Information processing device, information processing method, and information processing program

Abstract

An information processing device according to one aspect of the present invention comprises a task latent expression extractor, an operation generator, an intermediate body expression converter, and an operation discriminator. The task latent expression extractor extracts a latent expression vector from a language instruction given to a robot to be trained. The operation generator generates operation data representing an operation of the robot on the basis of the latent expression vector and information about the robot. The intermediate body expression converter acquires correct answer operation data of when a teacher operation subject having a body expression different from that of the robot has actually executed the language instruction, generates an intermediate body expression, which is an intermediate expression between the body expression of the teacher operation subject and the body expression of the robot, and converts each of the generated operation data and the correct answer operation data into an operation that uses the intermediate body expression. The operation discriminator discriminates the difference between the correct answer operation data after the conversion and the generated operation data after the conversion.

Description

Information processing device, information processing method, and information processing program

[0001] The present invention relates to an information processing device, an information processing method, and an information processing program.

[0002] One method for training a robot to perform a specific task involves collecting actual motion data using an actual robot. However, this method incurs high costs for actually operating the actual robot and also poses the risk of the actual robot breaking down. Another method involves preparing robot motion data using a simulator, but this only allows for limited tasks to be simulated within the simulator, making it difficult to prepare the wide variety of motion data required in reality. For this reason, a task capability transfer technology has been devised that transfers human task capabilities to a robot by having the robot learn the movements of a model human (expert). This technology uses, for example, Generative Adversarial Imitation Learning (GAIL) to train a humanoid robot to learn motions that mimic the movements of a model human.

[0003] In the above-described work capability transmission technology, a large amount of training data is required for a robot to autonomously perform a task, and this training data needs to be motion data of an agent with a similar shape to the robot performing the task. Furthermore, it is difficult to adapt training data prepared for training a humanoid robot to training a robot with a non-human shape. For example, it is difficult to adapt training data prepared for training a humanoid robot to training a quadrupedal robot. This makes it difficult to prepare training data for robots with a non-human shape.

[0004] Jingkai Sun et al., “Prompt, Plan, Perform: LLM-based Humanoid Control via Quantized Imitation Learning”, arXiv:2309.11359, 2023Jonathan Ho et al., “Generative Adversarial Imitation Learning”, arXiv:1606.03476, 2023

[0005] The present invention has been made in light of the above circumstances, and its purpose is to use movement data of bodily expressions that differ from those of a robot as learning data in a technology for transferring working ability from a human to a robot.

[0006] To achieve the above object, one aspect of an information processing device according to the present invention includes a task latent expression extractor, a motion generator, an intermediate physical expression converter, and a motion discriminator. The task latent expression extractor extracts a latent expression vector from linguistic instructions for a robot to be learned. The motion generator generates generated motion data representing the motion of the robot based on the latent expression vector and information about the robot. The intermediate physical expression converter acquires correct motion data obtained when a teacher subject, who has a physical expression different from that of the robot, actually executes the linguistic instructions, generates intermediate physical expressions that are intermediate expressions between the physical expression of the teacher subject and the physical expression of the robot, and converts each of the generated motion data and the correct motion data into motions using the intermediate physical expressions. The motion discriminator discriminates between the converted correct motion data and the converted generated motion data.

[0007] According to one aspect of the present invention, even if the physical expression of the subject of the action used as training data differs from that of the robot, by using an intermediate physical expression, it becomes possible to transfer abilities to a robot with a shape different from that of the subject of the training data, making it possible to transfer work abilities from a person to any robot.

[0008] That is, according to one aspect of the present invention, in a technique for transmitting working ability from a human to a robot, movement data of bodily expressions different from those of a robot can be used as learning data.

