Fast, strong and compliant pneumatic actuation for dexterous tendon
Contents
1. To Compressor PW Power Hardware details The pneumatic control unit Fig consists of the following subunits 1 Air preparation module consists of air compressor and LFR 3 4 D 5M MAXI A filter regulator with a flow rate of 7 600L min at 6 bar Detailed Specifications can be found in 8 2 Pneumatic control Valves Pneumatic control unit con tains 40 MPYE 3 5 proportional valves from FESTO Pneumatic valves MPYE 5 M5 010 B used for Hand muscle actuation unit have a flow rate of 100I min at 6bar bandwidth of 125Hz and weight 290g Detailed specifications can be found at 8 3 Air supply manifolds MPYE series proportional valves were designed by FESTO as stand alone high end devices to be used as individual units Thus no manifold solution have been made available for these valves After careful examination PAL manifolds of TIGER 2000 series 8 valves were modified for compatibility with MPYE se ries valves Pneumatic control unit contains 5 PAL 10 B manifolds 8 each serving 10 pneumatic control valves Evaluation and Experience Proportional valves were pre ferred over binary switching valves despite their moderate performance on parameters 8 and 9 This was driven primarily by our application requirement that called for fine grained control over flow rate for fast and precise actuator control Binary switching valves achieve variable flow rates using pulse width modulation PWM which restricts smooth and precise2. USA E mail vikash zhexu todorov cs washington edu how does one pull on all the tendons Here we offer a solution which we believe is quite universal and reflects the state of the art in actuation technology The rest of the paper is organized as follows In the next section we outline the design considerations and the choices we made We then describe the design of the new actuation system in detail followed by experimental results characterizing speed strength and compliance of the improved ShadowHand Finally we consider future simplifications which could make the system considerably less expensive while preserving its advantages We also summarize the application of our new actuation system to an independently developed 20 dof UW hand 5 II OUTLINE OF DESIGN CONSIDERATIONS AND CHOICES The requirement for high dexterity naturally leads to the choice of pneumatic actuation Indeed this may be the only available technology that combines speed strength and com pliance on the mechanism level with small and lightweight actuators in turn allowing portable drives with as many as 40 units to be built the ShadowHand has 40 tendons On the other hand the built in actuation system in the Shadow Hand was pneumatic and did not meet our expectations This however can be attributed to factors that can be avoided as follows First in order to obtain sufficient force from small air muscles one has to mount them so that they are p
3. control over flow rates Moreover PWM frequency is limited by the bandwidth of the device further constraining smooth behaviours A bank of binary valves can get very loud when switching at a fast rate PAL manifolds are capable of serving valves with flow rates up to 2600 min i e our manifold solution is capable of facilitating upgrades to all MPYE series models without any modification High flow rate 20X the requirement of the presently used MPYE 5 M5 010 B model manifolds are preferred since they act as pneumatic buffers thus avoiding air pockets under heavy flow requirements Our air supply manifold presently contains 2 expansion slots Addition of manifolds for future expansion is facilitated by the high level of modularity in the design within reasonable efforts Pneumatic latencies are a function of tube lengths between the actuators and the valves The air path between them was minimized to the extent possible Pneumatic manifolds and valves were configured to minimize the pneumatic interface face where valves expose their pneumatic connections to ac tuators surface area thus minimising the tube length between valves and respective actuators VII ELECTRONICS UNIT The electronics unit consists of the followings components 1 Sensors Cylinder pressure sensors and piston length sensors housed by Muscle actuation unit Figure 2 2 NI PXI Chassis with multiple A D and D A boards 3 Power Supply 4 Emergency Sto
