Motion data design system, robot design system, motion data design method, and robot design method

Abstract

This invention provides a motion data design system that generates motion data for designing a low-degree-of-freedom robot to perform a task, based on a task performed by a robot capable of high-degree-of-freedom movements. [Solution] The motion data design system 100 of the present invention comprises an observation unit 110 that observes task execution movements by a robot or living organism including a human capable of high-degree-of-freedom movements, a dimensionality compression unit 120 that extracts features from the high-degree-of-freedom motion data observed by the observation unit 110 and compresses the dimensions into a trajectory in latent space, and a motion data generation unit 130 that generates low-degree-of-freedom motion data specific to the task based on the trajectory in latent space.

Description

[Technical Field]

[0001] The present invention relates to a system and method for generating motion data for designing a low-degree-of-freedom robot from tasks performed by a high-degree-of-freedom robot or a living organism such as a human, and a system and method for designing a robot using the generated motion data. [Background technology]

[0002] With an aging population and labor shortages, the use of robots to assist people is desired, for example, in nursing care settings. However, while high-degree-of-freedom robots can mimic and perform various human tasks, they are expensive, making social implementation difficult from a cost perspective. To further promote the spread of robots, the use of low-degree-of-freedom robots, which use fewer motors, is essential. The robots that are currently becoming widespread are low-degree-of-freedom robots specialized for specific tasks such as cleaning and food delivery, and are simple, low-cost, and easy to maintain.

[0003] There is no established method for efficiently designing low-degree-of-freedom robots specialized for such specific tasks.

[0004] Non-patent document 1 discloses a method for compressing high-dimensional human motion data into a low-dimensional latent space and identifying human motion within that latent space.

[0005] Patent Document 1 discloses a motion model learning device that improves the accuracy of posture estimation without manually specifying joint positions. The motion model learning device, during learning, constructs a single latent variable model M from information about all the joints in the body, and also selects a portion of the body's joints as points to learn a latent variable model M1 from information about the selected points. Through learning, it obtains a motion model that associates this latent variable model M with the latent variable model M1. As a result, during estimation, by referring to this motion model, it is possible to obtain the positions of all joints in the latent space of the latent variable model M based on the movement of a portion of the joints that are easy to recognize and track in a two-dimensional image. [Prior art documents] [Patent Documents]

[0006] [Patent Document 1] Japanese Patent Publication No. 2012-83955 [Non-patent literature]

[0007] [Non-Patent Document 1] Mariho Oyama, Ichiro Kobayashi, "An Effort Towards Generating Robot Motions Based on Human Motion Recognition in Latent Space," The 30th Annual Conference of the Japanese Society for Artificial Intelligence, 2016. [Overview of the project] [Problems that the invention aims to solve]

[0008] The present invention provides a motion data design system capable of generating motion data for designing a low-degree-of-freedom robot that can perform a task, based on a task performed by a robot or living organism such as a human capable of high-degree-of-freedom movements. [Means for solving the problem]

[0009] The motion data design system of the present invention comprises: an observation unit that observes task execution movements by a robot or living organism including a human capable of high-degree-of-freedom movements; a dimensionality reduction unit that extracts features from the high-degree-of-freedom motion data observed by the observation unit and compresses the dimensions into a trajectory in a latent space; and a motion data generation unit that generates task-specific low-degree-of-freedom motion data based on the trajectory in the latent space. [Effects of the Invention]

[0010] According to the present invention, motion data for designing a low-degree-of-freedom robot can be generated using motion data of a task performed by a robot or living organism such as a human, which is capable of high-degree-of-freedom movements.