[0009] FIG. 1 is a diagram illustrating an example of the functional configuration of an information processing device according to a first embodiment. FIG. 2 is a flowchart illustrating an example of the processing procedure of a learning process executed by the information processing device according to the first embodiment. FIG. 3 is a diagram illustrating an example of how an intermediate physical representation is generated using the information processing device according to the first embodiment. FIG. 4 is a diagram illustrating an example of how an intermediate physical representation is generated using an information processing device according to a first modified example of the first embodiment. FIG. 5 is a flowchart illustrating an example of the processing procedure of an intermediate physical representation generation process executed by an information processing device according to a second modified example of the first embodiment. FIG. 6 is a diagram illustrating an example of the functional configuration of an information processing device according to a second embodiment. FIG. 7 is a flowchart illustrating an example of the processing procedure of a movement data generation process executed by the information processing device according to the second embodiment. FIG. 8 is a diagram illustrating an example of the hardware configuration of an information processing device according to an application example.

[0010] An information processing device, an information processing method, and an information processing program according to an embodiment of the present disclosure will be described in detail below with reference to the drawings. In the following description, components having substantially the same functions and configurations are denoted by the same reference numerals, and redundant description will be provided only when necessary.

[0011] First Embodiment (Configuration Example) FIG. 1 is a diagram illustrating an example of the functional configuration of an information processing device 10 according to a first embodiment. The information processing device 10 is a task capability transmission device that causes a robot to learn the movements of an expert as a model to perform a specific task. The information processing device 10 may also be referred to as an imitation learning device that generates robot movements by imitating human movements. The information processing device 10 of this embodiment is a learning device that includes a generator and a discriminator and trains the generator and discriminator using generative adversarial imitation learning (GAIL). The information processing device 10 acquires the movements of an expert as a model as training data, generates robot movements based on linguistic instructions, and compares the generated robot movements with the training data to perform learning using adversarial imitation learning.

[0012] 1, an information processing device 10 includes a task latent expression extractor 11, an action generator 12, an intermediate body expression converter 13, an action classifier 14, and an update unit 15. The action generator 12 functions as a generator in adversarial imitation learning. The action classifier 14 functions as a classifier in adversarial imitation learning.

[0013] The task latent expression extractor 11 converts linguistic instructions for the robot to be trained into latent expression vectors, thereby extracting latent expression vectors from the linguistic instructions.

[0014] The robot to be trained is a robot with physical expressions that can perform a specific task. Hereinafter, the robot to be trained will be simply referred to as a robot. The robot may be, for example, a humanoid robot or a robot with a non-humanoid shape. The robot may be, for example, a guide robot that moves around to guide a person to a destination, or an arm robot that performs tasks using an arm. For example, Ugo (registered trademark), TEMI (registered trademark), or Unitree (registered trademark) can be used as the guide robot. For example, Kinova or Universal Robot can be used as the arm robot. Note that the robot is not limited to the above examples, and any robot having physical functions that can perform the target task may be used.

[0015] The linguistic instruction is data in a language format that expresses the content of the task to be performed by the robot. The linguistic instruction is, for example, text data in a natural language format. The linguistic instruction is, for example, an instruction to a walking robot to specify a walking method, or an instruction to a guide robot to specify a guidance start position and a destination position. The linguistic instruction is a task to be performed by the robot as long as it is a task that the robot to be learned can perform.

[0016] The latent expression vector is a feature vector that represents the features of a task included in a linguistic instruction. The latent expression vector may also be called a skill. A known method can be used to convert a linguistic instruction into a latent expression vector.

[0017] The motion generator 12 generates generated motion data representing the motion of the robot based on the latent representation vector and information about the robot to be trained (hereinafter referred to as robot information). The robot information includes, for example, the size, shape, movable parts, skeleton, current posture, movement speed, and range of motion. For example, the motion generator 12 is a machine learning model trained to accept input of the latent representation vector converted from a linguistic instruction and robot information about the robot to be trained, and to output motion data for the robot to be trained to execute the task included in the linguistic instruction. For example, a known machine learning model used as a generator in an adversarial imitation learning method can be used as the motion generator 12.

[0018] The generated motion data is information about the motion of the robot when performing a task, including, for example, the movement trajectory of the robot's center position when performing a task, the positional relationship between the robot and other objects, the timing and speed of movement of movable parts, etc.

