974 3 David Barstow The PS Coding Expert A Knowledge Based Approach to A utomatic Coding Manuscript Submitted to Second International Conference on Automatic Coding October 1976 4 Bruce Baumgart GEOMED A Geometric Editor Stanford Artificial Intelligence Laboratory Memo AIM 232 Stanford Computer Science Report STAN CS 74 4 14 May 1974 5 T 0 Binford D D Grossman E Miyamoto Finkel E Shimano H Taylor C Bolles M D Roderick M S Mujtaba T A Exploratory Study of Computer Integrated Assembly Systems Prepared for the National Science Foundation Stanford Artificial Intelligence Laboratory Progress Report covering September 1974 to November 1975 6 T O Binford D D Grossman Liu Bolles R Finkel M S Mujtaba D Roderick B E Shimano R H Taylor R H Goldman J P Jarvis V Schdnman T A Gafford Exploratory Study of Computer Integrated Assembly Systems Prepared for the National Science Foundati
165. ok eight times longer to do the assembly job than the human operator did It should be emphasized that this study has indicated the presence of inefficiencies the present setup The quantitative determination and elimination of these inefficiencies the optimization of movements in itself another important research area and increased use of sensory feedback would bring about a reduction in the assembly time ACKNOWLEDGEMENTS The author would like to express his thanks to Richard Liu who suggested this particular assembly and for his valuable comments to Tom Binford whose valuable suggestions and advice provided new insights and to Dave Grossman for his editorial help in putting this paper together
166. om any frame to any other frame so that even small motions may require that large inertias have to be moved This fact suggests alternative arm designs with redundant degrees of freedom allowing small motions to be made with low inertia Some of the possibilities are described below a Extendible wrist which can elongate about 2 3 inches so that hand can move along the direction of approach without moving joints 1 2 or 3 b Extendible boom so that joint 4 can move out of the boom a distance of 1 2 inches without moving joints 1 2 or 3 c Independent finger movement so that to grasp something without moving it it would be sufficient to move only the fingers without having to use the CENTER command in which the whole arm has to move When necessary it would be possible to move both fingers together e g when the position of the hand is known precisely and it 15 desired to move the grasped object to the position defined by the location of the hand If redundant fine motions were provided in this fashion then one might consider providing detents for joints 1 through 3 so that these joints can stop only in a finite number of known positions say every 5 degrees or 1 degree apart which can be determined to a high degree of precision If there were no backlash and no static deflection of the arm components due to loading the use of stepper motors in joints I 2 and 3 would accomplish the same purpose The advantages of these changes w
167. ome queries is immediately available in the host machine that is the one with which the programmer is directly communicating At other tjmes the host computer will need to request information from or perform actions on data that is only available on the remote machine that is the one not currently discussing the knotty issues of bug control with the human guide In these cases the host will send a request to the remote machine in exactly the format that would be used by the human if he were working from that machine In fact there 15 no particular reason why several people cannot be simultaneously debugging from different portals each with his own debugging process and his own host machine A second feature of linking the compiler to the execution machine is that it provides mechanism whereby the entire execution can be controlled by a supervisory program residing on either machine Instead of going through the standard stages of writing code compiling it loading and then trying it all out using a different program for each stop on one machine or the other a unified command structure can control the entire program preparation endeavor A third happy result of the link is that complex forms of feedback for example visual feedback can be interfaced to running manipulator programs across this link The running program can wait for a picture to be taken and processed on the large machine and the results of this exercise can be tr
168. on Stanford Artificial Intelligence Laboratory Progress Report covering November 1975 to July 1976 7 Robert C Bolles Verification Vision Within a Programmable Assembly System Ph D Dissertation Summer 1976 8 Per Brinch Hansen Operating System Principles Prentice Hall Series in Automatic Computation Englewood Cliffs New Jersey 1973 9 0 J Dahl W Dijstra C A Hoare Structured Programming Academic Press New York 1972 10 Raphael Finkel Constructing and Debugging M anipulator Programs Ph D Dissertation Stanford Computer Science Department 1976 11 Robert Floyd Towards the Interactive Design of Correct Programs Stanford Computer Science Report STAN U 71 235 September 1971 12 Guiseppi Gini Gini and Marco Somalvico Emergency Recovery Intelligent Robotr Proceedings of the Fifth International Symposium on Industrial Robots September 1975 13 14 5 16 7 18 119 20 21 22 23 24 25 26 IV 56 Cordell Green et al Progress Report on Program Understanding Systems Stanford Artificial Intelligence Laboratory Memo AIM 240 Stanford Computer Science Report STAN CS 72 444 August 1974 David D Grossman and Russell H Taylor Interactive Generation of Object M odels With a Manipulator Stanford Artificial Intelligence Laboratory Memo AIM 274 Stanford Computer Science Report STAN CS 75 536 December 1975 Jame
169. onstants An advantage unique to execution time variables is the fact that values can be recomputed and saved when the program is run Thus our program will work correctly for many different initial box positions so long as the built in assumptions that the box is upright on the table in reach of the arm etc are not violated 9 Murphy 17 has investigated the reliability of systems in some detail Experience verified that his results apply with special force to manipulator programming Indeed one can write programs like the one developed in this section at one s desk The required location values can then be measured during initial setup For instance using a system like POINTY which is discussed in 14 6 There are number of tradeoffs involved in this mode of programming the principal advantage being the reduction of manipulator downtime while a new application is programmed and the principal disadvantage being the loss of immediate feedback while the program is being written 9 Actually the fact that AL preplans arm trajectories means that the underlying assumptions are rather more restrictive though still quite broad IV 9 In addition to the relatively broad assumptions about what the various runtime values are apt to be the program includes a number of much more restrictive assumptions about the accuracy of its runtime model If the values stored in the variables differ by even a small amount from the actual