Brief Description of the Drawings

[0011] [Figure 1] Block diagram showing the configuration of the operation data design system 100 and the robot design system 200 in the embodiment [Figure 2] Flowchart showing the operations of the operation data design system 100 and the robot design system 200 in the embodiment [Figure 3] Diagram showing the symmetric backward dressing method used as the dressing operation in the embodiment [Figure 4] Diagram showing the trajectory of the dimensionally compressed latent space in the embodiment [Figure 5] Diagram showing the automatic relevance determination (ARD) weights of each latent dimension in the embodiment [Figure 6] Diagram showing the directions of each axis for expressing the operation of the robot in the embodiment [Figure 7] Diagram showing the trajectory of the reconfigured operation data in the X-axis direction in the embodiment [Figure 8] Diagram showing the trajectory of the reconfigured operation data in the Y-axis direction in the embodiment [Figure 9] Diagram showing the trajectory of the reconfigured operation data in the Z-axis direction in the embodiment [Figure 10] Diagram showing the trajectory of the X-axis direction after performing the DTW processing in the embodiment [Figure 11] Diagram showing the trajectory of the Z-axis direction after performing the DTW processing in the embodiment [Figure 12] Diagram showing the outline of the cam mechanism in the embodiment [Figure 13] Diagram showing the cam profile created from the trajectory in the X-axis direction in the embodiment [Figure 14] Diagram showing the cam profile created from the trajectory in the Z-axis direction in the embodiment [Figure 15] Diagram showing the cam created using a 3D printer in the embodiment

Embodiments of the Invention

[0012] The embodiments of the present invention will be described below. However, the present invention is not limited to the embodiments described below.

[0013] (Embodiment)

[0014] [1. Structure] Figure 1 is a block diagram showing the configurations of the motion data design system 100 and the robot design system 200 in this embodiment. The motion data design system 100 includes an observation unit 110, a dimensionality compression unit 120, and a motion data generation unit 130. The robot design system 200 includes the observation unit 110, the dimensionality compression unit 120, and the motion data generation unit 130, in addition to a design data creation unit 140.

[0015] The motion data design system 100 generates motion data necessary for robot design. The robot design system 200 uses the motion data generated by the motion data design system 100 to design the robot's hardware.

[0016] The observation unit 110 observes task execution actions performed by a robot capable of high-degree-of-freedom movements and outputs data for learning the movements necessary for the robot to perform the task execution actions. A task execution action is a series of movements performed to achieve a predetermined purpose, such as a series of movements for assisting with dressing in elderly care, or a series of movements for making ramen or pasta. When observing complex task execution actions, imitation learning by a high-degree-of-freedom robot is efficient. However, the method of observing task execution actions is not limited to imitation learning by a high-degree-of-freedom robot; it may also be a method of observing task execution actions based on images of actual task execution actions performed by a person or animal, etc., taken from various angles.

[0017] There are two ways to represent robot motion: joint space representation, which uses joint angles, and task space representation, which uses the position and orientation of the end effector. When learning the latent space using task space, it is not possible to learn the variability needed to adapt to different orientations, so motion data is output in joint space.

[0018] The dimensionality reduction unit 120 extracts features from the high-degree-of-freedom motion data observed by the observation unit 110 and compresses the dimensions into a trajectory in latent space.

[0019] The dimensionality reduction unit 120 uses BGPLVM (Bayesian Gaussian Process Latent Variable Model) as a method for dimensionality reduction. BGPLVM can model complex data distributions with a small amount of data. BGPLVM is a dimensionality reduction method that maps high-dimensional observed data Y to low-dimensional latent variables X. BGPLVM is an extension of GPLVM (Gaussian Process Latent Variable Model). It is modeled by a Gaussian process, and observed data points y i The corresponding latent variable x i It is assumed that it is generated by [the following]. In BGPLVM, the relationship between observed data and latent variables is expressed by (Equation 1).

[0020]

number

[0021]

number

[0022] In (Mathematics 2), β -1 The accuracy of the observation noise is, I N The identity matrix is ​​Nx1(y i|0,K NN +β -1 I N ) has a mean of 0 and a covariance of K NN +β -1 I N Each shows a multivariate Gaussian distribution with a covariance of K NN K is an N×N covariance matrix composed of the kernel function k(x i ,x j ), and the automatic relevance determination (ARD) kernel can be applied and is defined by (Equation 3).

