Identifying Perceptually Indistinguishable Objects
Contents
1.2. HeadedToward other known PIO P gt return true if there was another PIO headed in the same location then the agent isn t sure else return unknown If none of the above there is no evidence that the two entities represent the same object and there has only been enough time to travel a short distance so assume that the two are not the same return false Conclusions and Future Work We are currently implementing this theory in a simu lated embodied robotic agent The base case of iden tifying immobile objects is currently implemented and we are working on the other base cases and expect to begin implementing the intermediate cases and the gen eral algorithm soon We still need to formalize and im plement some of the the support functions that we have assumed such as Disallow This paper has described a computational theory and algorithm of identifying an object which is perceptually indistinguishable from one seen before as either being the same as the one before or as being different The theory is built using the results of experiments with hu man subjects who did the same task It is cognitively plausible and designed to produce the same successes and failures as humans performing the same task References Brachman R J and Levesque H J eds 1985 Read ings in Knowledge Representation San Mateo CA Morgan Kaufmann Coradeschi S and Saffioti A 2003 An introduction to the anchoring problem Robotics
3. cabinet and the fancy china plate inside it as being at the same location however you cannot have two plates at the same location the one proposed in Lozano Perez amp Wes ley 1979 Once the path is available the length of the path is calculated from the lengths of the path s straight line segments Time E returns the time when the entity E was last encountered PIO Identification Functions Below is an algorithm based on human subjects experi ments for recognizing if a currently perceived object is new or is the same object as a PIO that was seen earlier The algorithm is given below as four functions writ ten in pseudo code The four functions presented are Recognize IdentifyNonMovingObjects IsSame and Is SameShortD The function Recognize takes an object description a set of positions of all of the current sightings of objects with that description the time of sighting and a possi bly empty set of entities representing objects with this description that have already been encountered The function returns a set containing mental entities corre sponding to each object currently seen Due to the si multaneous perception base case there will be one en tity returned for every element in the set of places function Recognize D P T E returns a set of entities inputs D description of an object in the world P set of places that agent currently perceives objects that have description D T the time of the
4. identify ing PIOs Continuous viewing like location is used to identify an object as being the same as a perceptually indistinguishable object seen earlier Pollock s reiden tification This ease of identification of object while under continuous observation seems to be implicitly as sumed in Coradeschi and Saffiotti s Coradeschi amp Saf fioti 2003 Track functionality More concretely the continuous viewing case applies if an agent views an object at position p and later ob serves an object that is perceptually indistinguishable at position po If the agent has continuously viewed the object as it moves from p to p2 the agent may assume with great certainty that the object it is currently seeing at p is the same object that it originally saw Human subjects tried to use this base case as of ten as possible when asked to follow a virtual robotic tour guide through a suite of rooms that also contained several perceptually indistinguishable robots serving as distractors Subject 7 after an early bit of difficulty says And I am following him very closely And I am not going to lose sight of him this time Subject 23 is also very specific about using continuous viewing So I m just staying uh close to this robot keeping my eye on him Intermediate cases of PIO identification What makes an intermediate case It has been pointed out that the base cases described in section represent primarily perceptu
5. perception and E set of previously created entities that have description D eSet First check the case of no PIOs first time the agent sees something with description D if IEl Of for each P P eSet eSet MakeEntity D P return eSet Next check the base case of unique objects and the intermediate case of those unique in the current context if IPI 1 amp IEl 1 amp IsUnique D V IsUniqueInContext D return E Now check the base and intermediate cases arising from objects that are not moving if NotAutoMobile ClassOf D return IdentifyNonMovingObjects D P E Next check the base case of continuously viewed objects for each P P En Continuously ViewedPIO D E P if not the base case consider the intermediate case of continuously perceived objects if Ep null En ContinuouslyPerceivedPIO D E P if En null eSet eSet En P lt P P For each remaining non base case object make a new entity and reason if that entity is the same as something seen before for each P P En lt MakeEntity D P for each Em E if IsSSame Em E Loc Em P Time Em T Believe equiv Em En RemoveEquivs E E eSet eSet En return eSet function IdentifyNonMovingObjects D P E returns a set of entities inputs D description of an object in the world P set of places that agent currently perceives objects that have description D E set of previously created entities