[0019] The intermediate physical expression converter 13 acquires correct action data. The correct action data is information about the action of the teacher action subject when executing a linguistic instruction. The teacher action subject is a person (expert) who has the work ability to be transferred to the robot for a specific task, and is a task performer who can execute the task with high accuracy. The teacher action subject is, for example, an expert or engineer who can execute the task of the linguistic instruction. The teacher action subject may have a physical expression different from that of the robot to which the work ability is transferred. The correct action data includes, for example, video recording the actions of the teacher action subject when executing the task, skeletal information of the action performing body, the movement trajectory of the center position of the action performing body, the positional relationship between the teacher action subject and other objects, the timing and movement speed of the movement of the teacher action subject's movable parts, etc.

[0020] Furthermore, the intermediate physical representation converter 13 generates an intermediate physical representation that is an intermediate representation between the physical representation of the teacher's subject and the physical representation of the robot. The intermediate physical representation is a physical representation that is composed of common elements between the physical representation of the robot and the physical representation of the teacher's subject. In this way, the intermediate physical representation converter 13 generates an intermediate physical representation that makes it possible to compare generated movement data having different physical representations with correct movement data.

[0021] The intermediate body representation is composed of, for example, one or more objects that approximate the body representation of the robot and the body representation of the teacher action subject. The object shape may be, for example, oval, rectangular, or circular, but may also be other shapes. The intermediate body representation may be composed of one object or multiple objects. The intermediate body representation may be generated using a rule-based method or a machine learning model (hereinafter referred to as a generative model) trained to output information for generating the intermediate body representation. Examples of generative models include generative AI technologies that generate images from text, such as diffusion models, and foundational models, such as large language models (LLMs). For example, when generating an intermediate body representation using a large language model, prompts such as "Please select the modal information required to perform task A" or "Please extract the information required to perform task A from the input data" are input to the large language model, and the intermediate body representation is generated using the response sentences from the large language model.

[0022] Furthermore, the intermediate body representation converter 13 converts the generated movement data and the correct movement data into the same intermediate body representation. At this time, the intermediate body representation converter 13 generates converted generated data by converting the generated movement data into a movement using the intermediate body representation, and converted correct data by converting the correct movement data into a movement using the intermediate body representation. As a method for converting into an intermediate body representation, for example, existing motion retargeting technology can be used. Alternatively, as a method for converting into an intermediate body representation, online motion retargeting using inverse kinematics or a conversion method based on deep learning can also be used.

[0023] The action classifier 14 determines the difference between the generated action data generated by the action generator 12 and the correct action data. In this case, the action classifier 14 determines the difference between the converted generated data obtained by converting the generated action data into an intermediate body expression and the converted correct data obtained by converting the correct action data into the intermediate body expression. The action classifier 14 is a machine learning model trained to accept input of the converted generated data or the converted correct data, determine whether the input data is the converted generated data or the converted correct data, and output the determination result. As the action classifier 14, for example, a known machine learning model used as a classifier in an adversarial imitation learning method can be used.

[0024] The update unit 15 updates the action generator 12 and the action discriminator 14 by an adversarial imitation learning method. At this time, the update unit 15 updates the action generator 12 so as to generate generated action data that cannot be detected by the action discriminator 14, and updates the action discriminator 14 so as to be able to accurately distinguish between the generated action data of the action generator 12 and correct action data. The update unit 15 alternately and repeatedly updates the action generator 12 and the action discriminator 14 until a specific number of times is reached or until a preset evaluation value threshold is reached.

[0025] (Operation Example) Next, an operation example of the information processing device 10 configured as above will be described. Fig. 2 is a flowchart showing an example of the procedure of the learning process by the information processing device 10. The learning process is a series of processes for updating the action generator 12 and the action discriminator 14 by an adversarial imitation learning method. Here, as an example, a case will be described in which a guide robot uses the action generator 12 that generates action data when performing the task of "guiding a person to be guided from point A to point B."