170. or this task expected locations accuracies etc are applicable to other assembly operations and the techniques used to represent planning information were developed without any particular task in mind When time came to write the automatic coding procedures described in Chapter 4 no substantial changes to the modelling mechanisms were required although a certain amount of bug killing was necessary However it is worthwhile to consider how hard it would be to add automatic coding procedures for other tasks As one might expect the easiest additions would be for variations of pin in hole such as screw in hole for which most of the analysis has already been done The principal additional difficulty that a screw in hole writer must handle would be figuring out how to pick up a screwdriver and how to load a screw onto it Since these are fairly specialized operations it seems reasonable to construct a small library containing the appropriate code for different drivers and screw dispensers We would also want to consider alternative methods such as using the hand to start the screw into the hole and then driving it down Almost as easy would be the task of fitting a nut or washer over a stud although keeping the fingers out of the way would probably be more of a problem Only slightly harder would be mating operations such as fitting a cover plate or gasket over aligning pins and operations such as putting a part into a v
171. ory for communications two fixed blocks of memory called noteboxes are reserved at all times One 15 for the PDP 10 to send little notes to the PDP 1 I and the other 18 for the PDP 11 to send notes to the PDP 10 As a sign that the note has been received and to clear the notebox for further communication the receiving processor overwrites the first word of the note setting it to zero The traffic in notes provides for agreements on the use of the larger memory for more substantial messages all actual allocation is treated by PDP 11 which honors requests for message space both from its own processes and from foreign notes T here are very few note types required to maintain the link since the allocation of message buffers 1s the prime activity at this level Allocation is non symmetrically handled by the PDP 11 so the two processors send different kinds of notes to each other Each note 1 at most three words long the noteboxes occupy very little space The first word identifies the type of note and the second two words provide room for arguments to the requests and responses 2 NOTESFROM THE PDP IO TO THE PDP II These are the note types that the PDP 10 can place in the PDP 1 I s notebox note type CETBUF lt Allocate a message buffer s bytes long The expected response is the BUFALC note note type USEBUF a The buffer that starts at address a 1s a message Look at it act on it and then reclaim the message buffer no
172. ould be faster nulling out of small errors and higher spatial resolution for given A D resolution Disadvantages would be that the programming language might become more highly hardware dependent and there would be times when additional gross motions would be needed to bring the fine motors back nearer the centerpoints of their ranges 24 It is not known whether or not the advantages outweigh the disadvantages In any event we do not plan to modify our arm hardware in the foreseeable future CONCLUSION This paper has discussed some of the problems encountered assembling a pencil sharpener with the six degree of freedom Stanford Yellow Arm and comparison of the assembly motions required with those of the human operator using MTM and Draper assembly primitives While arm resolution was lower than the clearances involved in the assembly the use of suitably designed fixtures enabled parts to be located to a high degree of precision by just dropping the part and nudging it into place rather than actually trying to position it precisely Analysis showed that the human operator is faster and requires fewer operations for the assembly process than the mechanical arm since the human operator makes more effective use of far more sensory feedback information and the human arm is lighter and more flexible and the hand is more dexterous and has more fingers than the mechanical counterpart With these handicaps it was found that the manipulator to
173. p surface of a small metal box Initially the box body sits on the work table at Typ and 1s subject to displacement errors of up to 20 3 inches along the x axis of the table and up to 0 2 inches along the y axis and to rotation errors of up to 5 degrees about the table z axis The hole 541 is located at Tbh with respect to the box the pin pind is held in a tool rack at and the manipulator bmanip is parked at bprrk where Such an approach is a natural extension to the present translation performed by the AL compiler 17 47 Twp trans nilrotn vector 45 2 102 0 Tbh trans nilrotn vector 3 85 3 20 4 90 Twp trans rot zhat 20sdeg vector 24 1 117 537 y bpark s trans rot yhat 80sdeg 43 5 56 9 10 7 From the initial computation we determine that direction axes parallel de 1 71 cm d 0 Arg 0 762 cm As JO degrees In other words the pin is expected to go 1 71 cm into the hole When we make the attempt if the pin tip is within 0 762 cm of hole center and the axes are within 10 degrees of parallel then the insertion operation will succeed If we miss then we won t go any distance into the hole at all Le we won t get stuck halfway in The pickup strategy generator now goes to work and decides on single grasping distance and a range of grasp angles derasp 3 54 100 degrees s 7 lt 180 degrees It then produces nine feasible pickup str
174. parts the assembly primitive the distance involved and specific cases involved in the the assembly primitive For example R 24 D means to REACH 24 inches to an object in a fixed location or to an object in the other hand or to an object on which the other hand rests The following is a partial list of primitives their abbreviations and a description A fuller description of the specific cases is given in Appendix 1 REACH R Move hand to destination or general location MOVE M Transport an object to a destination TURN T E Turn the hand by a movement that rotates the hand wrist and forearm about the long axis of the forearm GRASP G Secure sufficient control of one or more objects with the hand POSITION P Align orient and engage one object with another when the motions are minor RELEASE LOAD RL Relinquish control of an object by the fingers or hand APPLY PRESSURE AP Apply force along the axis of the forearm TURN amp APPLY TURN and APPLY PRESSURE are tabulated PRESSURE T amp AP together in MTM tables Draper Lab ASSEMBLY PRIMITIVES 71 While Draper Lab has defined 9 main primitives and 9 subprimitives for ACCOMMODATE only the ones used in this paper are described below GRASP Device uses tool to grasp part s to be assembled or to grasp another tool POSITION Device executes gross motion trajectory carrying tools and or parts INTERFACE Device goes from state of no contact between
175. pass through midair point and wind up at new box place The sequence of destination points is translated by AL into the corresponding joint behavior by a combination of compile time planning and execution time revision Although a simple list of destination points is sufficient for some purposes many tasks require a more detailed specification of how motions are to be performed Items of interest include the time to be spent on each motion segment forces to be exerted by the hardware external forces to which the manipulator is to be compliant and conditions to be monitored during the motion This information is supplied in AL programs by the use of clauses which modify the basic motion statement For example move carburetor to inspection station via unloading pint where force xhat 0 force yhat 0 duration gt 2 sec via approach point on force zhat gt 802 do stop on eectric eyeinterrupt do signal passed checkpoint might occur in a carburetor assembly program where a carburetor has been assembled in a fixture and now must be moved to an inspection station While the carburetor is removed from the fixture the arm is made compliant to forces in x and y and the motion is slowed down to take at least two seconds The carburetor is then moved to the inspection station via an intermediate approach point To avoid the possibility that a small positioning error might cause the manipulator to shove the carburetor through the table the mo