4. made Hardware details The actuator unit Figure 2 consists of double acting Airpel series cylinders commercially man ufactured by Airport Corporation Single acting cylinders are often air return or spring return and do not allow control over the return force Double acting cylinders were selected for complete control over the actuation force in both directions although this feature is not yet utilized The finger tendon s actuator unit has stroke length of 37 5mm can produce up to 42N of force and weights 45 7grams The wrist tendon s actuator unit has stroke length of 50mm can produce up to 125N of force and weights 95 7grams Detailed specifications can be found at 6 Double acting modules were used in single acting mode in the Muscle Actuation Unit Finger tendons were actuated using M9 cylinders and wrist tendons were actuated using M16 cylinders In tendon driven systems it seems preferable to have a small force on the actuators to avoid tendon slack when the muscles are in passive non pulling state However the non zero force from the passive tendon creates co contraction and adds undesirable stiffness to the joints The return force of single acting cylinders are fixed and non observable and cannot be compensated using pneumatic forces since both forces act in same direction Double acting cylinders with no return were employed to simulate virtual variable stiffness springs to intelligently handle tendon slack and minimiz
5. organization the two tendons in a pair will typically have different and possibly variable moment arms The FESTO MPYE valves are 5 3 valves and are in fact designed to power pairs of cylinders With these simplifications the cost can be reduced by half The ShadowHand skeleton without any actuation still costs around 60 000 However as shown in our companion paper 5 one can 3D print a hand with comparable dexterity with cost of materials around 100 and a couple of days of assembly work Combining these two advances it should be possible to make dexterous robotic hands whose performance exceeds any product available on the market to today for less than 40 000 These hands rely on off the shelf components and 3D printing and could be built in academic labs Finally we are not certain that proportional valves are actually needed We clearly need valves that respond quickly and have high flow rates but what if they were binary and thus presumably a lot cheaper In principle proportional valves provide smoother movement but given that air dynamics introduce low pass filtering it remains to be seen how much the performance of the to be developed control schemes will degrade in the presence of binary valves If the degradation turns out to be negligible this will result in substantial further reduction in the cost of the robotic hands we envision XI SUMMARY The development of the present system was motivated by our desire to
6. shutdown in case of operational emergency e g robot performing an unde sirable movement However careful observations listed below revealed that powering down the pneumatics is a more reliable and faster option e Valve charging up its actuator s pressure when emer gency was triggered Due to pneumatic buffers feed ing the control valves the actuator will continue charging up till the pressure drops in the buffers and then actuator s residual pressure flushes out e Actuator is pressurized and valve is closed to main tain the chamber pressure when emergency is trig gered This pressure gets trapped in the cylinder and emergency not avoided The correct way to process an operational emergency is to exhaust the input pressure by flushing out the pneumatic buffers using a high flow exhaust port and opening the valves to exhaust out the cylinder pressure This requires a pneumatic shutdown while valves maintain their input operating power For a non operational emergency circuit breakers are provided at all levels from top which powers down the entire system to bottom level which powers down individual subcomponents VIII COMPUTATIONAL UNIT Design parameters Design parameters established for the computational unit have been listed below in decreasing order of priority 1 Reliable communication 2 Minimum communication latency 3 Maximum computational capacity 4 Maximum data bandwidth Hardware details NI PXIe PC
7. tendon driven direct actuation under all configurations single double acting bundled distributed are supported The last three modules can be used for any pneumatically actuated system with appropriate configuration of the actuator unit V ACTUATION UNIT Presently the actuation unit is configured as a tendon driven hand thus we call it Muscle Actuation Unit It consists of the following sub modules 1 Hand muscles actuation unit 2 Arm muscles actuation unit In addition to the above the muscle actuation unit houses two components of electronic units the pressure and length sensor These are discussed in detail in section Muscles acting over the finger and wrist tendons form hand muscle actuation unit It consists of Actuator units and Housing assembly Muscles acting over the lower arm elbow upper arm and shoulder form Arm muscles actuation unit A Actuator unit Design parameters The selection of actuators was based on the design parameters listed below in decreasing order of priority 1 Minimum friction and stiction 2 Maximum force output 3 Fast with minimum response time 4 Minimum weight of actuators 5 Compactness of actuator unit 6 Compatibility with sensory devices 7 Availability in different configurations for different use cases 8 Durability It should be noted that these principles focus on generic actuation mechanisms No restrictions specific to tendon driven systems were