[0026] The processing procedure for the reliability determination process described below is merely an example, and each process can be modified as appropriate as possible. Furthermore, steps in the processing procedure described below can be omitted, replaced, or added as appropriate depending on the embodiment.

[0027] In the learning process, first, the task latent expression extractor 11 extracts a latent expression vector from a task instruction (step S101). At this time, the task latent expression extractor 11 extracts a latent expression vector from a task instruction including the task of "guiding a guide target person from point A to point B." The task latent expression extractor 11 transmits the extracted latent expression vector to the action generator 12 and the action discriminator 14.

[0028] Next, the motion generator 12 generates motion data for the guide robot based on the latent expression vector and the robot information of the guide robot (step S102). The robot information includes the size, center position, skeleton, etc. of the guide robot. The motion generator 12 generates generated motion data as motion data for the guide robot when it executes the task of "guiding a person to be guided from point A to point B." The generated motion data includes motion information of the body expressions possessed by the guide robot. The motion generator 12 transmits the generated motion data to the intermediate body expression converter 13.

[0029] Next, the intermediate physical expression converter 13 acquires correct movement data (step S103). At this time, the intermediate physical expression converter 13 acquires, as the correct movement data, information about the movements of a guidance expert (hereinafter referred to as a guide) when performing the task of "guiding a person to be guided from point A to point B." The guide is an example of a teaching action subject. The correct movement data includes the guide's skeleton, a video captured from behind the guide, the movement trajectory of the guide's skeleton, and the movement trajectory of the person to be guided.

[0030] Next, the intermediate body representation converter 13 generates intermediate body representations of the guide robot and the guide (step S104). Here, as shown in Fig. 3, an intermediate body representation C is generated that approximates the body representation A of the guide and the body representation B of the guide robot. The intermediate body representation C is composed of a combination of elliptical shapes that approximate the skeletal structures and center positions of the guide and the guide robot. In the example of Fig. 3, the intermediate body representation C is composed of two objects, an elliptical object C1 and an elliptical object C2.

[0031] Next, the intermediate physical representation converter 13 converts each of the correct movement data and the generated movement data into intermediate physical representations (step S105). At this time, the intermediate physical representation converter 13 converts the movement data defining the movement by the guide's movement representation A into movement data defining the movements of the ellipse objects C1 and C2 of the intermediate physical representation C. Similarly, the intermediate physical representation converter 13 converts the movement data defining the movement by the robot's movement representation B into movement data defining the movements by the ellipse objects C1 and C2 of the intermediate physical representation C. The intermediate physical representation converter 13 transmits the converted correct movement data and generated movement data to the movement classifier 14.

[0032] Next, the action classifier 14 acquires either the correct action data converted into an intermediate body expression or the generated action data converted into an intermediate body expression, and determines whether the acquired action data is the correct action data or the generated action data (step S106). The action classifier 14 transmits the determination result to the update unit 15.

[0033] Next, the update unit 15 updates the action generator 12 and the action classifier 14 by the adversarial imitation learning method using the discrimination result of the action classifier 14 (step S107). The update unit 15 repeatedly executes the processes from step S101 to step S107 until a learning termination condition is met (step S108). The learning termination condition is set using, for example, the number of repetitions or an evaluation value for the action generator 12. As a result, the update of the action generator 12 and the update of the action classifier 14 by adversarial imitation learning are executed alternately and repeatedly.

[0034] (Operations and Effects) The information processing device 10 of this embodiment includes a task latent expression extractor 11, a motion generator 12, an intermediate body expression converter 13, and a motion discriminator 14. The task latent expression extractor 11 extracts latent expression vectors from linguistic instructions for a robot to be trained. The motion generator 12 generates generated motion data based on the latent expression vectors and robot information. The intermediate body expression converter 13 acquires correct motion data. The correct motion data is motion data obtained when a teacher action subject, who has a different physical expression from the robot, executes the linguistic instructions. The intermediate body expression converter 13 generates intermediate physical expressions that are intermediate expressions between the physical expressions of the teacher action subject and the physical expressions of the robot. The intermediate body expression converter 13 converts each of the generated motion data and the correct motion data into motions using the intermediate physical expressions. The motion discriminator 14 discriminates the difference between the converted correct motion data and the converted generated motion data. The information processing device 10 of this embodiment also includes an update unit 15 that updates the motion generator 12 and the motion discriminator 14 using adversarial imitation learning.

