Whole-body control method and control device of robot, storage medium and robot

Abstract

The invention provides a whole-body control method and device of a robot, a storage medium and the robot, and belongs to the technical field of humanoid robot control. The method comprises the following steps: training a basic whole body control model by using an AMASS data set; training a sub-control model corresponding to each motion type on the basis of the basic whole body control model by using the plurality of clustering sub-data sets; after the trained sub-control model is deployed to a control module of the robot, an execution track sequence of the robot is collected by controlling the robot to move; utilizing the collected execution track sequence to train an action increment model corresponding to the sub-control model so as to adjust the corresponding sub-control model; and fusing the adjusted sub-control models by using a knowledge distillation technology to obtain an optimized whole-body control model. According to the embodiment of the invention, a'general-expert-general 'training framework is provided and is used for learning a general whole-body control strategy, and the problem of interference among different motion types is effectively relieved.

Description

Technical Field

[0001] This application belongs to the technical field of humanoid robot control, and particularly relates to a full-body control method, a control device, a storage medium and a robot for a robot. Background Art

[0002] Due to the morphological advantages of the human body structure, humanoid robots are capable of performing a variety of complex daily tasks, such as grasping, carrying, walking, and crossing irregular terrains. However, the full-body control problem of humanoid robots remains a major challenge in current robot control research. The main reasons are the limited hardware of humanoid robots, the high dimensionality of the action space, the complex control coupling, etc. Especially when performing agile actions with high degrees of freedom and high dynamic requirements, higher requirements are put forward for the generalization ability, robustness and accuracy of the control strategy.

[0003] The related technologies to solve the above problems are mostly based on dynamic modeling and optimal control technologies, but it is difficult to adapt to the changes in complex task scenarios. Reinforcement learning combined with simulation-to-real technology shows good prospects in controlling complex full-body actions. However, such methods usually have difficulty in taking into account the requirements of different action types. Summary of the Invention

[0004] The purpose of this application is a full-body control method, a control device, a storage medium and a robot for a robot, aiming to effectively alleviate the interference problem between different motion types.

[0005] According to the first aspect of this application, a full-body control method for a robot is provided. The full-body control method includes: obtaining an AMASS dataset and training a basic full-body control model using the AMASS dataset; structuring and dividing the AMASS dataset using a clustering algorithm to obtain multiple clustered sub-datasets, and using the multiple clustered sub-datasets to train sub-control models corresponding to each motion type on the basis of the trained basic full-body control model; after deploying the trained sub-control models to the control module of the robot, collecting the execution trajectory sequence of the robot by controlling the movement of the robot; training an action increment model corresponding to the sub-control model using the collected execution trajectory sequence to adjust the corresponding sub-control model through the trained action increment model; using knowledge distillation technology to fuse the adjusted sub-control models to obtain an optimized full-body control model; and using the optimized full-body control model to perform full-body control on the robot.

[0006] In an alternative embodiment, after obtaining the AMASS dataset, the full-body control method further includes: obtaining the HumanML3D dataset; and encoding the AMASS dataset and the HumanML3D dataset based on motion features and action text descriptions using an autoencoder and a self-attention mechanism network to complete the training of the basic full-body control model.

[0007] In an alternative embodiment, encoding the AMASS dataset includes: obtaining the SMPL-format action sequence in the AMASS dataset; and encoding the action sequence using a self-attention mechanism network structure, including converting joint axis angles and root coordinates into three-dimensional coordinates of joint points based on the SMPL-format action sequence, deleting redundant joint points, and introducing the moving speed of the corresponding part relative to the world coordinate system.

[0008] In an alternative embodiment, encoding the HumanML3D dataset includes: serializing the HumanML3D dataset using a BERT model to obtain a text sequence; and encoding the text sequence using a self-attention mechanism network structure.

[0009] In an alternative embodiment, training the action increment model corresponding to the sub-control model includes: constructing a simulator. Re-executing the corresponding motion in the simulator, and training the action increment model based on the collected execution trajectory sequence, including: initializing the first state collected based on the collected execution trajectory sequence; calculating a reward signal to minimize the difference between the simulated next state and the collected next state; and optimizing the action increment model using the PPO algorithm.

[0010] In an alternative embodiment, adjusting the corresponding sub-control model by the trained action increment model includes: constructing a new simulator; and in the new simulator, fixing the parameters of the action increment model and adjusting the corresponding sub-control model.