[0023]

Equation

[0024] In (Equation 3), α 2 f is the amplitude parameter of the kernel, α q indicates the relevance of each dimension, and Q is the dimension of the latent variable. By using the ARD kernel, the model can ignore the dimensions of latent variables with low relevance and enable learning of more concise and interpretable latent variables. The dimensionality reduction unit 120 can extract the characteristics of the task execution operation by performing dimensionality reduction using BGPLVM

[0025] The motion data generation unit 130 reconstructs the robot's motion from the trajectory in the latent space. The dimensionality reduction unit 120 calculates the trajectory in the latent space using BGPLVM, and the motion data generation unit 130 generates the robot's motion data from the trajectory in the latent space by performing a reverse process to the dimensionality reduction unit 120 using BGPLVM

[0026] The motion data generation unit 130 can generate robot motion data not only from the trajectory of the latent space generated by the dimensionality compression unit 120, but also from trajectories similar to it. This allows the generation of various motion data necessary to realize task execution, not just the task execution actions performed during observation. Based on the multiple motion data generated, the motion data generation unit 130 extracts the actions necessary to realize task execution. The motion data generation unit 130 outputs the extracted motion data as motion data necessary for robot design.

[0027] The design data creation unit 140 performs the hardware design of the robot that will perform the operation based on the operation data output by the operation data generation unit 130.

[0028] The motion data design system 100 and the robot design system 200 consist of memory and a processor, and each functional unit constituting the motion data design system 100 and the robot design system 200 is implemented by software in which the memory and processor work together. The motion data design system 100 and the robot design system 200 may also be implemented as programs that run on a computer. Furthermore, the implementation method for each functional unit is not limited to software; some parts may be implemented in hardware.

[0029] [2. Overview of Operation] The operation of the motion data design system 100 and robot design system 200, configured as described above, will now be outlined. Figure 2 is a flowchart showing the operation of the motion data design system 100 and robot design system 200 in this embodiment.

[0030] When generating robot motion data using the motion data design system 100, the first step is to determine the specific method of operation for the target task (S201). For example, if the task is to perform dressing assistance, the first step is to determine what specific actions will be taken to assist in dressing a person.

[0031] The observation unit 110 acquires operation data for the specific operation determined to perform the task execution operation (S202).

[0032] The dimensionality reduction unit 120 extracts features from the operation data output by the observation unit 110 and compresses the dimensions into a trajectory in the latent space (S203). The dimensionality reduction unit 120 extracts the features of the task execution operation by performing dimensionality reduction using BGPLVM.

[0033] The motion data generation unit 130 reconstructs the robot's movements from the trajectory in the latent space (S204). The motion data generation unit 130 generates robot motion data from the trajectory in the latent space by performing processing in the reverse direction of the processing performed by the dimensionality compression unit 120 using BGPLVM. The motion data generation unit 130 generates multiple motion data necessary to realize the task execution movement from multiple trajectories in the latent space. The motion data generation unit 130 extracts the movements indispensable for realizing the task execution movement from the generated multiple motion data and outputs them as motion data necessary for robot design.

[0034] The motion data design system 100 sequentially executes the operations S201 to S204 described above to generate the motion data necessary for robot design.

[0035] In addition to the steps in S201 to S204, the robot design system 200 performs hardware design for the robot that will perform the operation based on the motion data output by the motion data generation unit 130, and outputs the design data (S205).

[0036] [3. Specific Examples] The following describes, in more detail, the operation of the motion data design system 100 and the robot design system 200, focusing on the case where clothing assistance is performed as a task execution action.

[0037] [3-1. Determining Task Execution Behavior] There are two methods for using robots to assist with dressing: one using one arm, and another using both arms.

[0038] In the method using one arm, a mobile manipulator is used. The mobile manipulator puts the sleeves of clothing onto each arm. The advantage of the one-arm method is that it can reliably put the sleeves on the arms even if they are not symmetrical. However, there are challenges in that controlling the robot is very complex and it takes a long time to dress the person. Furthermore, the one-arm method is mainly for people who can move their arms and bodies freely, and the user is required to provide assistance such as holding the clothing in place. This is especially physically burdensome for the elderly, as it is difficult for them to raise their arms.