[0035] With the above configuration, the information processing device 10 of this embodiment converts the robot's motion data (generated motion data) generated by the motion generator 12 and the expert's motion data (correct motion data) into the same intermediate body representation. This allows the motion discriminator 14 to compare the correct motion data with the generated motion data, even if the expert and robot have different shapes. This allows the motion data of a subject with a different shape from the robot to be used for learning the generator that generates the robot's motion, thereby increasing the amount of learning data for the target task. In other words, the information processing device 10 of this embodiment enables the transfer of capabilities to a robot with a different shape from the subject of the training data through the intermediate body representation, thereby enabling the transfer of work capabilities from a person capable of performing a target task to any robot capable of performing the target task. Note that the above method is not limited to the transfer of work capabilities from a human to a robot, but can also be applied to the transfer of work capabilities between subjects with different body representations. For example, the above method can also be applied to the transfer of work capabilities from a non-human subject to a robot, as long as actual motion data can be obtained from the subject.

[0036] Note that an imitation learning method other than adversarial imitation learning may be used. For example, an imitation learning method other than adversarial imitation learning that directly learns the behavior of training data may be used.

[0037] (First Modification of First Embodiment) In the above embodiment, a task of guiding a person has been described as an example of a task performed by a robot. In this modification, a pick-and-place task performed by an arm robot capable of performing a manipulation function will be described as another example of a task. The pick-and-place task is a task of grasping a specific object and transporting it to a specific position. The arm robot only needs to have a body expression that allows it to grasp the object to be transported.

[0038] In this modification, in the learning process of FIG. 2 , the task latent expression extractor 11 extracts latent expression vectors from task instructions including the task of "moving apple A from position A to position B" (step S101). Next, the motion generator 12 generates motion data representing the arm robot's execution of the task of "moving apple A from position A to position B" as generated motion data (step S102). Next, the intermediate physical expression converter 13 acquires, as correct motion data, motion data representing the execution of the task of "moving apple A from position A to position B" by an executor of the task of grasping and moving an apple (hereinafter referred to as a "movement executor"). The movement executor is an example of a teacher motion subject. The correct motion data includes the movement executor's skeleton, a video recording of the movement executor's motion, a trajectory of the movement executor's skeleton, the shape of apple A, and position information of apple A.

[0039] Next, the intermediate body representation converter 13 generates an intermediate body representation F that approximates the body representation D of the moving executor and the body representation E of the arm robot, as shown in Fig. 4 (step S104). The intermediate body representation F is composed of a combination of objects that approximate the gripping portion of the moving executor's hand and the gripping portion of the arm robot. In the example of Fig. 4, the intermediate body representation F is composed of three objects: an ellipse object F1, an ellipse object F2, and a fulcrum object F3. The ellipse objects F1 and F2 rotate around the fulcrum object F3.

[0040] Next, the intermediate body representation converter 13 converts the movement data defining the movement of the movement executor's body representation D into movement data defining the movements of the ellipse objects F1 and F2 of the intermediate body representation F, and converts the movement data defining the movement of the robot's body representation E into movement data defining the movements of the ellipse objects F1 and F2 of the intermediate body representation F (step S105). After that, the movement discriminator 14 repeatedly discriminates between the correct movement data converted into the intermediate body representations and the generated movement data converted into the intermediate body representations, and the update unit 15 repeatedly updates the movement generator 12 and the movement discriminator 14.

