BUILDING PLANS FOR HOUSEHOLD TASKS FROM DISTRIBUTED

KNOWLEDGE

INVENTORS: RAKESH GUPTA AND CHIRAG SHAH

CROSS REFERENCE TO RELATED APPLICATIONS

[0001] This application claims priority under 35 USC § 119(e) to U.S. Provisional

Patent Application No. 60/697,843, titled "Building Plans for Household Tasks from Distributed Knowledge," filed July 8, 2005, and U.S. Utility Patent Application No. 11/296,020, titled "Building Plans for Household Tasks from Distributed Knowledge," filed December 6, 2005, which are hereby incorporated by reference herein in their entirety. [0002] This application is related to U.S. Patent Application Number 11/378,063, entitled "Commonsense Reasoning About Task Instructions," filed on March 16, 2006, which is incorporated by reference herein in its entirety.

[0003] This application is related to U.S. Patent Application Number 11/046,343, titled

"Responding to Situations Using Knowledge Representation and Inference," filed January 28, 2005, which is incorporated by reference herein in its entirety.

FIELD OF THE INVENTION

[0004] The present invention generally relates to the field of autonomous machines, and more specifically, to enabling mobile robots to perform tasks in constrained environments.

BACKGROUND OF THE INVENTION

[0005] Humanoid robots, for example, robots having human characteristics, represent a major step in applying autonomous machine technology toward assisting persons in the home or office. Potential applications encompass a myriad of daily activities, such as attending infants and responding to queries and calls for assistance. Indoor humanoid robots may be expected to perform common household chores, such as making coffee, washing clothes, and cleaning a spill. Additional applications may include assisting elderly and handicapped persons. Humanoid robots will be expected to perform such tasks so as to satisfy the perceived desires and requests of their users. Through visual and voice recognition techniques, robots may be able to recognize and greet their users by name. In addition, robots should be able to learn through human interaction and other methods. In order to meet these needs, such robots must possess the requisite knowledge.

[0006] In particular, to accomplish an indoor task, an autonomous system such as a humanoid robot needs a plan with steps. It is desirable that the robot be able to derive such a plan dynamically, that is, taking into consideration aspects of the actual indoor environment when the task is to be performed. It is also desirable to derive plans based on "common sense" knowledge. For example, steps for executing common household or office tasks could be collected from non-expert human volunteers using distributed capture techniques. [0007] Human actions have been analyzed on a variety of levels. At the most basic level, execution of an action can be represented by a motor response schema of sensory motor mapping. A description of this can be found in R. A. Schmidt, A Schema Theory of Discrete Motor Skill Learning, Psychological Review, 82(4):225-260, 1975, which is incorporated herein by reference in its entirety. At the most abstract level, concepts such as scripts and Memory Organization Packets (MOP) have been proposed to represent the organization of well-learned activities such as going to a restaurant or visiting a doctor for surgery. A description of this can be found in R. C. Schank and R. Abelson, Scripts, Plans, Goals and Understanding, Lawrence Erlbaum Associates Ltd., Hove, UK, 1977, and in R. C. Schank, Dynamic Memory: A Theory of Reminding and Learning in Computers and People, Cambridge University Press, Cambridge, 1982, both of which are incorporated by reference herein in their entirety. When a MOP is activated, only one step is generally carried out at a time, but steps can sometimes be combined with other activities. For example, one can read while waiting at doctor's office.

[0008] Between these extremes lies a range of practical, well-learned activities like making breakfast, cleaning one's teeth, dressing and so on. At this mid-level, Cooper and Shallice presented a computational model for selection of steps for routine tasks based on competitive activation within a hierarchically organized network of action schemas. Their activation model for sequential step selection was based on the Contention Scheduling theory of Norman and Shallice. A description of this can be found in Richard Cooper and Tim Shallice, Contention Scheduling and the Control of Routine Activities, Cognitive Neuropsychology, 17(4):297-338, 2000, and in D. Norman and T. Shallice, Attention to Action: Willed and Automatic Control of Behavior, pages 1-18, Plenum Press, New York, 1980, both of which are incorporated by reference herein in their entirety. The Cooper- Shallice model was demonstrated for the routine task of preparing coffee. Under normal functioning, the model was able to generate a sequence of simple actions (pick up spoon, dip spoon in sugar bowl, etc.) culminating in a drinkable cup of coffee.