[0039] In the method using both arms, the robot grasps both ends of the clothing while dressing, making it primarily suitable for users who can maintain a symmetrical posture. Compared to the method using one arm, the method using both arms restricts the robot's movement due to the clothing, allowing for symmetrical movements, thus eliminating the need for high-degree-of-freedom control. For example, dressing is possible even if both arms perform the same movement.

[0040] Considering these advantages, a method using both arms is employed, which is less burdensome for the user and does not require complex movements. However, even with the two-arm method, the user still needs to raise their arms to shoulder height, which is burdensome for the user. Therefore, a new method called the symmetrical backward dressing method is adopted, which does not require the user to raise their arms high and allows the robot to automatically dress the user by first bringing their hands to the cuffs. As shown in Figure 3, the symmetrical backward dressing method is a method in which the robot dresses the user from behind, from the arms to the shoulders. The clothing to be dressed is limited to front-opening clothing. When dressing, the user needs to maintain a symmetrical seated posture.

[0041] The conditions under which this dressing method can be applied to users are as follows: • Can maintain a symmetrical posture • The shoulder joint can be flexed backward at least 15 degrees.

[0042] [3-2. Data Acquisition for Task Execution] The symmetrical backward-facing dressing method is likely to result in users adopting a variety of postures. Therefore, data will be collected for two postures: the standard posture with a straight back, which is typical for the average person, and the kyphosis posture, which is a posture with a significantly curved back that may be adopted by people with health problems.

[0043] In this specific example, with the cooperation of a nurse with 26 years of nursing experience, motion data is acquired by having the nurse teach the robot how to perform dressing assistance actions, such as putting the sleeves of a garment over a patient (a mannequin used for nursing training called "Yururi-chan") and pulling the garment up to the front of the shoulders. The robot is started in puppet mode (symmetric mode), and when one arm moves, the other arm also performs a symmetrical movement. The observation unit 110 acquires motion data of the robot's 14 axes through a learning approach based on demonstrations in which the robot actually performs dressing assistance actions.

[0044] The observation unit 110 resamples each of the 14 acquired robot joint data points to 200 points, creates training data by concatenating the two datasets, and provides it as input to the dimensionality reduction unit 120.

[0045] [3-3. Dimensional compression into latent space] The dimensionality reduction unit 120 uses BGPLVM to extract features from the high-degree-of-freedom motion data observed by the observation unit 110 and compresses the dimensions into a trajectory in latent space.

[0046] The initial values ​​of the latent space are set using data-based principal component analysis (PCA). The latent space to be learned is set to 5 dimensions. 100 induction points are specified, and an automatic association detection (ARD) kernel is used. To ensure efficient learning convergence, the variance is constrained by setting the signal-to-noise ratio (SNR) to 10 dB in the first 20 iterations. After that, 2000 learning iterations are performed without constraints to obtain the latent space.

[0047] Figure 4 shows the trajectory of the latent space after dimensionality reduction by the dimensionality reduction unit 120. In Figure 4, the two trajectories are plots of data learned in the standard posture and data learned in the kyphosis posture. Figure 5 shows the ARD weights for each latent dimension. The ARD weights show that the two latent dimensions have significant values, indicating that the robot's movements can be represented in two dimensions in the latent space.

[0048] [3-4. Reconstructing robot motion from trajectories in latent space] The motion data generation unit 130 uses BGPLVM to reconstruct the robot's motion data from the trajectory in the latent space. It traces the trajectory in the latent space and reconstructs the robot's joint angles based on that trajectory. In fact, when the motion data generation unit 130 moves the robot based on the robot motion generated from the trajectory in the latent space, it can be confirmed that it is possible to perform clothing-wearing operations. It can be confirmed that by tracing the trajectory of a standard posture in the latent space, clothing-wearing operations can be performed for a user in a standard posture, and by tracing the trajectory of a kyphosis posture in the latent space, clothing-wearing operations can be performed for a user in a kyphosis posture.