8. Fast strong and compliant pneumatic actuation for dexterous tendon driven hands Vikash Kumar Zhe Xu and Emanuel Todorov Abstract We describe a pneumatic actuation system for dexterous robotic hands It was motivated by our desire to improve the ShadowHand system yet it is quite universal and indeed we are already using it with a second robotic hand we have developed Our actuation system allows us to move the ShadowHand skeleton faster than a human hand 70 msec limit to limit movement 30 msec overall reflex latency generate sufficient forces 40 N at each finger tendon 125N at each wrist tendon and achieve high compliance on the mechanism level 6 grams of external force at the fingertip displaces the finger when the system is powered This combination of speed force and compliance is a prerequisite for dexterous manipulation yet it has never before been achieved with a tendon driven system let alone a system with 24 degrees of freedom and 40 tendons I INTRODUCTION The unique capabilities of the human hand have long inspired roboticist in their pursuit to develop manipulators with similar dexterity We use this term here to refer to a combi nation of features many independently controlled degrees of freedom dofs speed strength and compliance Simple and isolated tasks such as grasping can of course be accomplished by simpler devices Nevertheless if robots are to perform a wider range of tasks in less structured envi
9. Health and the US National Science Foundation Thanks to Alex Simpkins and Rob Mabery for their help with the design and machining REFERENCES 1 Shadow Robot Company 2 S Jacobsen E Iversen D Knutti R Johnson and K Biggers Design of the utah m i t dextrous hand in Robotics and Automation Proceed ings 1986 IEEE International Conference on vol 3 apr 1986 3 A Deshpande Z Xu M Weghe L Chang B Brown D Wilkinson S Bidic and Y Matsuoka Mechanisms of the anatomically correct testbed act hand JEEE ASME Trasactions on Mechatronics 2011 E Chao J Opgrande and F Axmear Three dimensional force anal ysis of finger joints in selected isometric hand functions Journal of Biomechanics vol 9 no 6 pp 387 396 1976 5 X Zhe V Kumar and E Todorov A low cost 20 dof anthropomor phic robotic hand Design actuation and modeling manuscript under review 6 Airpel http www airpot com html airpels html 7 Airpel AntiStiction http www airpot com html anti_stiction html 8 Festo Company 9 PhaseSpace Company 10 National Instruments Company 4 11 Omron Company 12 SMC Company 13 SICK Company 14 A D Deshpande R Balasubramanian R Lin B Dellon and Y Mat suoka Understanding variable moment arms for the index finger mcp joints through the ACT hand in JEEE BIOROB 2008 15 X Zhe V Kumar Y Ma
10. Ie8371 x4 MXI Express is used to communicate with the chassis using a high bandwidth PCI Express link Any normal desktop or server computer with PCI express slot can serve as a Computational unit 3D motion tracking system communicates using a standard ethernet port PCIe link is used to retrieve pressure sensor and length sensor data and to command the Pneumatic control assembly Design evaluation and experiences Data communication using PCIe link and standard ethernet link ensured that our system is compatible with any standard computer without any special requirements Presently a 3 0Ghz AMD Phenom tm II X6 1075T 8 00 GB machine with Win7 OS is used to control the system IX DESIGN EVALUATION Final hardware was evaluated on various design parameters using two tendon driven hands Actuation system was perfectly compatible with both hand designs without modification 1 24 dof ShadowHand developed by ShadowHand com pany 2 20 dof UW hand being developed at our lab independent of the actuation system 25 00 26 00 45 00 920 00 Top View c Fig 6 a ShadowHand mounted on the actuation unit b UW hand mounted on the actuation unit c ShadowHand finger dimensions A Force and compliance System force and compliance characteristics were studied using the ShadowHand and UW Hand An external force of 6 grams for ShadowHand and 8gms for UW Hand at the index finger tip was enough to flex the MCP joint th