[0009] In contrast, work in artificial intelligence (AI) planning falls under the category of goal-controlled exploratory behavior. Attempts are made to reach the goal using knowledge of different plans, and a successful sequence is selected. Such planning is important during execution of tasks. A description of this can be found in Daniel S. Weld, Recent Advances in AI Planning, AI Magazine, 20(2):93-123, Summer 1999, which is incorporated by reference herein in its entirety.

[0010] One conventional approach has utilized expert systems to encode the steps for accomplishing a task algorithmically. A key component was the capture of human expert knowledge using a laborious manual process. A description of this can be found in D. A. Waterman, A Guide to Expert Systems, Addison Weseley, 1986, which is hereby incorporated by reference in its entirety. A disadvantage of this approach is that not everything that humans learn is taught by experts. Most day-to-day activities, e.g., tying shoe laces, are learned by observations of and interaction with non-experts.

[0011] According to Rasmussen et ah, human activity in such routine tasks is goal- oriented and controlled by a set of proven rules. The sequence of task steps is typically derived empirically, communicated from another's knowhow or a "cookbook" sequence. A description of this can be found in Jens Rasmussen, Skills, Rules and Knowledge: Signals, Signs, and Symbols, and Other Distinctions in Human Performance Models, IEEE Transactions on Systems, Man and Cybernetics, SMC-13(3):257-266, May/June 1983. [0012] One source of common sense knowledge is the Worldwide Web ("web"). For instance, websites such as eHow.com list the steps to perform activities. Intel Corporation developed a system called Probabilistic Activity Toolkit (PROACT) to build activity models. They automatically identified activities by observing the objects involved in the activity. They also found the relevance of various terms to a given activity from the web. For instance, the word "cup" is highly related to the activity "making tea" because "cup" occurs frequently on web pages about making tea. A description of this can be found in Matthai Philipose, Kenneth P. Fishkin, Mike Perkowitz, Donald Patterson, and Dirk Haehnel, The Probabilistic Activity Toolkit: Towards Enabling Activity-Aware Computer Interfaces, Technical Report IRS-TR-03-013, Intel Research Laboratories, November 2003, and in Mike Perkowitz, Matthai Philipose, Kenneth Fishkin, and Donald J. Patterson, Mining Models of Human Activities from the Web, Proceedings of the 13th Conference on World Wide Web, pages 573-582, ACM Press, 2004, both of which are incorporated by reference herein in their entirety.

2006/023822

[0013] Using the web as an open information source for building plans for tasks is very attractive. However, the extracted knowledge exhibits high variance and "noise," e.g., extraneous or erroneous information, and documents may be prohibitively large. An alternative is a distributed information source such as the Open Mind Indoor Common Sense (OMICS) database. In compiling this database, volunteers are prompted with household tasks and asked to provide steps to accomplish them. A description of this can be found in R. Gupta and M. Kochenderfer, Common Sense Data Acquisition for Indoor Mobile Robots, Nineteenth National Conference on Artificial Intelligence (AAAI-04), July 25-29 2004. However, even with this approach, semantic information must be extracted from the steps provided. [0014] From the above, there is a need for a practical system and method for building plans from distributed knowledge for enabling autonomous machines such as humanoid robots to perform tasks in constrained environments such as indoor environments.

SUMMARY OF THE INVENTION

[0015] The present invention meets these needs with a method and apparatus for providing autonomous machines such as humanoid robots with plans for performing tasks in constrained environments, such as indoor environments. According to one aspect of the invention, plans for performing such tasks are extracted or generated automatically from a knowledge database. The knowledge in the database is contributed by human subjects using distributed knowledge capture techniques. The Worldwide Web ("Web") is optionally used as a distributed knowledge capture and transfer medium, and as such, collection from a great number of people may be practical. As a benefit, the knowledge base embodies "common sense," that is, the consensus of the contributing subjects.

[0016] The knowledge database comprises one or more tasks, each task comprising one or more plans, and each plan comprising a sequence of instructions. According to one embodiment, the present invention chooses the "best" plan from the plans in the knowledge database as the plan that represents the majority consensus. In this embodiment, hierarchical agglomerative clustering may be used to group similar plans and then merge similar groups into larger groups. This is referred to as the discriminative approach.