[0049] By reconstructing robot motion data from trajectories that pass between learned trajectories, it is possible to generate unknown clothing-wearing movements. These movements can accommodate moderate kyphosis. In fact, when the robot is moved based on motion data generated from trajectories that pass between two learned trajectories, it can be confirmed that it can perform clothing-wearing movements for users with moderate kyphosis. This demonstrates that the latent space learned by BGPLVM can generate unknown clothing-wearing movements.

[0050] The robot's motion is reconstructed from the latent space by tracing the trajectories near and between the trajectories shown in Figure 4. In this specific example, 10 trajectories are specified in the latent space, and 10 robot motions are generated. Since the robot was taught in puppet mode (symmetric mode) during imitation learning, the robot motions generated from the latent space are also symmetric. In this specific example, the robot's hardware design focuses on the motion of one arm. In hardware design, it is difficult to design with low degrees of freedom if the generated joint space motion is the focus, so the task space motion is focused.

[0051] Figure 6 shows the directions of each axis to represent the robot's movement in the task space. Figures 7 to 9 show the trajectories of the robot's tip generated by the motion data generation unit 130. Of the generated trajectories, Figure 7 shows the trajectory in the X-axis direction, Figure 8 shows the trajectory in the Y-axis direction, and Figure 9 shows the trajectory in the Z-axis direction. In Figures 7 to 9, minimum-maximum normalization is performed on the time axis direction. The generated trajectories show characteristic trajectories in the X-axis and Z-axis directions.

[0052] The generated trajectory reveals that in the latent space, important features are learned for determining the movement of the end effector in the X and Z axes, while the Y axis is less important with respect to clothing. Based on this understanding of the latent space, the hardware is designed focusing only on the movement in the X and Z axes.

[0053] To further clarify the relationship between the movements of the X and Z axis trajectories, dynamic time stretching (DTW) is used to visualize the trajectory correspondence. The representative DTW trajectory is determined by calculating the DTW for all trajectories and selecting the one with the smallest distance from other trajectories. Figure 10 shows the DTW-processed X-axis trajectory, and Figure 11 shows the DTW-processed Z-axis trajectory. In Figures 10 and 11, the thick lines represent the representative trajectory, and the DTW is calculated based on this trajectory. Based on the information obtained from the representative trajectory, the robot hardware is designed to execute various trajectories by controlling the rotational speed in the time axis direction.

[0054] [3-5. Robot Hardware Design] As an example of designing hardware based on typical trajectories, we design a cam mechanism. Figure 12 shows an overview of the cam mechanism. The cam mechanism consists of a cam whose shape is determined by its cam profile, and a follower that generates a one-dimensional trajectory by contacting the cam. The follower's operation depends on the cam design. The cam mechanism can generate rich one-dimensional trajectories, and the cam profile can be efficiently designed based on time-series data. To design a smooth cam, a time average with a window size of 100 is applied. An offset is set in the smoothing data, and the minimum value is set as the minimum radius of the cam. The basic cam profile equation is as follows: x = r(θ)·cos(θ) y = r(θ)·sin(θ) Assuming an angular velocity of 1, θ = t. The time-series trajectory is assumed to pass through 0 to 270 degrees, and the trajectory from 270 to 360 degrees returns to the initial position. r(θ) = r(t) is r(t) = trajectory(t) between 0 and 270 degrees, where trajectory(t) is all the trajectory data. r(θ) from 270 to 360 degrees is linearly preserved. The cams created based on the above method are shown in Figures 13 and 14. Figure 13 shows the cam profile created from the trajectory in the X-axis direction, and Figure 14 shows the cam profile created from the trajectory in the Z-axis direction.

[0055] [3-6. Evaluation of Robot Hardware Design] Based on the calculated cam profile, a 1 / 10 scale cam mechanism is designed using a 3D printer, and its operation is verified to evaluate the design method. The cam mechanism created using the 3D printer is shown in Figure 15.