6. that have description D In this function the agent considers those objects which cannot move themselves if Immobile ClassOf D V pi OtherPossMover ClassOf D for each P P En PIO AtLoc ClassOf D E P if E null eSet eSet En else eSet eSet MakeEntity D P return eSet The function IsSame checks to see if two entities repre sent the same object Three return values are possible A return value of true means that the agent believes that e2 A return value of false means that the agent believes that e 4 e2 A return value of unknown means that the agent does not believe it has enough informa tion to decide whether e1 e2 If the agent must make a decision with an answer of unknown it will probably have to select randomly function IsSame Ey E2 Py Po Ty T2 returns three value boolean inputs E E2 Two entities P1 Po the place where each was perceived T1 Tp the time of each perception ratel RateOf E rate2 RateOf E gt Since the assumption is that an object s speed is constant if the rates of speeds differ then the objects differ if ratel rate2 return false if the agent doesn t know the speed of the objects in this non base case situation the agent cannot make a good decision about the identity of the objects if ratel unknown return unknown next check to see if an object with the known speed could have traveled from its previously known positio
7. the world that might or might not be the same as a previously perceived object in the world Its task is to decide whether the entity eg think of the author of Waverly corresponding to the newly perceived object is coreferential with an entity e1 think of Scott that corresponds to a previously perceived object When an agent wants to identify an object it must ac complish two things First the agent must identify what kind of object it is sensing The agent should use its sen sors and its understanding of what things look like to those sensors to identify the type or kind of thing that it the agent is looking at The agent must then reason about what actual object it is looking at A simple so lution and one easy to implement might be to assume that all objects that look the same are in fact the same object but this is clearly not the case A better solu tion discussed in Shapiro amp Ismail 2003 is whenever an agent looks for an object with certain properties it conceives of a new entity with only those properties When the agent finds a real world object that has those properties it should recognize if it already has a men tal entity corresponding to the object it just found If it does have such an entity then it should adopt a be lief that the object looked for is the same as the one that was found This approach has two drawbacks First it sidesteps the issue of how the agent reasons about ob
8. How did he come up that way The subject clearly seems to be using the same mental entity for both robots and believes that there is only one robot 3We are ignoring the use of illusions with mirrors and other deliberate attempts to make a single object appear to be multiple objects Base case 3 Immobile objects Immobile objects are defined here as those objects which cannot move or be moved We re also including those objects which humans expect cannot be moved even if such an object might be moved by using a rarely used technique For example people do not expect things like houses and other buildings or even large trees to be moved intact from one place to another even though it is possible Since the location of an immobile object does not change location is the most important feature which al lows an agent to identify immobile PIOs In order to identify an immobile PIO the agent must first recognize what kind of object it is seeing Then the agent needs to reason or realize that objects of this kind are immobile Then the agent cognizes the location of the object At this point the agent can identify the object Either the agent knows about an immobile object of this kind at this location in which case it now identifies the current object using the entity that denotes that previously seen object or the agent has never encountered one of this kind of object at this location in which case the agent identifies the objec
9. Identifying Perceptually Indistinguishable Objects John F Santore and Stuart C Shapiro Department of Computer Science and Engineering and Center for Cognitive Science University at Buffalo The State University of New York 201 Bell Hall Box 602000 Buffalo NY 14260 2000 jsantorelshapiro cse buffalo edu Abstract This paper describes a cognitively plausible computa tional theory of identifying perceptually indistinguish able objects PIOs based on a set of experiments which were designed to identify the knowledge and perceptual cues that people use for this purpose Identifying a PIO in this context means connecting sensor data from some physical object either to a new mental level symbol or to the correct preexisting one and is part of the solution to the symbol anchoring problem We discuss several base cases in the identification process some related in termediate cases and the knowledge that is needed for the general case An algorithm for identifying PIOs is included Introduction People often encounter objects that are perceptually in distinguishable from objects that they have seen before When this happens how do they decide whether the ob ject they are looking at is something never before seen or if it is the same one they encountered before How should an agent a person or a robot identify a percep tually indistinguishable object PIO Identifying a perceptually indistinguishable object means deciding if th