[0011] In an alternative embodiment, the use of the knowledge distillation technique to fuse the adjusted sub-control models to obtain an optimized full-body control model includes: taking each adjusted sub-control model as a teacher model, and constructing a student model and a dynamic gating network. Using the dynamic gating network, generate the weight distribution of each teacher model. Synchronously perform the following operations within a preset time window: drive each teacher model and the student model to generate a predictive control sequence and a state trajectory; fuse the outputs of each teacher model according to the weight distribution to generate a reference control sequence and a reference state trajectory; and calculate the spatio-temporal joint loss between the output of the student model and the reference control sequence. Deploy the student model in a simulation environment to perform closed-loop control, calculate the performance loss based on the collected trajectory data, and jointly optimize the student model with the spatio-temporal joint loss; and use the optimized student model as the optimized full-body control model.

[0012] According to a second aspect of the present application, there is provided a control device, the control device comprising: a memory, a processor, and a computer program stored on the memory and executable on the processor, the processor executing the computer program to implement the above-mentioned full-body control method.

[0013] According to a third aspect of the present application, there is provided a machine-readable storage medium, on which instructions are stored, and the instructions cause the machine to execute the above-mentioned full-body control method.

[0014] According to a fourth aspect of the present application, there is provided a robot, the robot comprising the above-mentioned control device.

[0015] Through the above technical solutions, the whole-body control method for a robot provided by the embodiments of the present application includes: obtaining the AMASS dataset and training a basic whole-body control model using the AMASS dataset; structuring and partitioning the AMASS dataset using a clustering algorithm to obtain multiple clustered sub-datasets, and using the multiple clustered sub-datasets to train sub-control models corresponding to each motion type on the basis of the basic whole-body control model; after deploying the trained sub-control models to the control module of the robot, collecting the execution trajectory sequence of the robot by controlling the robot's motion; using the collected execution trajectory sequence to train an action increment model corresponding to the sub-control model, so as to adjust the corresponding sub-control model through the trained action increment model; using knowledge distillation technology to fuse the adjusted sub-control models to obtain an optimized whole-body control model; and using the optimized whole-body control model to perform whole-body control on the robot. The embodiments of the present application provide a training framework of "general-expert-general" (i.e., "basic whole-body control model - multiple sub-control modules - optimized whole-body control model") for learning general whole-body control strategies, effectively alleviating the interference problem between different motion types. The embodiments of the present application also provide a clustering method based on an autoencoder (AE), which combines the action features of each part and the action text description to structurally partition the dataset, so as to better train the expert whole-body control strategy. By clustering the action data, for each motion subset, the corresponding action increment model is trained separately. Since the same type of actions has consistency in dynamic characteristics, this expert training can significantly improve the fitting accuracy of the action increment model and more efficiently correct the difference between the simulation and the real environment. Compared with the general action increment model, the action increment model provided by the embodiments of the present application has significant advantages in terms of action compensation accuracy, training convergence speed, and final control performance. In the general strategy training stage, the embodiments of the present application fuse the knowledge of each expert strategy through expert knowledge distillation to obtain a general whole-body control strategy that can handle various complex and agile motions.

[0016] Other features and advantages of the present application will be described in the following specification, and will become apparent in part from the specification, or can be understood by implementing the present application. The objectives and other advantages of the present application can be achieved and obtained through the structures and processes pointed out in the specification and the drawings. Brief Description of the Drawings

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following will briefly introduce the drawings required for use in the description of the embodiments or related technologies. Obviously, the following drawings are some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0018] Figure 1 It is a schematic flowchart of a full-body control method for a robot provided by an exemplary embodiment of the present application.

[0019] Figure 2 It is a schematic flowchart of the full-body control method provided by an exemplary embodiment of the present application. Detailed implementation manners

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Apparently, the described embodiments are some but not all of the embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts shall fall within the protection scope of the present application.

[0021] As described above, reinforcement learning combined with simulation-to-real technology shows good prospects in controlling complex full-body movements. However, such methods usually have difficulty in taking into account the requirements of different action types. The control focus of different action types varies significantly. For example, jumping actions pay more attention to high-torque control, while stable walking pays more attention to balance and continuity. Directly training a unified policy is likely to cause gradient conflicts between tasks, resulting in performance degradation.

[0022] In response to this, the embodiments of the present application provide a full-body control method for a robot, which is optimized from two directions of model structure and data structure to provide a general full-body control strategy for humanoid robots for executing diverse and agile actions.