11. ds onboard electronics Designing onboard electronics can definitely make the overall system compact and independent These advantages from design of onboard electronics would have come at the cost of reconfigurability and extendability of the hardware Furthermore it would have conflicted with our overall design parameter 4 of using off the shelf components to the extent possible All connections were made using TBX 68 a DIN rail mount screw terminal connector block from NI for accessability reconfigurability and debugging purposes 1 Sensors a Pressure Air path adds latency to any pneumatic system Since our actuators are driven using off site high flow valves pressure sensors were placed closest possible to the actuators to account for the pneumatic latency Length Length sensing capability was a difficult choice as it posed several challenges at multiple lev els Piston length sensing capabilities were added to cater to some of our recent efforts in the direction of bio mimetic tendon driven systems 15 In such systems it is particularity difficult to have joint angle sensors Access to tendon excursions helps with kinematic modeling 15 Size of Actuator unit The size of the actuation unit had to be increased to accommodate for length sensing without which the diameter of the hand muscle actu ation unit would have been comparable to the typical dimensions of a human forearm Number of Analog Input channels Due to hi
12. e joint co contraction Variable stiffness simulated springs facilitated control over the return force which would not have been possible if single acting cylinders were used For compatibility with length sensors magnetic cylinder pistons were used Fig 2 Cylinder unit AC Actuator LS Length Sensor PS Pressure Sensor Design evaluation and experience The actuator selection was the most critical and challenging selection of the entire design More than 16 models from 6 different manufacturers were rigorously tested over the listed design parameters Though all the considered models performed well on most deign parameters requirements 1 and 3 were exceptionally severe The models from Airpel significantly outperformed others on criteria 1 and 3 The stiction and friction values for these models were exceptionally small the piston fell under its own weight when the cylinder was not horizontal This is possible because the traditional pneumatic seals have been replaced with air seals with precision fit graphite pistons that slide freely without lubrication inside a Pyrex glass cylinder providing unique ability to impart smooth motion at very low pressures slow speeds and short strokes There is a small air leak of about 2SL min 7 however this is not an issue when using high flow rate valves To sum up the Airpel cylinder we selected rates well on all parameters except 4 and 5 on which it rates moderately B Housin
13. g assembly Design parameters The design parameters for the housing assembly are listed below in decreasing order of priority 1 Strength load bearing and stability 2 Minimum off axis actuator loads 3 Compactness of Hand muscles actuation unit 4 Weight 5 Ventilation 6 Compatibility with different hands 7 Ease of component assembly 8 Machinability Hardware details Figure 3 shows the final housing as sembly without the actuator unit The assembly contains 36 of the M9 Airpel actuator units for finger tendons and 4 of the M16 Airpel actuator units for wrist tendons Tendon driven robotic hands usually route finger tendons via the center of wrist joint in order to minimize the moment arms of finger tendons on the wrist joint As a result all finger tendons come out of the hand via an opening at the wrist To reduce off axis actuator loads it is desirable to mount the cylinders so that they all point to this opening suggesting a concave mounting plate The back plate is free of cables and connectors and has mounting holes for attachment to a robot arm It is compatible with the Shadow Arm robot which we have also redesigned with air cylinders but this will be described elsewhere Figure 4 shows the complete Muscle actuation unit without the back plate Note that if we did not attach a length sensor to each cylinder which doubles the cylinder diameter the diameter of the assembly could be reduced roughly by