176. powerful technique is to use a procedure that checks the consistency of the World and to call this procedure from the debugger at each of the test breakpoints Charles Simonyi personal communication Procedure calls that are expected not to be relevant to the problem are executed as a single step Finally the error 1 localized and the programmer convinces himself that his code in fact makes a mistake It is surprising how resistant the human mind 1s to the suggestion that a perfectly straightforward piece of code might fail under some circumstances Once the error is found it is fixed in place if at all possible and it often is possible if the debugger is capable of patching one piece of code in place of another and tested by backing up once again and seeing if the same fatal error occurs as before A more cautious technique of debugging is to step through all new pieces of code one instruction at a time in order to make sure they do not fail The methods for determining sources of error and fixingthem are similar to the outline above This method works well if the failing program 15 not heavily context dependent or if the error is so severe that the program never works A later stage of debugging deals with programs that fail only occasionally stepping through such programs to find bugs is tedious and generally unproductive In these cases debugging often proceeds by setting breakpoints and trying to figure out the exact situation
177. precise location of the spindle in its hand and had to grope for the hole to insert the spindle shaft in total assembly by the manipulator took longer than by a human by a factor of 8 times when the assembly was done neglecting overhead The same factor was expected from theoretical analysis DESCRIPTION OF THE ASSEMBLY TASK The parts of the assembly are shown in Figures 1 2 and 3 Four parts were assembled together the handle crank body of the sharpener base spindle assembled with the cutters in place and the shell NS foe wa eo 5 te P 3 Bed ELE eee en dM sauadsvyg jo e t 2x ve 1 e e sa a tot r t we x 2 m ea es Sees meer du a 0 S i ee ee e t gt gt 1 3 A A e e 2 St B E E 5 x t LI a as 89 en e cc PET gt EET gt ee a d FI 1 i E i r r gt es ene i 4 i E v ate eee e c 1 E n J of
178. r response these references are BACKQ and BACKR which take numeric arguments Therefore if the user examines the value of a frame and then decides to add a small vector to it the dialog may look like this user SHOW VALUE SNAME DEST alaid FRAM ROT XHAT 180 0EG VECTOR 1 2 3 user SET VALUE SNAME DEST SCODE BACKR 1 VECTOR 0 0 1 For this reason BACKQ and BACKR are not allowed as variables in the AL language 111 38 4 THE AL PORTAL AL programs can talk to ALAID in the same manner as the user The AL statement ALAID lt string gt sends the string to ALAID and waits for a response the string that contains the response is the value of the call to ALAID In this way a program can itself make use of the state saving features of ALAID and it can call in a new load module 5 ABBREVIATIONS Certain frequently used commands have standard abbreviations For example to look at a long set of consecutive pseudo code instructions it is awkward to repeat SHOW CODE PADDRESS 132004 SHOW COOE PADDRESS 132006 SHOW COOE PADORESS 132010 Instead the simple command linefeed suffices after the first location has been opened Also lt control P gt can be used for the PROCEED command Other abbreviations can be introduced as needed 11 39 CHAPTER 6 CONCLUSIONS In this report we have seen an approach to debugging that extends to control of the entire programming process During debugging the compiler is avai
179. rdware system was implemented including manipulators with tactile sensors and TV cameras tools fixtures and auxiliary devices a dedicated minicomputer and atime shared large computer equipped with graphic display terminals An advanced software system called AL was developed as a research tool for studying problems in assembly automation During the past year the AL system has been debugged improved and extensively documented In the near future AL will be used to program some simple assembly applications examples From a succession of applications generic assembly routines can be identified and accumulated in a library Additionally a design review of AL has recently been started to identify its strengths and weaknesses As a preliminary to this review a questionnaire concerning the potential use of AL at other laboratories has been circulated and responses are being received Work has begun on the classification of assemblies assembly processes and manipulators Improved languages and interactive programming techniques for assembly and vision are also being studied It 1 hoped that this work will lead to the extension of computer vision to areas that are currently infeasible such as picking discrete parts out of a bin Research has also included geometric modeling of objects and constraints assembly simulation control algorithms and adaptive methods of calibration This overview offers a concise summary of recent progress by the v
180. re and it is not possible to figure out who the culprit is A condition monitor that waits for that signal can then immediately stop execution and let the programmer poke around and find out what is happening Tracing evanescent situations is the most efficient debugging use of condition monitors A monitor can be used to trace the forces on an arm so that the programmer can figure out what a reasonable threshold might be for stopping the arm Condition monitors are also capable of testing variables and complaining if the values are bad but variable testing 15 the special forte of affixment structures A special changer can be associated with a suspect variable whenever its value is changed the changer can either trace the current value or it can make a validity check and complain if the value is bad Both of these programmable side effects find debugging use therefore in testing suspicions and running traces during the execution of the program In this sense they are very like output statements with which a programmer peppers his code in order to get some flavor of how the ingredients of his program are interacting In order to use this approach it is necessary to have some suspicions to modify the program in such a way that the suspicions can be tested and to then try out the code anew Interactive debugging usually does not follow such a tracing paradigm the debugger itself is used to do the necessary tracing If we aliow the d
181. red into a synch command involves signals that have short lifetimes If no process has received a signal after a period of time the signal disappears and the signalling process repeats the signal Backing past such a piece of code 15 easy since its effect 1s transitory This solution cannot be generalized to the types of events present in most parallel structures including those found in AL Section 3 Levels of Detail Another problem presented by AL is the fact that the code and data exist at several levels of detail Each statement in the source language is translated to a set of pseudo operations for the target machine that machine is implemented in PDP 11 assembler language in which each pseudo operation is expanded to a hand coded procedure Variables are used to clothe many disparate entities including program variables and expressions which are in many ways symmetric to program variables Condition monitors are also implemented as a funny kind of variable the value of which determines the state of the monitor and the code that it executes Force feedback is implemented through yet more complicated variables The question is at what level the user would like to carry on his extermination activities For errors in logic of his program he would most like to work in terms of the source language If 11 110 he wants to know the current status of the affixment structure information not available in the source language