10. al cases of identi fying PIOs and that there were likely to be simple cases that do not rely on purely perceptual mechanisms for the identification of PIOs In looking at our experimental data we found evidence of non perceptual cases that are similar to the base cases Each perceptual base case had one non perceptual simple case which can be closely identified with the base case We call these associated non perceptual cases intermediate cases They are so named because they are between the largely perceptual base cases and a mostly cognitive general PIO identifi cation mechanism Intermediate Case 1 rapid perceptions The first intermediate case is related to the base case of simultaneous perceptions In that case seeing mul tiple objects at once was sufficient to assure that there are multiple objects in the world In the rapid percep tions case on the other hand the objects usually two of them are not perceived at the same time but rather in rapid succession with no PIO encountered between the two perceptions As in the case of simultaneous percep tions the rapid perception case is used to prove to the agent that two objects are not the same Participants in the experiment sometimes used rapid perceptions to disprove a hypothesis of unique appear ance as Subject 18 does in the following transcript ex cerpt Going into the next room there is a multi colored robot and one who looks like the last one I m turnin
11. and Autonomous Systems 43 2 3 85 96 Feigl H and Sellars W eds 1949 Readings in Philosophical Analysis New York Appleton Century Crofts Feigle H and Sellers W eds 1949 Readings in Philisophical Analysis New York Appleton Centuray Crofts Frege G 1892 On Sense and Nominatum 85 102 Reprinted in Feigle amp Sellers 1949 Jain R Kasturi R and Schunck B 1995 Machine Vision New York McGraw Hill Johnson S P 1998 Object perception and object knowledge in young infants a view from studies of visual development 21 1 239 Printed in Slater 1998 Kuipers B and Beeson P 2002 Bootstrap learn ing for place recognition In Proceedings of the Eigh teenth International Joint Conference on Artificial In telligence IJCAI 02 174 180 San Francisco CA Morgan Kaufmann Kuipers B and Byun Y T 1991 A robot explo ration and mapping strategy based on a semantic hier archy of spatial representations Journal of Robotics and Autonomous Systems 8 47 63 Lehmann F ed 1992 Semantic Networks in Artifi cial Intelligence Oxford Pergamon Press Lozano Perez T and Wesley M A 1979 An algo rithm for planning collision free paths among polyhe dral objects Communications of the ACM 22 10 560 570 Maida A S and Shapiro S C 1982 Intensional concepts in propositional semantic networks Cogni tive Science 6 4 291 330 Reprinted in Brachman amp Levesque 1985
12. d into that last room that I just searched but it looks like there are two multi colored robots Human subjects also use information from observa tions of the specific objects being identified Beliefs formed from these observations include beliefs about where and when the agent last encountered a PIO that the subject believes might be the PIO that the subject is currently looking at Subject 25 counting robots with two distinct groups of perceptually indistinguish able robots I am entering the third room 5 I can find the third robot but I guess this is the same one as the first one but the room is different Another be lief formed about the object itself is the answer to the question Does the object appear to have a particular purpose or motivation and if so what is it Subject 10 following a tour guide There are a total of three robots in here now But and they seem to be mov ing randomly The direction or trajectory that the ob ject is moving in is important when an agent is trying identify a PIO only a relatively short time after encoun tering another PIO Subject 18 following a robot He hasn t changed directions so I can still tell which one is him It is also important for the agent to have some awareness of where other PIOs are in the area to make sure that it doesn t get unnecessarily confused if the ob ject it is focusing on moves too close to one of the oth ers Subject 23 foll
13. e object just encountered is a new never before seen object or if it has been previously encountered which previously perceived object it is This task is required in both the Track and Reaquire functionalities that Coradeschi and Saffiotti 2003 de fine as part of a solution to the symbol anchoring prob lem Symbol anchoring is defined as the process of creating and maintaining the correspondence between symbols and sensor data that refer to the same physi cal objects The Track functionality assumes that the object being tracked is continuously observed by the agent This continual observation will certainly make the task of identifying the object and anchoring its de scription to the appropriate mental symbol easier The Reaquire functionality in particular however needs a complete theory of identifying PIOs Reaquire is the case in which an object is reobserved after some time Copyright 2004 American Association for Artificial Intel ligence www aaai org All rights reserved Coradeschi amp Saffioti 2003 p 91 In order to know that the same object has been reobserved after some time rather than a new object that looks just like the old one the Reaquire functionality requires a mecha nism for identifying PIOs Identifying a PIO is a sub problem of object identi fication rather than object recognition Object recogni tion is defined in a computational vision textbook Jain Kasturi amp Schunck 1995 a