14. gh impedance of length sensors we observed ghosting 17 In case of ghosting every sensor at i channel of DAQ gets coupled with the one connected to every i 1 channel Most acceptable and efficient way to eradicate ghosting is to interleave null ground channels between each sensor channels Null channel interleav ing completely eliminated ghosting but resulted in 2X number of DAQ channels 2 Chassis 2X AI channels were used for length sensors b to mitigate the effects of ghosting This requirement was met by replacing one PXIe6363 module with PXIe6255 module that has more number of channels but a lower sampling frequency This reduced the rate at which Hand actuation unit s length sensor can be sampled from 32KHz to 9Khz 3 Power supply Power is distributed via modular power strips Individual sensors actuation banks can be turned on off to minimize noise floor and system load when not in use Modular power distribution facilitates addition of different power supply units for individual sensor actuator submodules thus facilitating upgrades at all levels 4 Emergency Stop Control valves provide unreliable flow rates when pneumatic input is fed without operating power Unreliable flow rates can damage the robot by pushing it outside its stability regime or setting joints into oscillations Emergency module was designed to allows pneumatic flow input only when control valves are powered up Initially we considered a complete
15. half a b Fig 3 Housing assembly a CAD model b Final machined assembly Design evaluation Machining the housing assembly and in particular the curved plate turned out to be harder than it first appeared but was eventually successful It weighs 660grams and can sustain about 75N from each actuator with a factor of safety 3 When the actuators and the ShadowHand are mounted the entire system weighs 4 5kg When attached to a robot arm most of this mass is near the base elbow thus we do not expect it to be problematic VI PNEUMATIC CONTROL UNIT Design Parameters For the pneumatic control unit the following design parameters in decreasing order of priority Fig 4 Muscle Actuation Unit a CAD model b Final assembly without back plate FP Front Plate HM Hand Mount LS Length Sensor AC Actuator PS Pressure Sensor PC Pneumatic Connectors were established 1 High flow rate 2 High update frequency 3 Minimum pneumatic latency 4 Minimum pneumatic bottlenecks and avoid air pockets 5 Independent of type of actuation unit 6 Modularity 7 Facilitate modification and accommodate extension 8 Compact and light weight 9 Cost Fig 5 Pneumatic Control Unit 2ft x 2ft x 2ft CH To chassis BP Breakout Panel CB Circuit Breaker PE Pneumatic E stop AF Air Filter MF Air Manifold HD Hand PC Pneumatic Connections HM Hand Mount HA Hand muscle actuation unit VL Valve ES E Stop CM
16. n The actuation system s speed capa bilities were evaluated using a simple open loop bang bang control strategy over the index finger MCP joint The goal was to achieve full stroke movements joint limit to joint limit at maximum frequency Control switching frequency was gradually increased until finger started making incomplete strokes i e reversed before hitting the joint limits Using this simple strategy a frequency of about 7Hz was achieved We are working towards a more principled way to further improve actuation speed by carefully modelling valve and pneumatics of our system Further details are provided in 18 C System latency and event timings System responsiveness was evaluated using a reflex experi ment During the experiment small external disturbances were applied at the finger tip of the middle finger The system was programmed to detect the disturbance using the joint angle sensors in the ShadowHand and react by extending the index finger The system was found capable of detecting minute dis turbances and reacting very quickly One can consider multiple definitions of response latency The first change in pressure is observed 8 msec after the disturbance this corresponds to reflex latencies defined in terms of muscle activity in the biological motor control literature See Fig 8 When measured in terms of the resulting motion the latency is around 29 msec D video attachment A supplementary video is attached
17. oard mounting means that there is an air line between the valve and the cylinder and the longer the line is the slower the actuator becomes However for the combination of line lengths flow rates pressures and cylinder volumes used here we were able to achieve maximum force in the cylinder around 10 msec after sending a command to the valve which is faster than the response of human hand muscles to neural input Another deviation from the ShadowHand design is that we attached a linear magnetic sensor to each cylinder Even though the ShadowHand has a joint angle sensor in each dof we reasoned that sensing the cylinder positions directly is useful for avoiding tendon slack and also calibrating the tendon moment arms and that other hands may not have joint angle sensors indeed we have already built an alternative hand UW Hand 5 which falls in the latter category Overall our philosophy was to design and build the best performing actuation system we could afford on a budget of 60 000 This includes a National Instruments PXI systems allowing us to sample 48 pressure sensors at 32KHz and 48 linear position sensors at 9KHz and average within batches resulting in very low noise measurements Later in this paper we discuss options for building a less expensive system with similar performance III DESIGN OBJECTIVES We envisioned our system to be a general purpose actuation module capable of actuating most pneumatically driven robot