[0041] In this modification, too, by converting the movement data of the arm robot generated by the movement generator 12 (generated movement data) and the movement data of the moving executor (correct movement data) into the same intermediate body representation, even if the shape of the teacher movement subject used as teacher data differs from that of the robot, it becomes possible to compare the correct movement data with the generated movement data by the movement discriminator 14. Therefore, in this modification, too, movement data of a teacher movement subject with a shape different from that of the robot can be diverted as teacher data for learning by the generator that generates the robot's movements, making it possible to increase the amount of learning data for the target task.

[0042] (Second Modification of First Embodiment) In the above embodiment, a method for generating intermediate body representations on a rule basis has been mainly described. In this modification, an optimal intermediate body representation for a specific task is automatically searched for by repeatedly generating and selecting body representations.

[0043] In this modification, the information processing device 10 generates multiple intermediate body representation candidates (hereinafter referred to as intermediate body representation candidates) using a generative model, evaluates the performance of a search movement generator (hereinafter referred to as a search generator) trained using each intermediate body representation candidate, selects the most appropriate intermediate body representation candidate, and optimizes the generative model using the selection result. The information processing device 10 repeatedly generates intermediate body representation candidates and optimizes the generative model, and generates an optimal intermediate body representation using the optimized generative model.

[0044] As the generative model, for example, a generative AI technology that generates images from text, such as a diffusion model, or a base model, such as a large language model (LLM), can be used. In this modification, one search generator and one search classifier (hereinafter referred to as the search classifier) ​​are prepared for each intermediate body expression candidate.

[0045] (Operation Example) Next, an operation example of the information processing device 10 of this modified example will be described. FIG. 5 is a flowchart showing an example of the procedure of intermediate physical representation generation processing by the information processing device 10. The intermediate physical representation generation processing is processing for automatically searching for an appropriate intermediate physical representation for a specific task. Note that the processing procedure in the reliability determination processing described below is merely an example, and each process can be modified as appropriate as possible. Furthermore, steps in the processing procedure described below can be omitted, replaced, or added as appropriate depending on the embodiment.

[0046] In the intermediate body representation generation process, the information processing device 10 first generates multiple intermediate body representation candidates using a generative model (step S111). At this time, a latent representation vector is input into the generative model to obtain multiple intermediate body representation candidates. Next, the information processing device 10 updates the search generator using the adversarial imitation learning method described in the first embodiment (step S112). At this time, the information processing device 10 performs the adversarial imitation learning method using each intermediate body representation candidate instead of the intermediate body representation, and updates the search discriminator and the search generator. Next, the information processing device 10 selects an optimal intermediate body representation candidate using the updated search generator (step S113). At this time, the information processing device 10 evaluates the performance of each search generator updated using the adversarial imitation learning method, selects the search generator with the highest performance, and selects the intermediate body representation candidate used as the intermediate body representation in the learning of the selected search generator as the optimal intermediate body representation candidate. When evaluating performance, it is advisable to compare, for example, whether a specific task can be executed within a certain time, the time required to execute a specific task, whether multiple tasks can be executed, robustness, and the like.

[0047] Next, the information processing device 10 optimizes the generative model using the selection result of the optimal intermediate body representation candidate (step S114). At this time, the information processing device 10 updates the generative model so that an intermediate body representation candidate close to the optimal intermediate body representation candidate is generated. As a method for optimizing the generative model, for example, an evolutionary algorithm or Bayesian optimization can be used.

[0048] The information processing device 10 repeatedly executes the processes from step S111 to step S114 until the optimization of the generative model satisfies a termination condition (step S115). The termination condition is set using, for example, the number of repetitions or an evaluation value for the generative model. As a result, the generation of intermediate body representation candidates using the generative model and the optimization of the generative model are repeatedly executed, and a generative model that can generate a more appropriate intermediate body representation is automatically searched for.

[0049] In the information processing device 10 according to this modification, the intermediate body representation converter 13 can generate intermediate body representations using an optimized generative model. Furthermore, by using an optimized generative model, intermediate body representations optimized for a specific robot and a specific task can be used in learning by the action generator 12 using the intermediate body representations, thereby further improving the learning performance of the adversarial imitation learning method using intermediate body representations.