[0056] The system uses the HEBI (HEBI Robotics) robot module to drive the system. Motor speed is determined based on time expansion and contraction information obtained during DTW analysis. To demonstrate whether this mechanism can dress a human body and adapt to different postures such as kyphosis, a 1 / 10 scale mannequin will be made to assume a standard posture and a kyphosis posture. In this demonstration, the robot will not wear clothes, and the possibility of dressing will be evaluated by observing the trajectory of the robot's tip. The robot speed is determined by dividing the dressing motion into 10 equal phases and calculating the ratio of compression time to expansion time in each phase.

[0057] In the standard posture, the robot assumes a position that approaches the user, and at this point, the vertical displacement is small. After the robot has approached the user sufficiently, it begins to displace upward. On the other hand, in the kyphosis posture, the upward displacement begins earlier than in the standard posture.

[0058] Demonstrations using a 1 / 10 scale prototype of the created cam confirmed that it could generate trajectories that followed a person's arm in both a standard and kyphosis posture. These trajectories reproduced the taught movements performed by actual humans, particularly the movement of putting clothing along the arm and over the shoulder. The trajectories showed that the characteristics of clothing movements are represented in a low-dimensional latent space, demonstrating that the designed cam has the ability to represent characteristic movements, thus confirming the effectiveness of the method.

[0059] [4. Other specific examples] In the above example, motion data was acquired for two postures during the dressing process: a standard posture and a kyphosis posture, and two trajectories were generated in the latent space. The number of trajectories in the latent space is not limited to two; there may be only one, or three or more. Because BGPLVM is used for dimensionality reduction, by acquiring data for the dressing process in at least one posture and generating a trajectory in the latent space, the robot's motion can be reconstructed by tracing the vicinity of that trajectory.

[0060] The above example illustrates the use of a cam mechanism as the hardware for a robot. However, the hardware implementation method is not limited to a cam mechanism. For example, it could be implemented using a combination of wires and motors.

[0061] [5. Effects] The motion data design system 100 in this embodiment includes an observation unit 110 that observes task execution movements by a robot or living organism including a human capable of high-degree-of-freedom movements, a dimensionality reduction unit 120 that extracts features from the high-degree-of-freedom motion data observed by the observation unit 110 and compresses the dimensions into a trajectory in latent space, and a motion data generation unit 130 that generates task-specific low-degree-of-freedom motion data based on the trajectory in latent space.

[0062] This allows for the generation of motion data for designing low-degree-of-freedom robots using motion data from tasks performed by robots or living organisms such as humans that are capable of high-degree-of-freedom movements. By designing the robot's hardware using this motion data, the system can be simplified, costs reduced, and maintenance made easier, accelerating the social implementation of robots.

[0063] The dimension reduction unit 120 may use BGPLVM to perform compression to a lower-dimensional latent space.

[0064] By using the statistical method BGPLVM to reduce dimensionality, it is possible to generate trajectories in the latent space without requiring the large amount of motion data necessary for training in neural networks and the like.

[0065] The motion data generation unit 130 may generate multiple motion data corresponding to a task from multiple trajectories in the latent space, determine the degrees of freedom necessary to realize the task based on the generated motion data, and generate motion data for the determined degrees of freedom.

[0066] Multiple motion data generated from multiple trajectories in the latent space allow us to understand which axis of motion is important in the latent space, based on the features learned to determine the movement of the end effector in the task space. This enables low-degree-of-freedom hardware design that focuses solely on the motion of that axis.

[0067] The observation unit 110 outputs task motion data in the joint space, which is represented by joint angles, and the motion data generation unit 130 may generate motion data in the task space, which is represented by the position and orientation of the end effector, when generating multiple motion data corresponding to a task from multiple trajectories in the latent space.

[0068] By using joint space when learning the latent space, the problem of not being able to learn the variability to adapt to different poses in the task space can be avoided. Furthermore, by generating motion data in the task space from the trajectory in the latent space, it becomes easier to design with low degrees of freedom.