18. p Design parameters 1 Sensors a Observability of entire state of the system b Resolution bandwidth and range c Compatibility with generic pneumatic actuators d Stand alone compact and light weight 2 3 4 e Flexibility and ease of replacement Chassis a Minimum latency b Sensing resolution bandwidth and range c Data bandwidth for communication with computational unit d Support for multiple communication protocols e Expandable and reconfigurable f Support for different operating systems g Compact enclosed and safe Power Supply a Compatible with sensor and actuator requirements b Clean and reliable c Peak load capacity d Electromagnetically decoupled output channels e Overload protection over voltage protection and short circuit protection f Indication monitor Emergency Stop a Minimum latency b User and hardware safety Hardware details 1 2 3 4 Sensors a Pressure Electronics unit contains 48 solid state SMC pressure sensor 40 sensors are housed by Hand mus cle actuation unit and rest 8 by Arm muscle actuation unit b Length Electronics unit contains 48 Sick magnetic piston length sensor 40 sensors are housed by Hand muscle actuation unit and rest 8 by Arm muscle actuation unit c 3D tracking system consists of active marker motion tracking system from PhaseSpace 9 Chassis National Instrument s 9 Slot 3U PXI Expres
19. re stretched meaning that even when the system is inactive there is a lot of passive co contraction Since the tendons unavoidably slide over multiple surfaces putting them under tension causes so much friction that the compliance advantage of pneumatic actuation is lost Another example where passive tensioning increases friction is the ACT hand where the electric motors need to be augmented with springs so as to prevent the tendons from slipping off pulleys the ShadowHand does not use pulleys The solution then is simple replace the air muscles with air cylinders which do not need tensioning One caveat here is that most cylinders have pneumatic seals with unacceptable friction levels However we found a cylinder AirPel 6 that is effectively frictionless weighs 46 grams and generates 42 N of linear force at 100 PSI Second the ShadowHand uses small inexpensive valves mounted on board Such valves have insufficient flow rate resulting in sluggishness that matches the reputation pneumatic systems have in robotics Yet nowadays one can find qualita tively better valves fast proportional and with high flow rate We selected the FESTO MPYE series although there may be other comparable choices To be clear there is no perfect valve at the moment Apart from the higher price high quality valves are large and must be mounted off board This alone is not an issue given that the compressor is also off board but off b
20. ronments than what is currently possible they are likely to need manipulators approaching human levels of dexterity The specific motivation behind the work described here was somewhat accidental We purchased a ShadowHand with the goal of developing and testing advanced control schemes for object manipulation After experimenting with it briefly we concluded that at least for the unit we received the actuation needs to be substantially faster and more compliant in order to support dexterous object manipulation We then disconnected the actuators air muscles and observed that when we pulled on the tendons manually the resulting finger motion was very fast and compliant Thus we decided to return the built in actuation system and develop an alternative Upon further reflection it became clear that such a development is very much needed in the field Indeed there exist multiple tendon driven hands including the Utah MIT hand 2 the ACT hand B as well as cadaver hands 4 that are comparable to human hands kinematically yet the lack of suitable actuation has hindered control applications Furthermore 3D printing technology has made it surprisingly easy to design and build new hands and possibly other mechanisms with large numbers of joints and tendons attached to them see below The question then is The authors are with the Departments of Computer Science amp Engineering and Applied Mathematics University of Washington WA 98195