14. en the computational agent identifies a perceptu ally indistinguishable object using a base case it does not form a new mental entity for the object and then try to find an existing entity with an equivalent exten sion The agent only creates new entities as needed for Perceiving could be done using any sense but in this pa per we will often use looking as a euphemism for any type of perceiving gt We are currently preparing a paper describing the results of these experiments cognizing information Maida amp Shapiro 1982 The object that the agent is perceiving is either the one that it has seen before or a new object If the object is the one then the agent ought to use the original mental en tity for it and not conceive of something new which the agent then believes is really the same thing in the world If the object is a new one a new mental entity is created for the new object in the world that our agent conceives of Each of the four base cases is described in its own subsection below Base case 1 Simultaneous perceptions If an agent perceives two perceptually indistinguishable objects in its sensory field at the same time the agent can trivially conclude that the two are not the same object Unlike some of the base case strategies sub jects were conscious that they were using this strategy of simultaneous perceptions while they used it While counting moving robots Subject 37 states I m trying to
15. g back that robot is still in the other room so I know that these are two distinct robots Prior to this excerpt Subject 18 has seen only one robot a silver gray robot As he enters another room Subject 18 sees a multi colored robot as well as a silver gray Our thanks to the anonymous reviewer who did so robot In order to identify this silver gray robot as a new never before seen robot Subject 18 looks back to ward the place where he last saw a silver gray robot When he sees a silver gray robot in the previous loca tion as well Subject 18 assumes correctly in this case that the robot seen in the current room is different than the one he looked back to see In order to take advantage of this rapid perceptions case an agent must see an object Q then must turn at least as fast as objects of type Q can move turning no more than 180 and must see another object that looks like Q If all of these conditions hold the agent can determine with a very high degree of confidence that there are two different PIO objects in the world Intermediate Case 2 Locally Unique Objects An agent can often easily identify an object without the object being truly unique in the world or even believed to be by the agent The agent must believe that an ob ject is unique in the current context For example if you know identical twins but one of them is in the army posted abroad for the next six months if you see some one that looks l
16. ike these twins in town tomorrow you can immediately assume that you see the second twin Of course the simultaneous perceptions base case de scribed above will trump a belief that an object has a unique appearance Subjects seemed to use this assumption that unique objects can be effortlessly identified as the same mental entity when they could Sometimes the assumption of uniqueness of appearance would be limited to a single room Subject 12 while following a robotic tour guide in a suite of rooms with PIOs as distractors says I m stuck ok but there is only one robot so I can follow it Subject 23 doing the same task says something similar There aren t any other robots in this room so it s a little easier to follow Intermediate Case 3 Stationary Objects The next intermediate case is related to the base case of immobile objects Stationary objects are those ob jects which cannot move themselves and are not easily moved by a breath of air A helium filled balloon is not a stationary object even though it can not move it self On the other hand many of the objects that we come into contact with in our daily lives are stationary lamps computers textbooks and similar objects are all stationary objects Their position will not change or at least people do not expect it to change unless there is an animate object to move the stationary object sub ject 31 explicitly pointed this out in a retrospective after count
17. ile counting immobile objects ahhh ok it seems to me I ve already seen this room and counted this one glass here Earlier it was noted that the use of a single entity is contingent on an agent correctly identifying its cur rent location Our subjects were vulnerable to mis taking one room for another if the two looked simi lar Kuipers and his colleagues Kuipers amp Byun 1991 Kuipers amp Beeson 2002 call this sort of mistake per ceptual aliasing and have discussed the problem and a solution for robotic agents When our subjects fell vic tim to perceptual aliasing use of location information to identify immobile objects was fallible Sometimes sub jects would notice the possible aliasing such as Subject 20 while counting glasses who says I m just just curi ous to whether or not this is the same room So I m go ing to go back and retrace that my steps Subjects who fell victim to perceptual aliasing and never realized it generally failed at the counting and identification tasks Base Case 4 Continuous viewing Pollock 1974 has discussed reidentification of objects a subproblem of identifying PIOs He notes that an ob ject under continuous observation can be reidentified at a later time as being the same object in particular that continuity of appearance is a logical reason for reiden tification Continuous viewing of an object also appeared in the human subjects trials as a base case for