[0017] According to another embodiment, the present invention generates, i.e., derives, a new plan from the plans in the knowledge database. This is referred to as the generative approach. Each plan may be modeled as a first-order Markov chain, wherein each step depends on the previous one, with no hidden states. According to one embodiment of the

present invention, the generated plan takes into consideration environmental constraints, such as whether a particular appliance, e.g., a coffeemaker, is present within the indoor environment. This enables a humanoid robot to generate optimum task plans in real time, based on the robot's perception or information of such constraints. [0018] Pairs of actions and objects are extracted from the knowledge database to formulate task steps. For example, for the task "wash clothes," the action-object pair corresponding to one task step might be (collect, clothes). A subsequent step might be (move to, washing machine). Furthermore, relationships between actions and objects for a given tasks are derived from the knowledge database. For example, the conditional probability of collecting clothes given that clothes are present is derived. A particular task may comprise multiple, potentially hundreds, of plans, each consisting of a sequence of steps. A plan may comprise any number of steps, preferably five to seven.

[0019] According to one embodiment, the possible steps of the possible plans corresponding to a particular task are captured in a graph. The graph comprises nodes and links. The nodes represent action-object pairs, i.e., steps, and the links represent transitions between steps. A given node / step may be common to multiple plans. The graph begins with a common "start" node, from which links connect to all possible first steps of the task. Associated with each link is a relative likelihood, or probability, of the corresponding transition, as determined from the knowledge database. This topology repeats for the successive nodes and links.

[0020] This graphical methodology advantageously provides an efficient basis for deriving or generating all possible plans corresponding to a task. The graphic model can be conveniently modified when plan updates become available. Plans may be determined from the graph based on a variety of strategies, such as local or global optimality. According to one embodiment, in the event of a tie between two candidate plans, the plan having the shortest sequence is chosen. According to another embodiment, environmental constraints are taken into account. For example, if it is known that the actual environment has an available washing machine, plans that utilize washing machines will be favored, as opposed to, say, plans that involve dry cleaning.

[0021] The features and advantages described in the specification are not all inclusive and, in particular, many additional features and advantages will be apparent to one of ordinary skill in the art in view of the following drawings, specification, and claims. Moreover, it should be noted that the language used in the specification has been principally selected for

readability and instructional purposes, and may not have been selected to delineate or circumscribe the inventive subject matter.

BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The invention has other advantages and features which will be more readily apparent from the following detailed description of the invention and the appended claims, when taken in conjunction with the accompanying drawings, in which:

[0023] Figure 1 illustrates a system according to one embodiment of the present invention.

[0024] Figure 2 illustrates a method of creating a task knowledge database, according to one embodiment of the present invention.

[0025] Figure 3 illustrates a technique for building a plan for executing a task according to one embodiment of the present invention.

[0026] Figure 4 illustrates a portion of a task model according to one embodiment of the present invention.

[0027] Figure 5 illustrates results of experimental evaluation of the present invention.

[0028] Figure 6 illustrates p-values corresponding to experimental evaluation of the present invention.

DETAILED DESCRIPTION

[0029] Reference will now be made in detail to several embodiments of the present invention, examples of which are illustrated in the accompanying figures. It is noted that wherever practicable similar or like reference numbers may be used in the figures and may indicate similar or like functionality. The figures depict embodiments of the present invention for purposes of illustration only. One skilled in the art will readily recognize from the following description that alternative embodiments of the structures and methods illustrated herein may be employed without departing from the principles of the invention described herein.

[0030] Now referring to Figure 1, a system according to one embodiment of the present invention is shown. Computer system 110 comprises an input module 112, a memory device 114, a storage device 118, a processor 122, and an output module 124. In an alternative embodiment, an image processor 120 can be part of the main processor 122 or a dedicated device to perceive situations as digital images captured in a preferred image format.