21. s PXTe 1078 module with 1GB s is configured as Chassis Each slot has a bandwidth of 250MB s Present con figuration of modules sample 40 hand length sensors at 9kHz and 48 pressure and 8 arm length sensors at 32kHz Chassis is also equipped with Imbps CAN module which is used to communicate with ShadowHand sensors Computational unit uses one 798MB s bandwidth PXIe PCle data channel for complete control over the chassis Detailed specifications can be found at 10 Power Supply 24 V DC 10Amps S8VS 24024A switching power supply from OMRON is used as power source One electromagnetically separated channel powers the sensors while the other powers the Pneumatic control valves Emergency Stop A hybrid combination of software and pneumatic e stop is used as an Emergency unit Pneu matic stop has a flow rate of 6 500 l min at 6 bar and weighs 600gms Design evaluation and experiences TABLE I SENSOR SPECIFICATIONS SEDA ms Measuring Range Operational Voltage 32 256mm 15 30 VDC ripple P P 10 0 10 VDC 0 05mm 0 3mm 0 1mm 1 ms 25mA 2kQ N A 3m s 0 1 Mpa 12 24 VDC ripple P P 10 1 5 VDC lt 2 lt 0 7 lt 0 2 N A lt 15mA 1kQ 4 6g without wires N A Analog output Resolution Linearity Repeatability Sample Time Current consumption Output impedance Weight Max Speed Since the intended applications of our system was towards research applications there was never a preference towar
22. s with an added emphasis towards tendon driven systems The intended use is limited to research settings at the present level of development Design choices listed below in decreasing order of priority were made while designing the system 1 Compatibility with requirements of most pneumatically driven robots with focus towards tendon driven systems 2 Minimum hardware bottlenecks 3 Maximum computational capability 4 Use of off the shelf parts 5 Modularity in design at all levels 6 Accommodate extensions and facilitate modifications at all levels 7 Intended use under research settings 8 Safety 9 Weight 10 Cost IV DESIGN The hardware Fig Ip consists of the following modules 1 Actuation unit 2 Pneumatic control unit 3 Electronics unit 4 Computational unit Out of the four units only the actuator unit is specific to the robot being driven by the system Presently this unit is configured for tendon driven hands However it should Computational Unit Electronics Unit Pressure ensor CON J a oe Actuator Unit Air Preparation Unit Pneumatic Control Unit Actuation Unit Fig 1 Hardware design schematics Air pathway is represented using blue lines Electrical pathway is represented using purple lines Digital data pathway is represented using pink lines be noted that the actuator units bear no restriction towards tendon driven systems All actuation mechanisms
23. solve complex dexterous manipulation problems We emphasize that we are not building hardware for the sake of building hardware Indeed the primary focus of our research group is control see for examples of recently developed control schemes applicable to complex robots If the robotic hardware we need already existed we would be more than happy to focus on using it and making progress in terms of control Unfortunately suitable hardware in terms of robotic hands does not appear to exist with the possible exception of the DLR hand which is not available commercially and so we were forced to develop the system described here This development is now complete and we are ready to make a transition to control experiments We are also finalizing the re design of the ShadowArm robot with similar actuation and will soon be able to mount the hand assembly on the new arm Once the entire system is functional and used in specific manipulation experiments we will be able to characterize its capabilities in context But the basic tests performed here already illustrate that the system is very capable We hope that other research groups as well as commercial entities will be interested in building similar actuation systems This paper contains a lot of technical details that should help in replication efforts and we are happy to provide additional details upon request XII ACKNOWLEDGEMENTS This work was supported by the US National Institutes of