18. ing glasses in task of the experiment Experimenter What strategies did you use to do this task Subject 31 Mmm I guess I just kind of based it on the fact that they would be stationery through out the rooms and there was nobody else in there In the absence of a mover stationary objects can be treated just like immobile objects that is location be comes the paramount criterion for identifying the ob ject The lack of another agent capable of moving a sta tionary object is something that a PIO identifying agent must reason about Intermediate Case 4 Continuously Perceived Objects It is well known Johnson 1998 that young children will identify object that are briefly occluded as the orig inal objects Our subjects overwhelmingly did likewise Though subjects may have briefly lost sight of the focus object by looking away or having the object occluded the subjects nonetheless knew where the object was and looked for it where it ought to be when they viewed the object again Most of the time subjects were not even aware that they had lost sight of the object in ques tion Identifying PIOs in general While identifying PIOs is trivial and intuitive when one of the base cases can be applied and only a little more difficult in one of the intermediate cases when one of the base cases does not hold the task can be much harder An agent usually requires several more pieces of knowledge to identify mobile object
19. ity E has OtherKnownPIOs E E gt returns true if the agent knows about other entities that are percep tually indistinguishable from the two enti ties E Ey and which are not coreferential with either otherwise returns false OtherPossMover Class returns a possibly empty set of all of the entities that the agent knows are in the area that could have moved something of the class Class PIO AtLoc Class ESet Loc returns Ifan object of Class has been seen at location Loc be fore then the entity from the set of enti ties ESet denoting that object is returned otherwise null the non entity is returned Note The agent may still fall victim to perceptual aliasing Kuipers amp Byun 1991 Kuipers amp Beeson 2002 and so if the agent is mistaken in its beliefs about the lo cation then the agent is likely to reach the wrong conclusion about the identity of the object RateOf E returns the speed or rate of movement of the entity E RemoveEquivs E ESet effect removes from the set of entities ESet all entities that are coref erential with entity E ShortestKnownPathBetween P P2 returns the length of the shortest route that the agent knows about between the two positions P and Py This function is done at a sub cognitive level The path is calculated by using a simple path planning algorithm like Only one object of a particular class can be at a single location at any one time One might think of a
20. ject identity Second even though the agent may now correctly believe that the two entities refer to the same object in the world there are times when a new entity is unnecessary It would be better to simply use the agent s original entity if at the time of the second sighting the agent can instantly reason that the object is the same as one it has seen before In the remainder of this paper we will discuss four base cases in the identification of PIOs and then intro duce the knowledge used in more complex cases The base cases and the more general knowledge are all based on evidence drawn from protocol analysis style exper iments done with human subjects to see how humans identify PIOs We will then give an algorithm for iden tifying a currently perceived object as being the same or different from any PIOs that the agent has been encoun tered previously Finally some conclusions drawn from the work so far are discussed The base cases in the identification of perceptually indistinguishable objects What makes a base case Experiments with human subjects showed that there are four conditions under which human subjects find the identification of perceptually indistinguishable objects to be very easy We ll call these four conditions the base cases of the identification task Subjects actively tried to put themselves into a position where they could use one or more of these base cases to identify the PIOs in the experiment Wh