182. rently running in the AL context and with extensions to it both simple and complex 11 3 CHAPTER 2 CLASSICAL INTERACTIVE DEBUGGING Success is the mother of disaster V Cerf The art of debugging programs has developed as a bastard son of the art of computer programming The first debugging was done by staring at the code until a bug was found by inserting intermediate output statements to test hypotheses concerning the expected state of the computation entire generation of programmers became familiar with the core dump either in machine representation it is said that one can even come to love hexadecimal or with some preliminary transcription into instructions or more often ASCII or EBCDIC tcx t rcpresen tation Programming in machine language has given rise to such interactive debuggers as DDT which allow the user to interrupt his program investigate it make changes and then allow it to continue The fact that in machine language program and data are represented by the same forms namely machine words makes such debugging especially natural Fundamental to such debugging 15 the concept of breakpoints which are locations in the program of interest to the programmer during his debugging When breakpoints are encountered the interactive terminal is connected directly to the debugger and execution of the program is suspendedl During this time the user can examine the values of variables set and remove
183. roblems This is accomplished by subtracting the directed length of the last three joints from the position of the hand This gives us three equations representing the position of the end of the prismatic joint These equations are only functions of the first three joint variables Denoting the end of the prismatic joint by T y Ty T7 and combining equations 4 amp 5 8 9 12 amp 13 we get Ty Tyg 7 5015 cO 80753 14 Tay 5 0015 0150 5 15 Tz Ty 69 5 S 16 We can now obtain a quation in the single variable 9 by simultaneously eliminating s and S from equations 14 and 15 After substituting tangent of the half angle equivalents for the sine and cosine of 0 the equation becomes a quadratic polynomial in tan whose solution follows 111 3 z Tx 5 2o 5 Ty This equation will yield two values for 0 corresponding to the two configurations obtainable by flipping the prismatic Joint over thus changing from a right to left shouldered arm or vise versa The solutions computed using the sign in front of the radical will produce positive rotation angles for joint 2 whereas the solutions using the negative sign will product negative values of 0 Since we always operate our arms with 0 in the range r 175 5 we will use the negative form of the solution If S T is equal to zero two special cases must be considered either
184. s Low Automatic Coding Choice of Data Structures Ph D Dissertation Stanford Artificial Intellgence Laboratory Memo AIM 242 Stanford Computer Science Report STAN CS 74 452 August 1974 Zohar Manna and Richard Waldinger Knowledge and Reasoning in Program Synthesis Stanford Research Institute Artificial Intelligence Center Technical Note 98 November 1974 C Murphy The Reliability of Systems unpublished manuscript date unknown J L Nevins D E Whitney Doherty D Killoran P M Lynch D S Seltzer S N Simunovic R Sturges P C Watson E A Woodin Exploratory Research in Industrial M odular Assembly The Charles Stark Draper Laboratory Inc Prepared for the National Science Foundation Memo No R 800 covering June 1973 to January 1974 March 1974 Memo No R 850 covering February 1974 to November 1974 December 1974 Richard Paul Modelling Trajectory Calculation and Servoing of a Computer Controlled Arm Stanford Artificial Intelligence Laboratory Memo AIM 177 Stanford Computer Science Report STAN CS 72 311 November 1972 Richard Paul Manipulator Path Control Proceedings of the 1975 International Conference on Cybernetics and Society 1975 pp 147 152 C Rosen D Nitzan R Duda G Gleason J Kremers W Park R Paul Exploratory Research in Advanced Automatton Prepared for the National Science Foundation Stanford Research Institute Project 4391 Fifth Report January 1976 Earl D Sac
185. s problem the well for the roller handle was tapered as shown in Figure 7 so that when the handle was dropped from about half an inch above the well it landed in roughly the right place and just needed to be tapped into place by having the manipulator hand rest on it and drag it towards the center by friction until the hole end of the crank was flush with the fixture The base was put into place by wobbling it a little while moving it down and stopping motion of the arm when a force was encountered after releasing the base the hand was lifted closed and tapped down on the base The spindle presented special problems since the clearance between the shaft and the hdle into which it was inserted was 0 004 inches while the arm reading was given in terms of 0 01 inches although noise in the A D channels and devices resulted in an uncertainty of 0 04 inches This meant that two successive readings without movement would indicate the arm to have moved by as much as 0 04 inches The spindle shaft was touched against the sides of the main fixture in order to locate more precisely the position of the hole and then a spiral search in steps of 0 04 inches was done to actually insert the spindle shaft into the hole While this small step may appear to be very close to the uncertainty of the A D channels it had to be used since larger steps would have resulted in the arm making serious overcorrections Once the first part of the insertion had bee
186. s for which AL was designed the hand is the only part of the manipulator that interacts directly with other objects The position of the rest of the arm is generally irrelevant so long as it doesn t collide with anything Thus AL programs describe motions by specifying a sequence of frame values through which the hand must pass For instance move blue to box grasp via grasp approach Since the purpose of manipulation 15 to move objects rather than to get the manipulator s hand to particular places thts concept has been generalized to allow the user to describe motions in terms of frames other than the hand itself Thus affix box to blue move box to new box place via m idair point Here the affix statement tells the system that changes in the value of the blue hand are to eme quevP Gey AEN Ce GEE GE qa SER Gp Ge Ge Ge Gres OE GD Gan am 2 Actually this is an over simplification box grasp would merely be marked as invalid and a new value recomputed when required 3 The manipulator hardware at Stanford consists of two Scheinman arms one of which is anodized blue and the other gold Thus blue and yellow are predefined AL frames corresponding to the hands of the two arms At the time of this writing only blue has been interfaced to the runtime system IV 4 cause corresponding changes in the value of box This information is used to produce hand positions that cause box to
187. s of freedom and locator C restricts the last degree of freedom The form of the locator selected depends onthe condition of the reference surface of the part finished surfaces can be supported on a surface rather than suspended on points while rough surfaces are given as few points of contact as deemed necessary for stability of the part Clamps hold the workpiece firmly against the locators provided and resist all forces introduced by the operation In the assembly of the pencil sharpener fixtures were used to locate the parts precisely since the forces encountered in the assembly process were small much smaller than in the case of machining clamping was not done by external clamps Instead the manipulator was used to provide the necessary reaction against the locators Fixtures were made by casting plaster of Paris in a box and dipping the parts suitably covered with modelling clay and masking tape and coated with a thin layer of petroleum jelly into the plaster to make a mold for the part The plaster was then machined to provide space for the manipulator fingers to be inserted around the part to be gripped Since the surfaces of the sharpener were smooth surface contact was used instead of point contact Drawings of some of the fixtures used are shown in Figures 5 6 and 7 The advantage of using plaster of Paris was that making the initial mold and machining was a very simple and economical process which did not require specia