[0069] The robot design system 200 in this embodiment includes: an observation unit 110 that observes task execution movements by a robot or living organism including a human capable of high-degree-of-freedom movements; a dimensionality reduction unit 120 that extracts features from the high-degree-of-freedom movement data observed by the observation unit 110 and compresses the dimensions into a trajectory in latent space; a movement data generation unit 130 that generates low-degree-of-freedom movement data specific to the task based on the trajectory in latent space; and a design data creation unit 140 that creates design data for a robot that realizes the task based on the low-degree-of-freedom movement data generated by the movement data generation unit 130.

[0070] This allows for the hardware design of low-degree-of-freedom robots using motion data from tasks performed by robots capable of high-degree-of-freedom movements or by living organisms such as humans. This simplifies the system, reduces costs, simplifies maintenance, and accelerates the social implementation of robots.

[0071] The design data creation unit 140 may create design data for a robot using a cam mechanism.

[0072] By using a cam mechanism, a rich one-dimensional trajectory can be generated, and cam profiles can be efficiently designed based on time-series data.

[0073] The motion data design method in this embodiment comprises: an observation step (S202) of observing task execution movements by a robot or living organism including a human capable of high-degree-of-freedom movements; a dimensionality reduction step (S203) of extracting features from the high-degree-of-freedom motion data observed in the observation step (S202) and compressing them into a trajectory in latent space; and a motion data generation step (S204) of generating low-degree-of-freedom motion data specific to the task based on the trajectory in latent space.

[0074] This allows for the generation of motion data for designing low-degree-of-freedom robots using motion data from tasks performed by robots or living organisms such as humans that are capable of high-degree-of-freedom movements. By designing the robot's hardware using this motion data, the system can be simplified, costs reduced, and maintenance made easier, accelerating the social implementation of robots.

[0075] The robot design method in this embodiment comprises: an observation step (S202) of observing task execution movements by a robot or living organism, including a human, capable of high-degree-of-freedom movements; a dimensionality reduction step (S203) of extracting features from the high-degree-of-freedom movement data observed in the observation step (S202) and compressing them into a trajectory in latent space; a movement data generation step (S204) of generating low-degree-of-freedom movement data specific to the task based on the trajectory in latent space; and a design data creation step (S205) of creating design data for a robot that realizes the task based on the low-degree-of-freedom movement data generated in the movement data generation step (S204).

[0076] This allows for the hardware design of low-degree-of-freedom robots using motion data from tasks performed by robots capable of high-degree-of-freedom movements or by living organisms such as humans. This simplifies the system, reduces costs, simplifies maintenance, and accelerates the social implementation of robots.

[0077] (Other embodiments) As described above, embodiments have been explained as examples of the present invention. However, the present invention is not limited thereto and can also be applied to embodiments that have been modified, replaced, added, omitted, etc.

[0078] Therefore, other embodiments are illustrated below.

[0079] In this embodiment, the dimensionality reduction unit 120 was described in a form that uses BGPLVM as the dimensionality reduction method. The dimensionality reduction method is not limited to BGPLVM; it may also be GPDM (Gaussian Process Dynamical Model) or machine learning methods such as neural networks. By using statistical methods such as BGPLVM or GPDM to learn the trajectory in the latent space, the large amount of motion data required for training in a neural network becomes unnecessary. When using a neural network, training with a large amount of motion data is required, but dimensionality reduction is still possible.

[0080] In this embodiment, a configuration targeting dressing assistance as the task execution operation was described. However, the task execution operation is not limited to dressing assistance; various operations are possible. For example, a series of operations for making ramen or pasta, operations for assisting a person with eating in caregiving, or a series of operations performed by a craftsman in a factory. In the case of a series of operations for making ramen or pasta, the series of operations performed by a person during the process are observed to acquire operation data, which is then compressed into a latent space, and the robot operation is reconstructed based on the trajectory in the latent space. Based on the reconstructed operation, a low-degree-of-freedom cooking robot for making ramen or pasta can be designed.

[0081] Task execution actions are not limited to actions performed by humans; they may also be actions performed by animals such as cats or dogs. For example, by observing a cat's affectionate behavior towards a human, motion data can be acquired, compressed into a latent space, and the robot's movements can be reconstructed based on the trajectory within that latent space. Based on the reconstructed movements, it is possible to realize a low-degree-of-freedom cat-like robot that performs affectionate behavior towards a human.