Similarly, memory device 114 may be a standalone memory device, (e.g., a random access memory (RAM) chip, flash memory, or the like), or an on-chip memory with the processor 122 (e.g., cache memory). Storage device 118 may be any bulk storage device such as a hard disk, DVD-R/RW, CD-R/RW or RAM. Likewise, computer system 110 can be a stand-alone system, such as, a server, a personal computer, or the like. Alternatively, computer system 110 can be part of a larger system such as, for example, a robot having a vision system. [0031] According to this embodiment, computer system 110 comprises an input module

112 to receive requisite information such as commonsense data from a database 140. Input module 112 may also receive digital images directly from an imaging device 130, for example, a digital camera 130a (e.g., robotic eyes), a video system 130b (e.g., closed circuit television), an image scanner, or the like. Alternatively, the input module 112 may be an interface to receive information from a network system, for example, a database, another vision system, Internet servers, or the like. The network interface may be a wired interface, such as, a USB, RS-232 serial port, Ethernet card, or the like, or may be a wireless interface module, such as, a wireless device configured to communicate using a wireless protocol, e.g., Bluetooth, WiFi, IEEE 802.11, or the like. An optional image processor 120 may be part of the processor 122 or a dedicated component of the system 110. The image processor 120 could be used to pre-process the digital images received through the input module 112 to convert the digital images to the preferred format on which the processor 122 operates. [0032] Requisite information is stored in the memory device 114 to be processed by processor 122. Processor 122 applies a set of instructions that when executed perform one or more of the methods according to the present invention, e.g., generating a plan for executing a task as described herein. Memory device 114 may, e.g., include a module of instructions 116 for generating a plan for executing a task.

[0033] Processor 122 may output information through the output module 124 to an external device 150, e.g., a network element or server 150a, a display device 150b, a database 150c or the like. As with input module 112, output module 124 can be wired or wireless. Output module 124 may be a storage drive interface, (e.g., hard-drive or optical drive driver), a network interface device (e.g., an Ethernet interface card, wireless network card, or the like), or a display driver (e.g., a graphics card, or the like), or any other such device for outputting the information or responses determined.

[0034] Referring now to Fig. 2, according to one embodiment of the present invention, a method 200 for creating a plan database is shown. Method 200 may be implemented, for

example, by processor 122 executing an instruction module 116 as described above. Knowledge database 206 is first assembled offline, and may correspond, for example, to database 140 discussed above. For each defined task, humans contribute commonsense plans to database 206, each comprising a sequence of steps. Knowledge database 206 may include plans for hundreds of tasks such as making coffee, cleaning the floor or washing clothes. A sample plan from knowledge database 206 is shown below: Task : wash clothes

Steps : collect clothes move to washing machine place clothes in washing machine add detergent to clothes close lid of washing machine start washing machine

[0035] One embodiment of the present invention uses the set of plans for a given task within knowledge database 206 to build a plan database 226. Plan database 226 may reside, for example, in storage device 118 or database 150c as discussed above. A particular plan is later extracted or generated from plan database 226, as will be described. A task of interest is chosen 212. The information in Knowledge database 206 is first pre-processed by extracting 216 a set of action-object pairs corresponding to the task. According to one embodiment, task steps are first parsed with Brill's part-of-speech (POS) tagger. A description of this can be found in Eric Brill, A Simple Rule-based Part-of-speech Tagger, Proceedings of ANLP-92, 3rd Conference on Applied Natural Language Processing, pages 152-155, Trento, IT, 1992. This method identifies the first verb as the action. If the verb is followed by a proposition, the preposition is combined with the action verb. Finally, the first noun phrase is identified as the object of the action. The result of parsing the above sample plan is shown below:

Task: wash clothes Action object pairs for steps: collect, clothes move to, washing machine place, clothes add, detergent close, lid start, washing machine

[0036] According to one embodiment, a conditional probability distribution is determined 220 for every object and corresponding action-object pair for a given task according to:

P _n ( _/ act j.ι-on I ) ob jj ■ect ,) _\ = — f(action —,ob -j —ect) - (1)

/(object) where/ (action, object) is the number of times action occurs with object and/ (object) is the number of times object occurs within all of the plans corresponding to the given task. Once determined, the task definitions and corresponding action-object pairs and conditional probabilities are stored 224 in plan database 226.

[0037] A simplistic method of determining a plan makes a random selection from the set of plans in the plan database 226. This approach serves as a basis for comparison with more sophisticated approaches discussed below, and shall be referred to hereinafter as Technique 1.