188. s used Draper and MTM and relationships to WAVE are shown in the following table Draper RELEASE GRASP POSITION ROTATE INTERFACE ACCOMMODATE CPX ACCOM INSERT DEPRESS PRECISE POSITION RETURN WAIT MTM HUMAN RELEASE GRASP MOVE REACH T amp AP POSITION REACH WAIT MTM YELLOW OPEN CENTER CLOSE GOTO MOVE T amp AP POSITION GOTO MOVE WAIT WAVE OPEN CENTER CLOSE GOTO MOVE CHANGE PLACE CHANGE GOTO PARK WAIT Elements is square brackets J indicate elements that were necessary as assembly primitives but were unavailable PRECISE POSITIONING OF THE MANIPULATOR The time taken for the manipulator to null out position errors while not obvious in the analysis slowed down the assembly process To speed up the assembly the motions which did not require precise positioning were performed without nulling out the final position errors Of special importance was screwing in the spindle since software limitations required that the rotation be made in steps of 120 degrees To null out the error at the end of each 120 degree twist meant that the motor ground away to achieve the last bit of unnecessary precision Not nulling the movement resulted in the handle turning a few degrees more less than the desired amount but this was not at all critical 21 It was found that asking the arm to move directly to the part to be picked up inevita
189. sily handled by checking to be sure that is less than the length of the fingers 90 degrees could be used here however the extra 10 degrees lessens the chance that the hand or wrist will interfere with something gt 1 33 In the second case where both ends of the pin will go into holes we lpin dr rtp 2dadi lfp TK Again if the interval 1s short its midpoint will be picked If the interval 1s longer then three values will be used Since the pin must be turned around the only value for 7 15 90 degrees Picking Values for 0 Once we have picked values for and d we still must determine the rotation value w it is necessary to consider actual arm solutions Unfortunately the only way presently available for doing this is to invent values and try them 19 Values of are considered in increments of 45 degrees For each value the grasping position is calculated and the arm solution procedure is called to see if the position 15 feasible In some cases we ma produce a great number of candidate grasping positions Therefore the solutions for all feasible positions are graded for toughness and non degeneracy and only the best few values are retained for further investigation The current rule for evaluating arm solutions is very crude the angle of the elbow joint 5 of the Scheinman arm is examined Angles near 45 degrees are considered best Our selection procedure looks someth
190. sor receives an answer it can use the communication number to determine which process within its domain made the request and is awaiting the response buff et header MESTYP This field distinguishes whether the message is an answer a query or a tidbit and whether it comes from the PDP 10 or from the PDP I 1 When a message arrives at a processor this information is used to decide what to do with the message buff amp header MESLTH The length in bytes of the message 11 20 5 A SAMPLE DIALOG BETWEEN MACHINES Suppose that the user has asked the 10 for the value of some variable The communication between the machines might look like this GETBUF 50 from PD P 10 BUFALC 50 20436 from 11 lt message at 20436 type request SHOW VALUE SNAMEvar1 USEBUF 20436 from POP 10 1 gets value makes an answer at 22312 type answer SCALAR 3 0 TAKBUF 22312 from POP 11 lt PDP 10 gets answer reports It to user RELBUF 22312 from PD P 10 6 ROUTINES RESIDENT ON THE TWO MACHINES Even though symmetry of form between the two processors is desired each has a different regime in force that constrains the implementation The runtime system is under the tutelage of the kernel which has control over the various processes and scheduling The part of ALAID that resides on the PDP 10 must be compatible with the compiler so it has been Written in SAIL the language in which the compiler is written using the SA
191. stinations i xf destinations i crlf k is step J until j do print if ksi then else location variable base destinations k v xf destinations k crlf prin t gualifiers erlf end end Here we first break the destination sequence up into subsequences with common controllable frames and then generate a motion statement for each subsequence There are several possible pitfalls since the semantics of two successive motion statements are not identical to a single statemerit especially where the qualifiers include stop on force tests At present this difficulty is solved by being careful that the procedure will not be called with arguments that split the motion at a bad point This solution was satisfactory for our present small set of code emitters but something better will have to be done in the long run An alternative approach would be to compute the relation between each controllable frame and the manipulator and then to write the motion purely in terms of the manipulator frame This solves the abovementioned difficulty but introduces additional complexity making the output programs harder to read A better fix would probably be to extend the syntax of AL to allow hybrid destination lists and then allow the AL compiler _ to worry about the relation to manipulator frames 4 8 Example The task strangely enough is insertion of an aligning pin into a hole drilled in the to
192. stinct ways for handling this One way is to use a trans variable to recompute each frame value each time it is needed Thus a user might write an expression like boxegrasp xf to specify the proper hand position for grasping the object whose coordinate system given Various flavors of these primitives come under many names See for instance 8 for further discussion 1V 3 by the frame box This approach can get tedious where the same frames are being referenced repeatedly and tends to hide the intent of a program behind a smokescreen of frame transformations The alternative method of using a separate frame variable for each frame of interest makes motion statements easier to read and write but means that all associated variables must be updated whenever something is changed The affix construct AL allows the user to specify that a variable is to be continuously computed from other variables For instance affix box grasp to box at grasp xf would cause the assignment statement box grasp boxegrasp xf to be performed automatically every time box is updated When one object is assembled to another or when an object is grasped by the manipulator it is customary to affix their location variables For example affix cover to box affix box to blue The data structures associated with affixments thus form another important part of an AL program s model of the world Motion Statements In the task