21. mbers of Class cannot move on their own ClassOf Desc returns the class of objects that look like Desc ContinuouslyPerceivedPIO Desc ESet P returns an entity from the set of entities ESet The returned entity corresponds to the object with the description Desc that the agent has perceived as being the same from some previous place to the place P This func tion is similar to Continuously ViewedPIO see below however there is no require ment that the object be viewed continu ously rather if sight of the object is lost for a few seconds the tracking mechanism will continue to track where it ought to be and continue to accept the object perceived there as the one being tracked If there is no such entity then the function returns null the non entity Continuously ViewedPIO Desc ESet P returns an entity from the set of entities ESet The re turned entity corresponds to the object with the description Desc that the agent believes it has viewed continuously from some pre vious known place to the place P If there is no such entity then the function returns null the non entity This function will rely on a FINST like Pylyshyn 1989 mecha nism for tracking continuously viewed ob jects Sandewall 2002 has implemented such a system using tracker objects and ac tivity demons to supervise the tracker ob jects Continuously ViewedPIO will return true if there is a tracker object that has con tinuously vie
22. n to the currently perceived position If not the two must be different objects possibleRange ratel T gt T1 if possibleRange lt ShortestKnownPathBetween P P2 return false If the agent knows a motivation of E which would prevent E from being at the place where E 2 is currently perceived then assume the two are different if Disallow MotivationsOf E P1 P2 T1 T2 return false If the distance that E could have traveled between sightings is greater than some environment specific large distance then the agent can t decide if possibleRange gt LargeD return unknown else Otherwise use a rule for identifying objects that might have only traveled a small distance return IsSameSmallD Eq E5 P5 T gt function IsSameSmallD E E gt P2 T2 returns three value boolean inputs Ej E gt Two entities Pa the place where the second was perceived T the time of the second perception If the agent knows about PIOs other than the two entities in question then if none of the others could reach the place of the second perception P3 by the time T gt then assume that E E if OtherKnownPIOs E E gt if A other known PIO suchthat CouldReach other known PIO P2 T2 return true Otherwise if the object represented by the first entity was headed toward the place of the second perception and there are no other PIOs known to be headed in that location E E gt if HeadedToward E P2 if A other known PIO suchthat
23. owing a robot So I just cut in front of that robot in order to keep following mine Suc cessful subjects like subject 23 would often keep some awareness of nearby PIOs and act to avoid occlusion of their focus object by other PIOs An algorithm Assumptions Let us suppose that our agent has seen an object with appearance d at location p at time t The agent needs to identify the object Either the object is a new one or it is one that the agent has seen before in which case the agent already has an entity representing the object When the agent already has one or more entities repre senting objects with description d the agent must rec ognize which entity if any represents the same PIO us ing reasoning We assume for simplicity that the object s speed if known will be constant If the agent doesn t know the speed of the object it will probably not be able to de cide if the object it is perceiving has been previously encountered or not Helper Functions The PIO identification functions assume that several support functions are available Believe equiv E E gt effect creates a new belief that E and E3 refer to the same object in the world 5When sequences of two or more dots appear inside of a quote from a subject it indicates that the subject gave a no ticeable pause at that point The number of dots indicates the length of the pause NotAutoMobile Class returns true if the agent be lieves that me
24. pp 170 189 Pollock J 1974 Knowledge and Justification Prince ton Princeton University Press Pylyshyn Z 1989 The role of location indexes in spatial perception A sketch of the finst spatial index model Cognition 32 1 65 97 Russell B 1905 On denoting Mind 14 56 479 493 Reprinted in Feig amp Sellars 1949 Sandewall E 2002 Use of cognitive robotics logic in a double helix architecture for autonomous systems In Beetz M Hertzberg J Ghallab M and Pol lack M E eds Advances in Plan Based Control of Robotic Agents Lecture Notes in Computer Science 226 248 New York Springer Verlag Shapiro S C and Ismail H O 2003 Symbol an choring in a grounded layered architecture with inte grated reasoning Robotics and Autonomous Systems Shapiro S C and Rapaport W J 1992 The SNePS family Computers and Mathematics with Applications 23 2 5 243 275 Reprinted in Lehmann 1992 pp 243 275 Shapiro S C and the SNePS Implementation Group 2002 SNePS 2 6 User s Manual Department of Com puter Science and Engineering University at Buffalo The State University of New York Buffalo NY Shapiro S C 1998 Embodied Cassie In Cognitive Robotics Papers from the 1998 AAAI Fall Symposium Technical Report FS 98 02 Menlo Park CA AAAI Press 136 143 Slater A ed 1998 Preceptual Development Visual auditory and speech peception in infancy East Sus sex UK Psychology Press