[0050] Note that intermediate body representations optimized for a specific robot and a specific task can also be used to train a motion generator using intermediate body representations for other robots. In this case, by reusing the intermediate body representations, the process of searching for the optimal intermediate body representation can be omitted, thereby making learning more efficient.

[0051] The method of this modification may also be used to search for a robot's optimal body representation for performing a task. For example, multi-arm, multi-joint body representations with high degrees of freedom are generated as intermediate body representation candidates, and a search motion generator selects an intermediate body representation candidate that satisfies an evaluation value. A body representation with reduced degrees of freedom of the selected intermediate body representation candidate is then regenerated as the intermediate body representation candidate. By repeating the above process, intermediate body representation candidates that have the required task execution performance and the minimum necessary degrees of freedom are generated. The finally generated intermediate body representation candidate can then be used as the robot shape optimal for performing the task.

[0052] Second Embodiment Fig. 6 is a diagram showing an example of the functional configuration of an information processing device 20 according to a second embodiment. The information processing device 20 is an inference device that performs inference using a generator generated by adversarial imitation learning. The information processing device 20 is a movement generation device that generates a robot movement based on a specific linguistic instruction using a movement generator trained by adversarial imitation learning using the information processing device 10 described in the first embodiment.

[0053] 6, the information processing device 20 includes a task latent expression extractor 21 and an action generator 22. The action generator 22 is a generator trained by adversarial imitation learning.

[0054] The task latent expression extractor 21 converts linguistic instructions for the robot to be trained into latent expression vectors, thereby extracting latent expression vectors from the linguistic instructions.

[0055] The motion generator 22 generates generated motion data representing the motion of the robot based on the latent representation vector and information about the robot that is to execute a task (hereinafter referred to as robot information). The robot information includes, for example, the size, shape, movable parts, skeleton, current posture, movement speed, and range of movement. As the motion generator 22, for example, the motion generator 12 trained by the information processing device 10 of the first embodiment can be used.

[0056] (Operation Example) Next, a description will be given of an operation example of the information processing device 20 configured as described above. Fig. 7 is a flowchart showing an example of the procedure of the operation data generation process by the information processing device 20. The operation data generation process is a process of generating operation data using the operation generator 22 for causing an arbitrary robot to execute an arbitrary task.

[0057] The processing procedure for the reliability determination process described below is merely an example, and each process can be modified as appropriate as possible. Furthermore, steps in the processing procedure described below can be omitted, replaced, or added as appropriate depending on the embodiment.

[0058] In the action data generation process, first, the task latent expression extractor 21 extracts a latent expression vector from a task instruction (step S201). The task latent expression extractor 21 transmits the extracted latent expression vector to the action generator 22.

[0059] Next, the motion generator 22 generates motion data for the robot based on the latent expression vector and the robot information (step S202). The generated motion data is output to, for example, a control device for the robot and used to control the motion of the robot.

[0060] (Actions and Effects) The information processing device 20 of this embodiment includes a task latent expression extractor 21 and an action generator 22. The task latent expression extractor 21 extracts latent expression vectors from linguistic instructions for a target robot. The action generator 22 generates action data based on the latent expression vectors and robot information. With the above configuration, highly accurate action data can be generated by using the highly accurate action generator 22 trained using the method described in the first embodiment.

[0061] (Application Example) Next, an example of the hardware configuration of the information processing apparatuses 10 and 20 according to the above embodiment and the modified example will be described with reference to the block diagram of FIG.

[0062] As shown in FIG. 8, the information processing devices 10 and 20 include, for example, a CPU (Central Processing Unit) 101, a ROM (Read Only Memory) 102, a RAM (Random Access Memory) 103, a display 104, a communication interface 105, and storage 106.