193. t 8 oz do stop on arrival do begin This should never happen abort Help Help The box has been stolen end correction that inv spot on surface pinevector 0 0 0 move pin to n Aole position via hole approach on force pinezhat gt 8 0 do stop distance off zhat inv in Aole positionpinevector 0 0 0 correction et cetera Alternatively one could use correction to make an appropriate modification to the box or hole location For instance box box vector 0 0 correction It 15 possible to take advantage of affixment to do away with the need for any explicit mention of correction For instance affix spot on surface to box rigidly j move down until hit the spot move pin to spot_on_surface vector 0 0 1 0sinches on force pinszhat gt 8 oz do stop Say that s where we got to Spot on surface pin The rigid affixment asserts that whenever either frame is updated the other is to be update amp appropriately Thus the assignment statement will translate the box location to account for whatever distance the pin actually travelled This technique is easy to write since you don t have to invent variables or figure out complicated arithmetic expressions Also it is easy to read since the code is terser and the assignment statement more nearly reflects the intent of the motion statement which was to get the pin to spot on surface 2 3 2 Error Recovery So far we ve be
194. te type RELBUF a The PDP 11 sent the PDP IO a message at address a The PDP 10 has looked at it and 15 finished with it 3 NOTES FROM THE TO THE PDP JO These are the note types that the PDP 11 can place in the PDP 10 s notebox 11 19 note type BUFALC s gt A requested buffer has been allocated the PDP 10 s use It has size s bytes and is at address a note type TAKBUF The buffer that starts at address a is message for the PDP 10 which should look at it and act on it 4 MESSAGE BUFFERS Once the dickering between processors has established room for a message the responsible processor fills the given area known as a message buffer There are several kinds of messages The first is a request which is either a query for information or a directive to be obeyed The second message type is an answer which either contains information queried by some other message or indicates that some directive has been carried out A third type of message 15 the tidbit which is information possibly but not necessarily interesting to the destination processor and which may be ignored it is never acknowledged Each message buffer holds the contents of the message along with a few header words that indicate the nature of the communication buf fer header MESID This entry is the communication number of the message Answers to requests have the same number as the request in that way when a proces
195. ted Interactive debugging of programs in these languages is made difficult by the fact that the programmer must be able to associate machine code to source code This process is made easier by such debuggers DDT and RAID the latter being display version of DDT a standard debugger on the PDP 10 which can display memory cells in various modes including symbolic instruction mode and various arithmetic modes Another feature that these debuggers offer is that they can refer symbolically to locations in memory if the programmer has named a variable felicity then that name may be used during debugging as well If a program label is charity then that 15 how the debugger will refer to that section of code Thus these debugging programs contain disassemblers that can take assembled code and recreate the source that spawned it In order to allow the user to modify instructions classical debuggers also include primitive assemblers as well The use of symbolic names for instructions as opposed to numeric format for labels and for variables is a special case of a very important idea in debugging the code that is being debugged should be presented to the programmer in as close a way as possible to the code that he himself wrote Unfortunately DDT and RAID only work on the machine language level a level at which most programs are not written A significant effort in the direction of source language debugging is the debugger BAIL
196. the first three and last three joint transformations and hand code the computation of the resulting matrix elements The total arm transformation could then be determined by multiplying the two matrices together in the standard fashion A further improvement to this scheme has been suggested by Lou Paul Instead of forming the full 4x4 matrix representing the transformation for the last three joints only its last three columns are explicitly computed The last three columns of the full arm transform can then be determined by multiplying the two partial transformations together while the first column of the full transformation can be formed by taking the cross product of the second and third columns We have written a program to perform these operations for the Stanford Scheinman Arm and find that it executes in approximately a third of the time of our former method of multiplying the six link matrices together The nominal execution time for this new program is approximately 2 0 msec The two partial transforms used for this computation are presented below All four columns of the transform for the last three joints are presented for the sake of completeness Transform from Al to AS 6 08 s 50 5 0 50 5 c 010 50 59 50 50 5 0 5 0 50 c 0 5 0 Transform from A4 to A6 s cO 059 50 c0 s0 c0 cO 50 50 50 450 5 6 50 0 50 50 c Se 0 0 0 B AUTOMATIC FORCE WRIST CALIBR
197. the hole the limiting tilt angle is arc cos 1 C 8 degrees Note that while the Stanford Arm gripper was designed not to be compliant compliance was assumed at the gripping point for purposes of this discussion The spindle had a step and the hole had a chamfer so the Draper parameters are L g distance from spindle end to grasping position 2 inches 6 chamfer 0 025 inch Step 0 05 inch Dealing with the step as though it were a chamfer insertion all the way was possible without two point contact if the offset from the center 0 05 0 025 0 075 in in addition the entrance tilt lt 8 p uL will be less than 0 215 deg depending on the offset With the hardware available and the friction characteristics of the joint motors it was calculated that a minimum penetration of 0 8 inch using a nominal downward force of 10 oz was necessary to prevent jamming SENSING REQUIREMENTS Position Sensing Position sensing would be all that 15 necessary if the arm could be positioned with a tolerance of within 0 001 inch and tilt of 0 1 degree and if parts could be positioned to these tolerances within the fixtures However given the Stanford Yellow arm with a repeatability of 0 04 inch and possibility of specifying distances to within 0 01 inch it is essential that force and touch feedback be used Vision Sensing Verification vision would be useful in determining the initial process of inserting the spindle in
198. the offset in the right 8 bits In order to make a level offset pair uniquely refer to a variable it 15 necessary to know which of several parallel blocks that 15 which context contains it uniquely For example consider this piece of code 11 26 BEGIN level 0 name 1 SCALAR 51 offset 10 COBEGI BEGIN level 1 name 2 SCALAR S2 offset 10 SCALAR S3 offset 12 END BEGIN level one name 3 SCALAR S4 offset 10 END END SCALAR 55 offset 12 END and S4 because they are each the first variable in the first level During execution there 18 no possible conflict because two different interpreters are active the one that has access to S2 cannot see S4 and vice versa Therefore a third datum 15 necessary to distinguish variables the name of the interpreter or equivalently a context in which the variable is unambiguous Variable S2 is fully described by the triple 2 1 10 which gives the name the level and the offset variable 53 is 3 1 10 During execution the name is always implicit and no code need be generated However if ALAID wishes to access a variable it must specify the name as well either explicitly or implicitly The current context gives a partial implicit specification if it contains only one interpreter no interpreter name is necessary If there are several interpreters then those variables to which each has access need no interpreter names the others do The s