25. r reasoning from observations to beliefs such as Shapiro 1998 p138 has been provided The focus will be on reasoning with beliefs about the world in or der to identify PIOs Our agent s beliefs and reasoning are based on an in tensional representation Maida amp Shapiro 1982 In tensional representations model the sense Frege 1892 of an object rather than the object referent itself The terms of our representation language SNePS Shapiro amp Rapaport 1992 Shapiro amp the SNePS Implementa tion Group 2002 denote mental entities Some such entities are propositions others are abstract ideas oth ers are the agent s concepts or ideas of objects in the world This is important for the task of identifying PIOs because before the identification task is complete the agent may have two mental entities e and e2 that it might or might not conclude correspond to the same object in the world It is in a similar situation as George IV who wished to know whether Scott was the author of Waverly Russell 1905 p 108 Mental entities are the denotations of the symbols described by Coradeschi and Saffiotti 2003 as part of the symbol anchoring pro cess We will use object to refer to an object in the world and entity to refer to a mental entity that is the deno tation of a SNePS term The task is identifying per ceptually indistinguishable objects because the agent has perceived an object in
26. s able to quickly and easily decide that this is a different stamp Sometimes the task is not as easy Consider the case in which a woman puts her glass down on a counter at a cocktail party When the woman returns to pick her glass up again and finds more than one glass on the counter the woman will often find it difficult to decide which glass is hers Sometimes the woman is not able to decide with enough certainty which glass is hers even after thinking about it This paper proposes a theory of how agents particu larly artificial embodied agents such as robots can use reasoning to identify PIOs as well as humans do Let us first examine what is required to identify an object in the world An embodied agent gathers information about its world by observing the world with its sensors and us ing its effectors to move itself to a better observation point when necessary From its observations the agent forms beliefs about the objects in the world People use these beliefs in conjunction with their commonsense rules about the world to help them identify objects in the world Identifying PIOs relies entirely on this mech anism since there is no sensory information that will help to distinguish one PIO from another The designer of an artificial embodied agent must provide the agent a mechanism for both creating beliefs from observations and using those beliefs to reason In the remainder of this paper we will assume that a mech anism fo
27. s the process of finding and labeling objects in the real world based on known object models that is object recognition in computer vision is the process of deciding what category an ob ject belongs to By object identification we mean de ciding which individual object it is rather than deciding what category of objects it belongs to When an agent perceives an object it first uses its object recognition system to decide what category of thing it is then it uses its object identification routine to choose and an chor a mental concept to the object The object identifi cation system uses non perceptual properties and back ground knowledge to identify the object as being the same one that the agent perceived at some previous time or to identify it as something new that the agent has never thought about before This identification of ob jects across time is a necessary part of any solution to the symbol anchoring problem Sometimes identifying PIOs seems effortless Con sider the case in which a man has a pile of stamps of the same design in his drawer He opens the drawer takes out a stamp and puts it on an envelope and mails the envelope The next day he needs to mail another en velope and so needs a stamp He opens the drawer and takes out a stamp that looks just like the one used the day before The man never considers whether it is the same stamp even though there is no perceptual differ ence between the two stamps He i
28. s which are not continuously viewed If people need to identify an ob ject as the mental entity e experiments show that they use knowledge of how rare or common they believe ob jects that look like e are They will also use their beliefs about how fast the objects like e can move and the time between the time t that the agent last encountered an object it thinks might have been e and the time t2 that the agent sees e itself Humans will also use the mo tivations of the object being identified if they can infer any Humans subjects seem to use general beliefs formed from observations of the world The most salient is in formation about the class of objects to which the PIOs being identified belong These include things like how fast or slow do objects of this kind move Subject 8 while counting moving robots I think that s the guy I counted already because ah well he uh couldn t have moved that fast Has an object of this kind ever been known to change speed Subject 6 asked in a retro spective why subject chose to follow a particular robot It s possible that it changed speeds but it didn t really appear to do so throughout the game Have I ever iden tified more than one object that is perceptually indistin guishable from this one Subject 18 while counting robots in a condition with two distinct groups of per ceptually indistinguishable robots Because I thought maybe the multicolored robot had travele