DISCRIMINATIVE APPROACH

[0038] According to one embodiment of the present invention, the plan that represents the consensus of the human contributors is selected from plan database 226. This is referred to herein as the discriminative approach, as well as Technique 2. This technique is based on hierarchical agglomerative clustering method 300, as illustrated in Fig. 3. This comprises grouping similar plans present in plan database 226, and then clustering similar groups into larger clusters. A description of this can be found in Gelad Salton, ed., Automatic Text Processing: The Transformation, Analysis, and Retrieval of Information by Computer, Addison Wesley, 1989, which is incorporated by reference herein in its entirety. According to this embodiment, plan database 226 may reside in storage device 118 or may correspond to database 140, as discussed above. Method 300 may be implemented, for example, by processor 122 executing an instruction module 116 as described above Also, the results of clustering method 300 may be passed by output device 124 to network 150, and stored or forwarded as appropriate.

[0039] For each task in plan database 226, the similarity metric between pairs of plans

Pi and p _j in plan database 226 is determined 308 as

Sm(P ₁ ,p _j ) = \ / (2)

Un(P ₁ )Un(P _j ) where UnφJ is the length of, e.g., the number of steps in, plan/? _/ and Un(LMS) is the length

of the largest matching sequence, i.e., the largest number of common sequential steps in plans p, and pj. Pairs of plans are grouped 316 by specifying 312 a minimum similarity according to equation (2). Such groups are in turn hierarchically clustered 320 based on the average group similarity used to obtain a specified number of clusters. In other words, two levels of clustering are performed to obtain a desired number of hierarchical clusters. [0040] It has been empirically found that such clusters correspond to distinct techniques or categories for accomplishing a task. For example, for the task of making coffee, the various clusters correspond to using a coffee maker, instant coffee, and an espresso machine. Since a majority of volunteers contributed plans for making coffee using a coffeemaker, that category was the largest cluster. Plans for indoor tasks commonly fall into five or fewer categories. Therefore, the average group similarity is preferably set to obtain five or fewer hierarchical clusters, i.e., categories of plans, for each task. After hierarchical cluster formation, a particular plan is randomly selected 324 from the hierarchical cluster that corresponds to the largest number of contributed plans. GENERATIVE APPROACH

[0041] According to another embodiment of the present invention, a new plan may be generated from plan database 226. First, a task model is constructed for each task using first- order Markov chains. A Markov chain is a sequence of random variables. Application of Markov chains is useful due to the inherently sequential nature of the steps comprising a plan. According to one embodiment, each plan is modeled as a first-order Markov chain wherein each step depends on the previous step, with no hidden states. An exemplary task model 400 for the task "wash clothes" is illustrated in Fig. 4. Task model 400 comprises a graph wherein the nodes, also referred to as states or steps, represent task steps. The links of the graph represent transitions between successive steps. AU plans for the task begin and end with a common start step 410 and a common end step 490. A plan is represented by a particular sequence of alternating steps and links. An exemplary plan for washing clothes is partially represented within the dashed boundary 412. Method 400 may be implemented, for example, by processor 122 executing an instruction module 116 as described above. The generated plan may be passed by output device 124 to network 150, and stored or forwarded as appropriate. [0042] Steps 420 succeed start step 410 and comprise all possible first steps for all plans that can be derived from graph 400. For example, step 420b is the first step (after the start step) of the exemplary plan. Associated with each link is a transition probability, that is, a relative probability of the succeeding step given an occurrence of the preceding step. For

example, if 100 plans are represented in task model 400, and "collect clothes" is the first step of 30 of the plans, the probability associated with step 420b is 0.3. The probabilities for the other links in task model 400 are similarly determined.

[0043] The complete task model thus compactly represents the possible steps for all plans for a given task. Furthermore, the task model serves as a basis for generating new plans, as will be described below. Advantageously, the task model can be easily expanded or otherwise updated by adding or changing steps and links and adjusting the probabilities as new plans for the task become available.