[0063] The CPU 101 is an integrated circuit capable of executing various programs and controls the overall operation of the information processing devices 10 and 20. The ROM 102 is, for example, a non-volatile semiconductor memory and stores programs and control data for controlling the information processing devices 10 and 20. The RAM 103 is, for example, a volatile semiconductor memory and is used as a work area for the CPU 101. The display 104 is, for example, a liquid crystal display or a touch panel display and displays information. The communication interface 105 is, for example, an input / output component for exchanging information with the outside, such as a USB, an HDMI (registered trademark), or a network interface. The storage 106 is a non-volatile storage device. The storage 106 stores system software and the like for the information processing devices 10 and 20. The CPU 101 may also be called a "processor."

[0064] In the above embodiments, the CPU 101 of the information processing devices 10 and 20 may be other circuits (or processors). For example, the information processing devices 10 and 20 may include a GPU (Graphics Processing Unit), an NPU (Neural Network Processing Unit), an MPU (Micro Processing Unit), or the like instead of a CPU. Each of the processes described in each embodiment may be realized by dedicated hardware. The processes of the information processing devices 10 and 20 may be a mixture of processes executed by software and processes executed by hardware, or may be only one of them.

[0065] The technical idea of ​​the present application can be executed by a computer based on a program that causes the computer to execute instructions shown in the processing procedures shown in the above-described embodiments and modifications. The technical idea of ​​the present application can also be realized as a program that can be provided over a network. The technical idea of ​​the present application can also be realized as a recording medium on which the program is recorded.

[0066] This invention is not limited to the above-described embodiments, and the components can be modified and embodied in practice without departing from the spirit of the invention. Furthermore, various inventions can be formed by appropriately combining multiple components disclosed in the above-described embodiments. For example, some components may be omitted from all the components shown in the embodiments. Furthermore, components from different embodiments may be appropriately combined.

[0067] REFERENCE SIGNS LIST 10, 20... Information processing device 11, 21... Task latent expression extractor 12, 22... Action generator 13... Intermediate body expression converter 14... Action discriminator 15... Update unit 101... CPU 102... ROM 103... RAM 104... Display 105... Communication interface 106... Storage

Claims

1. An information processing device comprising: a task latent expression extractor that extracts a latent expression vector from linguistic instructions given to a robot to be trained; a movement generator that generates generated movement data representing the movement of the robot based on the latent expression vector and information about the robot; an intermediate physical expression converter that acquires correct movement data when a teacher action subject having a physical expression different from the robot actually executes the linguistic instructions, generates an intermediate physical expression that is an intermediate expression between the physical expression of the teacher action subject and the physical expression of the robot, and converts each of the generated movement data and the correct movement data into movements using the intermediate physical expression; and a movement discriminator that discriminates between the converted correct movement data and the converted generated movement data.

2. The information processing device according to claim 1, further comprising an update unit that updates the action generator and the action discriminator by adversarial imitation learning.

3. An information processing method in which a task latent expression extractor extracts a latent expression vector from linguistic instructions for a robot to be learned, a movement generator generates generated movement data representing the movement of the robot based on the latent expression vector and information about the robot, an intermediate physical expression converter acquires correct movement data when a teacher action subject having a physical expression different from that of the robot actually executes the linguistic instructions, generates intermediate physical expressions that are intermediate expressions between the physical expression of the teacher action subject and the physical expression of the robot, and converts each of the generated movement data and the correct movement data into movements using the intermediate physical expressions, and a movement discriminator discriminates between the converted correct movement data and the converted generated movement data.

4. An information processing program for implementing in a computer the following functions: extracting a latent expression vector from linguistic instructions given to a robot to be learned; a movement generator generating generated movement data representing the movement of the robot based on the latent expression vector and information relating to the robot; an intermediate physical expression converter acquiring correct movement data when a teacher action subject having a physical expression different from that of the robot actually executes the linguistic instructions, generating an intermediate physical expression that is an intermediate expression between the physical expression of the teacher action subject and the physical expression of the robot, and converting each of the generated movement data and the correct movement data into movements using the intermediate physical expression; and a movement discriminator discriminating between the converted correct movement data and the converted generated movement data.

Images