24. tsuoka and E Todorov Design of an anthropomorphic robotic finger system with biomimetic artificial joints in IEEE BIOROB 2012 16 National Instruments Company Knowledgebase 318ietlo How do i eliminate ghosting from my measurements 17 National Instruments X series user manual Multichannel scanning considerations 18 T Yuval T Wu J Movellan and E Todorov Modeling and identifi cation of pneumatic actuators manuscript under review 19 I Mordatch Z Popovic and E Todorov Contact invariant optimization for hand manipulation in Eurographics ACM SIGGRAPH Symposium on Computer Animation 20 Y Tassa T Erez and E Todorov Synthesis and stabilization of complex behaviors through online trajectory optimization 21 E Rombokas M Malhotra E Theodorou E Todorov and Y Mat suoka Tendon driven variable impedance control using reinforcement learning 2012
25. us confirming the exceptional compliance of the final system Typical characteristic force behaviours are summarized in Table II amp TABLE II ACTUATOR FORCE CHARACTERISTICS Specification No load connected Orientation Force jopnon pe Minimum external force to break Horizontal 2 5g ston and ition Piston falls under its own weight Minimum external force to break vertical stiction and friction TABLE III HAND FORCE CHARACTERISTICS Specification Finger and actuators oriented vertically Shadow UW Minimum actuation force at finger tip to move MCP 4 0g 2 0g jin t aim presao oo o o Minimum actuation force at finger tip to move MCP jin amin sack comeston presi S Maximum flexing force at Index finger tip 300 5g 705g Maximum extension force at Index finger tip 439 49 700g 90 fe Oo Pressure PSI Pressure PSI oo fe 0 30 60 90 Time ms a b 60 Time ms Fig 7 a Pressure behaviours while pressuring cylinder from zero using valve command V b Pressure behaviours while exhausting cylinder from zero using valve command V B Actuation Speed Our prime motivation for the development of this actuation system was to use tendon driven hands and perform dexterous hand manipulation experiments Any dexterous hand manipu lation demands agility and responsiveness from its actuation hardware These capabilities are evaluated in the present and the following sectio
26. with the manuscript The video demonstrates speed reflex and compliance properties of the actuation system with ShadowHand skeleton Speed behaviour is demonstrated using a sequence of flection and extension of joints limit to limit one at a time Entire sequence of flexing and extension for all 24 joints merely takes 2 44 seconds Each movement is roughly 70 milliseconds Hand reflex is demonstrated using the experiment mentioned in subsection above System compliance is demonstrated using 3 experiments First we demonstrate that actuation yields away to air blow from an average adult Second dead weights of lg 2g and 5g were dropped on the finger from a height of 5 cms to show compliance Third we demonstrate the effects of gentle interactions from an average adult 7s ETN Dl Index MCP Extensor Pressure A Middle MCP Joint velocity Purturbation Index MCP Joint velocity Reflex Time ms Fig 8 Time stamps from left to right T 1 Event Trigger Middle finger MCP movement detection 18 179ms T7 2 Actuation voltage written to valve 22 084ms 7 3 Pressure wave arrival 24 044ms T 4 Index finger MCP movement detected 46 943ms T 5 maximum pressure at the actuator 47 55 X HOW TO MAKE THE SYSTEM LESS EXPENSIVE The approximate cost breakdown of our present system is as follows Component Unit price Toa FESTO valves SICK sensors AirPel cylinders Pressure sensors NI PXI s
27. ystem Custom machining The most expensive components and thus primary candi dates for simplification are the valves linear sensors and electronics The PXI system we configured is an overkill considering that the sensors have low noise to start with so the high sampling rates and mini batch averaging are not essential and the bandwidth of the valves is only 125 Hz so there is no point in having a fast control loop indeed we are only using 200 Hz The National Instruments system has very mature drivers and is overall a great choice but one could build a considerably less expensive replacement for the present purposes perhaps using multi channel A D chips on custom circuit boards mounted in the forearm The biggest potential for savings are in the valves and position sensors Presently we use one valve and one sensor per cylinder This results in a universal pneumatic drive which can be used to actuate any mechanism with up to 40 tendons Note however that 40 is actually quite a lot and is only needed when using so called 2N designs where tendons act on individual joints and are arranged in agonist antagonist pairs as opposed to the more distributed action found in the human hand and in the ACT hand If we are willing to assume that most or all tendons will always operate in agonist antagonist pairs we could use one valve and one position sensor per tendon cylinder pair note that we still need all cylinders because even in this
Download Pdf Manuals
Related Search