GENERATIVE PLANS BASED ON LOCAL OPTIMALITY

[0044] According to one embodiment of the present invention, a new plan for a task may be generated from the corresponding task model by choosing successive steps according to highest probability. For example, in task model 400, state 420b would be chosen among states 420. Thus, the state at time t is determined by

NextStateif) = arg max p(s, | S _j ) (3)

where t is the time of the current step, S _j is the state at time t-1, and s, are the possible successor states of S _j , that is, the states linked to sj. Such a generative plan is said to have local optimality. This approach shall be referred to hereinafter as Technique 3. [0045] Once a candidate step has been selected, incoming links to that step within the task model other than the link from the preceding state are removed. This is done to avoid forming undesired cycles, that is, loops through the graph. It should be noted that in general it is possible for the same step to occur more than once in a plan. For example, for the task "wash clothes," the step "open lid" can occur before putting the clothes into the washing machine and after washing is done. A sample generated plan for the task "wash clothes" according to this embodiment is given below, with each step described by the format step -> transitionprobability -^step: start -> 0.36 -> get, clothes get, clothes -> 0.3333 -> locate, machine locate, machine -> 0.4285 -> move to, machine move to, machine -> 0.2727 -> fetch, clothes fetch, clothes -> 0.5 -> open, machine open, machine -> 0.8 -> put, clothes put, clothes -> 0.5294 -> add, detergent add, detergent -> 0.5 -> close, machine

close , machine - > 0 . 2857 - > start , machine start , machine -> 0 . 6666 - > end

GENERATIVE PLANS BASED ON GLOBAL OPTIMALITY

[0046] According to another embodiment of the present invention, a plan may be generated from a task model by choosing each step in a. manner that takes into consideration all preceding steps, not just the immediately preceding step. In this embodiment, the step at time t is determined using the equation

NextState(t) = arg max />(_?,• | S ₁ , S ₂ ,...,S^ ₁ ) (4) i where t is the time of the current step, S ₁ are all the possible successor steps of _? _/ _; _, and sj, s 2, ..., Si.] are the steps that occurred through time t-1 and are linked in the task model. If multiple candidate plans give the largest probability, that is, if there is a tie, a selection may optionally be made randomly. This approach is referred to as Technique 4.

[0047] According to another embodiment of the present invention, Technique 4 is followed, except that in case of a tie, the plan having the fewest steps is chosen. This approach is referred to as Technique 5.

GENERATIVE PLANS WITH ENVIRONMENTAL CONSTRAINTS

[0048] Techniques 1-5 may be implemented a priori, that is, without regard to the actual indoor environment in which the autonomous machine will operate. However, known aspects of the actual indoor environment may be taken into account in generating a plan. This promises a practicable plan. Such environmental information may, for example, be provided by the user of a robot. Alternately, environmental information may be determined by the robot via digital camera 130a, video system 130b, or other means. It may be advantageous to restrict or bias the generated plan to recommend the use of known available objects. This approach is referred to as Technique 6.

[0049] According to one embodiment of the invention, since task models are comprised of action-object pairs, the probabilities determined in Equation (1) may be utilized by assuming that the most likely action is the one that occurs most frequently with the object known to be present. More specifically, the most probable action to be associated with such an object is found according to

Most probable action = arg max P (action | object) (5) action

Even though observed objects are associated with their most likely actions, such actions are

not mandated within the plan. For example, the plan generation process may neglect such steps for consistency with other chosen steps.

[0050] A sample generative plan for the task "washing clothes" with the constraint that water, clothes and washing machine be used is given below:

Found restriction: "feed, water" with probability 0.2

Found restriction: "put, clothes" with probability 0.32

Found restriction: "start, washing machine" with probability 0.15

The plan: start -> 0.36 -> get, clothes get, clothes -> 1 -> put, clothes put, clothes -> 1 -> feed, water feed, water -> 1 -> feed, detergent feed, detergent -> 1 -> set, timings set, timings -> 1 -> start, washing machine start, washing machine -> 0.6666 -> end

To bias the method toward choosing steps that fulfill the restrictions, the corresponding transition probabilities are set to 1.0. As a result, this plan differs from the one derived above without constraints.

[0051] In experiments, task plans for a set of 105 household tasks were determined and evaluated according to techniques 1-5 as described above. Technique 6 was not evaluated for reasons discussed below. At least 25 plans were present in the knowledge database for each task. All determined tasks used the same knowledge and preprocessing procedure. The five techniques evaluated are summarized below:

Technique 1 (Random approach): a plan was selected randomly from the plan database.