[0082] In this embodiment, since the movement of the end-effector in the X and Z axes is important for clothing assistance operations, design data for a robot with two degrees of freedom was created based on motion data reconstructed from latent space. Depending on the task execution operation, three degrees of freedom or four or more, including the rotation direction of each axis, may be required. In such cases, the degree of freedom required to achieve the task may be determined from the motion data reconstructed from latent space, and design data with that degree of freedom (e.g., three degrees of freedom) (e.g., design data in the three axes of X, Y, and Z) may be created.

[0083] Depending on the target task execution motion, it may be difficult to reduce the degrees of freedom required to achieve that task based on motion data reconstructed from the trajectory in latent space. In such cases, the execution motion of the target task may be reviewed, the method of operation may be changed, and the process from acquiring task execution motion data to reconstructing robot motion may be repeated to determine whether the task can be achieved with a low degree of freedom. By repeating this procedure, a method for achieving the target task with a low-degree-of-freedom robot may be found. [Explanation of symbols]

[0084] 100 Operational Data Design System 110 Observation Unit 120-dimensional compression section 130 Operation data generation unit 140 Design Data Creation Department 200 Robot Design Systems

Claims

1. An observation unit that observes task execution actions by robots or living organisms, including humans, capable of high degrees of freedom, A dimensionality reduction unit extracts features from the high-degree-of-freedom motion data observed by the aforementioned observation unit and compresses the dimensions into a trajectory in latent space, A motion data generation unit that generates low-degree-of-freedom motion data specific to the task based on the trajectory in the latent space, A system for designing operational data equipped with the necessary features.

2. The operation data design system according to claim 1, wherein the dimensionality compression unit performs compression to a low-dimensional latent space using BGPLVM (Bayesian Gaussian Process Latent Variable Model).

3. The motion data design system according to claim 1, wherein the motion data generation unit generates a plurality of motion data corresponding to the task from a plurality of trajectories in a latent space, determines the degree of freedom necessary to realize the task based on the generated plurality of motion data, and generates motion data for the determined degree of freedom.

4. The observation unit outputs operation data of the task in the joint space, which is represented by the joint angle. The motion data design system according to claim 3, wherein the motion data generation unit generates motion data in a task space represented by the position and orientation of the end effector when generating a plurality of motion data corresponding to the task from a plurality of trajectories in the latent space.

5. An observation unit that observes task execution actions by robots or living organisms, including humans, capable of high degrees of freedom, A dimensionality reduction unit extracts features from the high-degree-of-freedom motion data observed by the aforementioned observation unit and compresses the dimensions into a trajectory in latent space, A motion data generation unit that generates low-degree-of-freedom motion data specific to the task based on the trajectory in the latent space, Based on the low-degree-of-freedom motion data generated by the motion data generation unit, a design data creation unit is provided to create design data for a robot that performs the task. A robot design system equipped with [features / equipment].

6. The robot design system according to claim 5, wherein the design data creation unit creates design data for a robot using a cam mechanism.

7. An observation step to observe the task execution actions of a robot or living organism, including a human, capable of high-degree-of-freedom movements, A dimensionality reduction step is performed to extract features from the high-degree-of-freedom motion data observed in the aforementioned observation step and compress them into a trajectory in latent space, A motion data generation step that generates low-degree-of-freedom motion data specific to the task based on the trajectory in the latent space, A method for designing operational data.

8. An observation step to observe the task execution actions of a robot or living organism, including a human, capable of high-degree-of-freedom movements, A dimensionality reduction step is performed to extract features from the high-degree-of-freedom motion data observed in the aforementioned observation step and compress them into a trajectory in latent space, A motion data generation step that generates low-degree-of-freedom motion data specific to the task based on the trajectory in the latent space, Based on the low-degree-of-freedom motion data generated in the motion data generation step, a design data creation step is performed to create design data for a robot that performs the task. A robot design method that includes [a specific feature / feature].

Images