29. see them simultaneously Subject 4 while doing the same task is even more explicit when stating The same two robots at the same time so I know that there are at least two robots here Base case 2 Objects with a unique appearance If the agent believes that an object has a unique appear ance and there are no distinct PIOs in the world then the agent can instantly anchor the object to its mental entity as soon as the agent perceives an object with that appearance The agent can and ought to use its original entity for the object in this case Subjects were often aware enough of their use of this assumption of unique appearances to try to verify the assumption when possible Subject 15 when counting robots when there were two groups of perceptually in distinguishable robots says And I see the clown dalek here aaand the little black and white one I don t annd a clown here is that the same clown The use of a single mental entity for an object be lieved to have a unique appearance was particularly noticeable when the subject s assumption that an ob ject has a unique appearance turned out to be incorrect While trying to follow a robotic tour guide who turns into a room on the left of a corridor Subject 42 says I can t catch up with you Where are you going And then a second later as a second robot emerges from a room on the right of the corridor a little further from the subject He came up that way
30. t as a newly encountered object and anchors a new entity to the object description Experiments with human subjects support the claim that location is of paramount importance in identifying immobile PIOs Human subjects find the use of lo cation information so intuitive that they rarely notice it at the conscious level When human subjects were asked to discuss what they were doing and why while counting immobile PIOs they never mentioned the ob ject s location as being important during the task even if they were clearly using location information How ever when asked in a retrospective interview subjects were able to articulate that location information was what they were relying on The following exchange is representative It was taken from a retrospective inter view following an experimental task in which subjects were asked to count glasses The glasses were immobile in this environment and recognized as such by subjects Experimenter how were you able to distinguish between the glasses even when they looked the same Subject 33 ah because they are lying in the dif ferent rooms That s why They are different An agent cannot use only an object s location to iden tify that object as the appropriate mental entity The agent must still recognize the object as belonging to the same class of perceptually indistinguishable objects as the previously encountered object There are two rea sons for this The first reason is
31. that any object might be destroyed a house might burn down a tree might blow down in a hurricane Since an object might be de stroyed and some other object take its place perhaps a gazebo in the place of an ancient tree the object it self needs to be identified as being perceptually indis tinguishable from the previously seen object The second reason not to rely on location alone is that location is only sufficient to anchor immobile objects So an agent must recognize the object as being of a class of objects that are immobile in order to take advantage of location as the distinguishing factor for anchoring an object The use of the original entity seems to be supported by the human subject data in the immobile object case as well While performing the glass counting task no subject who was sure about what room he she was in expressed doubt about the identity of a glass The glass was either referred to as the same one seen previously or it was referred to as a new glass This contrasts with mo bile objects where subjects often clearly seem to have more than one entity for an object and can talk about both entities For example when following a robotic tour guide in a suite with several perceptually indistin guishable distractors Subject 30 briefly loses the tour guide robot and then makes the following statement Where did the robot go I think this is the one In con trast Subject 55 makes the following typical statement wh
32. wed the object at P from some previous spot CouldReach E P T returns true if the agent believes that the object corresponding to entity E could have arrived at the place P by time T Disallow MSet P P gt T T2 returns true if the agent believes that any element of the set of motivations MSet of an entity would dis allow the entity from being at location P and time T given the agent s belief of a previous encounter at time T and location Py otherwise the function returns false HeadedToward E P returns true if the object cor responding to entity E was headed in the direction of the place P when the agent last observed the object otherwise returns false Immobile Class returns true if the agent believes that members of Class are immobile and false otherwise IsUnique Desc returns true if the agent believes that there is only one object in the world that has the appearance Desc otherwise the function returns false IsUniqueInContext Desc returns true if the agent believes that there is only one object in the current context that has the appearance Desc otherwise the function returns false Loc E returns the location where the entity E was last encountered MakeEntity Desc Loc returns a new mental entity to represent the object with the description Desc which is believed to be at location Loc MotivationsOf E returns a set containing any moti vations that the agent believes the ent
Download Pdf Manuals
Related Search