Technique 2 (Discriminative approach): a plan was selected from the largest cluster.

Technique 3 (Generative approach, local optimality): a plan was generated from the corresponding Task Model, with each step considered independently.

Technique 4 (Generative approach, global optimality, random tie resolution): a plan was generated from the Task Model by evaluating the probability of the sequence of steps as a whole. In case of a tie, a random selection was made from the best candidates.

Technique 5 (Generative approach, global optimality, deterministic tie resolution): a plan was generated from the Task Model by evaluating the probability of the whole sequence of steps and by choosing the plan having minimum length in case of a tie. [0052] The following criteria were used to evaluate and compare the results:

1. Completeness: a plan for a task should include a complete set of steps. For example, a plan for cleaning the floor that applied water and soap but did not mop the floor afterward would be considered incomplete.

2. Correct sequence: the sequence of steps should be consistent with the steps themselves. For example, a plan that poured coffee from a carafe into a mug before adding water to the coffee-maker would merit a low rating.

3. Sensibility: the overall plan should make sense. For example, if a coffee-making plan used both a coffee-maker and instant coffee, it would merit a low rating.

4. Concise description: for a given number of objects, a plan having a more concise description is preferred. For example, for the task "making coffee," the step "add filter" is preferable to the sequence "find filters," "take one filter," "add filter."

[0053] Since the "goodness" of a plan cannot be evaluated objectively, human subjects evaluated the results. Ten subjects ranked each determined plan for each of the 105 tasks. Technique 6 was not evaluated in this manner, since the use of environmental constraints was considered difficult to evaluate. However, technique 6 is a special case of technique 3 and thus "inherits" the advantages of that technique. The worst possible score for a plan was 5 and the best possible score was 1. The ten scores each plan received were averaged. The 105 average scores for each technique were then added. The results are summarized in Fig. 5, from which the following observations may be made:

• Technique 1 gives the worst performance; thus, reliance on the knowledge of just one person or source likely results in a poor plan.

• Technique 2 is superior to Technique 1. Thus, a plan chosen randomly from the consensus cluster is better than a plan chosen completely at random.

• Technique 3 offers the best performance. This technique is also attractive for other reasons. First, it considers steps as separate entities, rather than aspects of pre-existing plans (as with Techniques 1 and 2). Thus, the resultant plan is not limited to a pre-existing plan (as with Techniques 1 and 2). Second, this technique is able to remove some "noise," e.g., misspellings, mistakes, etc., and spurious data through learning. Third, it captures the consensus at the level of task steps and their sequencing.

• Techniques 4 and 5 do not perform as well as technique 3. This may be due to the lack of the number of plans used to effectively perform inferencing on long sequences of steps. [0054] To evaluate the confidence associated with the ranking in Table 1, a paired two- tailed t-test was performed. This test evaluates the level of confidence in the relative results of a pair of tests by determining whether the outcomes of the two tests result from the same probability distribution. If such is the case, then the results will be statistically meaningful, i.e., there will be high confidence in the evaluated relative performance of the two techniques. The p-values for each of Techniques 2-5 relative to Technique 1 are given in Fig. 6. For p- values less than 0.05, the confidence in the relative performance will be at least 95%. As shown in Fig. 6, the p-values for techniques 3 and 4 are less than 0.05. Thus, there is a confidence of at least 95% that techniques 3 and 4 perform better than technique 1. As described above, these tests ranked best and second best, respectively. [0055] Advantages of the present invention include a system and method for providing plans for autonomous machines such as humanoid robots to perform indoor tasks. The plans are based on commonsense, that is, on the consensus of non-expert human volunteers. This is an appropriate basis for routine household and other indoor tasks. The plans may be determined a priori, that is, without knowledge of the actual operational environment. Alternately, plans may be optimized to take into account the availability of objects in the actual operational environment. The plans are easily modified when new variations become available.

[0056] Those of skill in the art will appreciate additional alternative designs for a system and method for providing plans for autonomous machines to perform indoor tasks. Thus, it is to be understood that the invention is not limited to the precise construction and components disclosed herein and that various modifications, changes and variations which will be apparent to those skilled in the art may be made in the arrangement, operation and details of the method and apparatus of the present invention disclosed herein without departing from the spirit and scope of the invention as defined in the appended claims.