199. thods developed in my dissertation 28 we get H estimated position of hole with respect to work station H A Pinit estimated initial position of the pin i estimated accuracy of at runtime AH b APijn estimated runtime accuracy of P AP init subject to constraints on and For planning purposes we will mainly deal with the expected locations H 0 Pinit Pini Also we need several important paramters describing how the pin fits into the hole direction The end of the pin which is to be inserted into the hole The distance the pin is to go into the hole d The maximum sticking distance into the hole that the pin can jam without making it all the way to where it is supposed to go Thus ded represents a minimum threshold for telling whether the pin insertion is successful 3 Here we are being a bit sloppy in our use of H The difficulty is that we must deal with three separate entities representing the hole 1 the object model representation a LEAP item 2 our location estimation and 3 a variable in the output program Generally this discussion will center on I and 2 3 isn t needed until time comes to generate the actual program text IV 29 A0 Maximum possible axis misalignment for the pin insertion to succeed ok p g P Arg Maximum possible radius misalignment between the pin and
200. ting example In fact the program portal can be thought of as a potentially infinite set with portal for each process currently active 11 22 This would be the case when more than one process has hit a breakpoint Given the plurality of portals effort must be exerted to direct information to the right place Queries are signed by the originator and responses carry the same signature so it is straightforward to direct the answers back to the proper portal A harder question is to decide where to send a query that cannot be answered locally In the case of a two member ALAID the usual answer is to send the query to the other member If the query itself can be answered but to complete the answer requires some information that is not available then a query can be formed to get that information from the other member If that query fails then the user can be asked he is also a member of the ALAID community although perhaps not in such good standing In that case care must be taken to distinguish between the user s answers and further queries that he might make In this way portals can be characterized by their ability to make and respond to queries The portal that connects ALAID with a program is usually useful only for transmitting queries into ALAID and answers back out The user is also usually a questioner but in some situations he can be asked queries directly The other members have both properties of originating an
201. tion in its code CONVERT PADDRESS SADDRESS L1 The appropriate symbol table resides in the PDP 10 so it takes this request and finds that L1 1 at location 132012 It then responds to the PDP 11 PADDRESS 1320 12 Now that the label has been resolved the PDP 11 proceeds to place the breakpoint Having finished its task it responds to the PDP IO DONE Some tasks rquire more communication than this simple example demonstrates If user wishes to assign a value to a variable then the variable must be sought in a symbol table the value must be evaluated possibly involving compiling code and running it or else by repeated requests for the values of variables and finally the assignment can be made These examples point out the various portals to which each ALAID member must respond One portal is the user In the example above this portal is In use on the PDP 10 member only Each of the two members in the example has a portal devoted to the other member In addition ALAID can be used as an interface into AL from another program in this case a portal is devoted to that program One example is when the executing AL program encounters the breakpoint set in the scenario above It then informs the PDP 11 member of ALAID through its own portal ALAID then sends a tidbit to the other member More than one portal can be in use simultaneously as we have seen the PDP IO member is active on two portals during the breakpoint set
202. tion is terminated as soon as the force in the z direction exceeds a half pound Finally as soon as an electric eye detects something an external control signal is generated Tta jectories Ultimately ail manipulator motions must be described in terms of joint motions since joints are what the runtimc system can control However this representation is awkward for specifying motions and introduces a needless degree of hardware dependency if it is used Motions are specified in AL programs by giving a list of positions through which an object is to pass The requited coordination is achieved by solving the joint angle equations for each position These data points are then used to produce polynomials in time which describe the behavior of each joint Unfortunately the computation required for preparation of these polynomials is non trivial Consequently the compiler must pre compute trajectories based on a planning model of xhat yhat and that are unit vectors in the x y and z directions 5 This method was developed by R Paul and is reported in 19 More recent refinements may be found 5 and 10 In his recent work Paul has abandoned polynomials in favor of an intapolation scheme 21 20 IV 5 expected affixment and frame values These precomputed polynomials are modified by the addition of higher order terms just before the motion is executed so that the positions reached correspond to the actual runt
203. tions about the x y and g axes and are shown below o 0 0 M 0 0 1 010 0 0 I M 10 0 0 10 10 M 1100 1000 The constraints on the free variables are 5 deg sys 5 deg 0 25 degrees lt lt 0 25 degrees 0 25 degrees 0 25 degrees 0 25 degrees 1 s 0 25 degrees where ny nand n repres nt the hand rotation errors and represents the rotation error of the hole Solving we get 13 These parameters correspond to the best overall strategy found in Section 4 8 IV 39 09 354 30 306 60 306 90 354 120 306 150 306 Consquently 46 354 Once this value is computed we compare it to the allowabk limit 0 If the value is out of bounds then the pin to hole alignment may not be good enough to guarantee success Presently this is grounds for rejection of the strategy Other options would be to add another parameter to the search loop so that different pin orientations as well as different positions are tried to include smarter accommodation techniques or to attempt in some way to ascertain the pin hole orientation more accurately 4k 2 Error Along the Hole Axis Az is easily computed from A1 Recall that our in hole test examines how far the pin gets along the hole axis before being stopped If it doesn t get far enough then we assume that we hit the object and must try again For this test to work we must be sure that Az
204. to the hole Before insertion takes place it 1 assumed that the spindle and the hole each other and verification vision could tell how close they are and actually monitor the positions of the spindle and the hole as the spindle approaches the hole For the task given being able to sense a tolerance of 0 002 inch and an angle of 0 1 deg would enable decisions to be made as to which direction to move or tilt the spindle Resolution at a finer level would enable how much to be computed as well With verification vision a V 23 spiral search would not need to be done to locate the hole as was necessary in the assembly Force and Touch Sensing Touch sensing is necessary at the fingers together with ability to measure hand openings as a means of telling whether the object has been grasped at the right place if at all Since the parts are fairly rigid the touch resolution is not critical as the gripping force is about 5 Ib Force sensing to a resolution of 0 5 oz would allow the arm to know if the part has slipped out of its grasp by checking the weight at the end of the hand Force sensors behind the hole at the fixture and behind the spindle would be helpful in telling the forces and moments at the hole and the spindle and together help to prevent jamming ARM DESIGN Some of the problems in the movement of the arm stem from the fact that six degrees of freedom determine an essentially unique solution for motion fr
205. trol of machine interrupts The problem of non interference has been discussed by D Swinehart Swinehart 74 his interest is to allow the debugger to oversee and report on the state of continuing computations without interfering with them The method he employs is to make the debugger itself a process Just like the others The debugging process is a window into the inner secrets of any other process it chooses to examine It can link itself into the data structures of that process and therefore it has access to variables and code local to the object of its scrutiny If variables are held in common among several processes then the data structures within any one of them point to all the global variables Of course there is a problem of naming since different variables may be I 1 9 called the same thing in different blocks and in different threads This problem of non unique naming is fairly well understood standard PDP IO DDT and RAID have separate symbol tables for separate blocks and they maintain a current block If a variable is requested it is sought first in the current block and then in successively more global blocks Commands are provided to change the context by moving to other blocks This separate symbol table idea can be generalized to the domain of concurrent processes each one has its own symbol table that associates internal names those that the process itself uses when making reference to a variable with programmer d
206. ustrated in our discussion of the pin in hole example Among the more important considerations were 1 We made a number of unstated though obvious assumptions about the location of the various entities For example the hole was assumed to be unobstructed 1 the box better be right side up 2 In grasping the pin we had to consider whether the hand could reach the required locations If it is possible for the box to be in more than one position or orientation then this must be taken into account 3 We made use of the fact that the pin could be attached temporarily to the hand by grasping Similarly it 1s important to realize that a subsequent motion of the box will cause the pin to move too 4 The code contains many assumptions about the accuracy of our variables pin and hole In deciding whether tapping or a search were necessary for instance it 1 necessary to consider whether the along axis determination is good enough and whether in plane errors are within the capture radius required by the pin to hole geometry Any reasonably sophisticated maniputator language allows much of this information to be represented explicitly in the program In AL for instance object locations are represented by frame variables and attachments by affixments In writing programs it is thus necessary to keep track of programming information such as what variables have been declared and what calculations have been perfor
207. will send the user a message that identifies the process and where itis executing PROCEED Any halted process in the current context is allowed to continue execution Once the user is satisfied that the program is behaving properly in the region of a breakpoint this command is useful for proceeding to the next breakpoint HALT All processes within the current context are halted As they stop they send the user a message telling where they are This command will not stop a moving arm since the process controlling the arm is in the middle of a single pseudo operation JUMP lt address gt All processes in the current context stop executing at their current location and start executing at the given address If a block must be exited or entered before this jump can be done then the block exit entry code is executed appropriately Since this command is dangerous it is not honored if the current context contains more than one active process EXECUTE lt instruction gt All processes in the current context execute the given instruction It is not necessary that the processes be halted first as soon as they are finished with the current instruction they coun the given instruction and then proceed with whatever they were doing Of course it is most usual to give this command to a stopped process The instruction can be in SCODE or PCODE modes This facility is quite powerful any operation available to the AL language can be performed in th
208. y other one in such a way that the sender is identified it suffices One generalization of treating each processor as an indivisible entity is to treat each process as an entity The ALAID member on each process could be shared among all the processes within a single processor giving rise to a two level hierarchy of ALAID communication Once a hierarchy of processes 15 introduced it can be used as a general ordering technique for the entire cooperative computation with higher level processes charged with parcelling out tasks to inferior ones and directing communcations between those lower processes and the outside world Using ALAID based communication might be fruitful in this domain much further research is warranted to investigate these issues Another realm for further work is the dynamic redistribution of information Space constraints may force one Processor to relieve itself of some information of secondary importance by giving it to another processor to store Frequent references to some data may imply that thosedata should be made more accessible by copying them or moving them to the requesting processor To identify these situations set up new processors with ALAID members and transmit the information is a problem of some difficulty whose solution may prove quite powerful 11 25 Section 2 Sym bol Tables As we have seen debugging is an activity that requires repeated examination and modification of a test program
209. ymbol table for variables which is used to make correspondences in both directions is structured according to interpreters This structure implicitly includes the interpreter name in each entry The individual entries include source code name internal compiler name which is different and the level offset pair Search in the table for level offset pairs is conducted first in the top level interpreter of the current context and then if necessary in each of the next lower level interpreters The result of the search is either a level offset if that will suffice or a name level offset if necessary or an error condition found more than once in which case the context 1s not sufficiently specific If the variable is not found at the current interpreter then surrounding contexts are searched until the variable is found To find the source name for a given level bffset pair if the level 15 deeper than the current context then there is no unique solution Otherwise it is fairly easy to find the source name Once a name level offset triple has been found to find the value of that variable on the PDP 11 requires that all interpreters be accessible by name This is accomplished by keeping a linked list of all interpreters a tree structure would be more efficient if there are many and providing a special purpose pseudo op INAME that causes the interpreter that executes it to assume a new name 11 27 2 PROCESSES The previous discussion

EXPLORATORYSTUDY - The Stanford University InfoLab

Contents

Download Pdf Manuals

Related Search

Related Contents