[0023] Figure 1 It is a schematic flowchart of a full-body control method for a robot provided by an exemplary embodiment of the present application. The full-body control method may include the following steps:

[0024] Step S110: Obtain the AMASS dataset and use the AMASS dataset to train a basic full-body control model.

[0025] Among them, the AMASS (Archive of Motion Capture as Surface Shapes) dataset is a large-scale, unified-format 3D human motion capture (MoCap) dataset. The AMASS dataset can be obtained based on the AMASS database, and the AMASS database is a JSON-format file that stores the metadata of each motion capture dataset and the parameters of the SMPL human model.

[0026] In a preferred embodiment of the present application, after obtaining the AMASS dataset, the full-body control method may further include: obtaining the HumanML3D dataset; and encoding the AMASS dataset and the HumanML3D dataset using an autoencoder and a self-attention mechanism network based on motion features and action text descriptions.

[0027] Among them, in addition to the action sequences expressed in the SMPL format that come with the AMASS dataset, the embodiments of the present application also use the HumanML3D open-source dataset to provide text annotations and corresponding time frame ranges for most of the data in the AMMAS dataset.

[0028] In a preferred embodiment of the present application, encoding the AMASS dataset may include: obtaining the action sequences in the SMPL format in the AMASS dataset; and encoding the action sequences using a self-attention mechanism network structure, including converting joint axis angles and root coordinates to three-dimensional coordinates of joint points based on the action sequences in the SMPL format, deleting redundant joint points, and introducing the moving speed of the corresponding parts relative to the world coordinate system.

[0029] In a preferred embodiment of the present application, encoding the HumanML3D dataset may include: serializing the HumanML3D dataset using the BERT model to obtain a text sequence; and encoding the text sequence using a self-attention mechanism network structure.

[0030] Please refer to Figure 2 For example, taking the motion features of leg joint points and action text descriptions as an example, the embodiments of the present application use an autoencoder (AE) and a self-attention mechanism network to encode the AMASS dataset and the HumanML3D dataset. For example, encoding the action using a self-attention mechanism network structure, for example, converting joint axis angles and root coordinates to three-dimensional coordinates of joint points based on forward dynamics, deleting redundant joint points, and introducing the moving speed of the feet relative to the world coordinate system. For text data, it is serialized through the BERT model and also processed through a self-attention mechanism network structure to obtain a latent space representation corresponding to the action modality.

[0031] For example, after processing the AMASS dataset and the HumanML3D dataset, divide the training set, validation set, and test set; based on reinforcement learning for model training, a basic full-body control model (which can also be called a general full-body control strategy or a general expert strategy) can be obtained. It should be noted that the embodiments of the present application do not limit the above-mentioned reinforcement learning models (networks), and any model (network) that can be used for humanoid robot control belongs to the scope protected by the present application.

[0032] Among them, the loss function is, for example: L = MSE(z text , z motion ) + Huber(M' text , M') + Huber(M' motion , M'). Among them, MSE(z text , z motion ) represents the intermediate-layer latent variable of the text and the action modality, Huber() represents the loss function, M' is the key feature of the sequence, and M' text and M' motion are the reconstruction results of the text modality and the action modality, respectively.

[0033] As described above, different from the method of directly using an autoencoder to reconstruct the complete trajectory, the embodiments of the present application introduce text annotation to align actions at the semantic level, so as to ensure that actions with different styles but the same semantics are closer in the latent space. For example, the walking action can be moving straight forward or walking in a circle. If only relying on autoencoder training, it is difficult to reflect the similarity between the two, while through text semantic alignment, this problem can be effectively overcome. The embodiments of the present application effectively improve the aggregation of semantically similar actions in the latent space by introducing text alignment. Further, the embodiments of the present application do not directly use the action sequence expressed in the SMPL format for reconstruction. Since SMPL mainly represents joint angles and base coordinate transformations, this representation method is difficult to intuitively reflect the concerned action features. Therefore, the embodiments of the present application, based on forward dynamics, convert joint axis angles and root coordinates into three-dimensional coordinates of joint points; delete redundant joint points, and introduce the moving speed of the corresponding part (for example, the foot) relative to the world coordinate system to enhance the discrimination ability of action categories such as jumping, standing, and moving. The reconstruction content only includes some key joint points, foot speed, root node speed, and rotation information, so as to focus on the core action features.

[0034] Step S120: Structurally divide the AMASS dataset using a clustering algorithm to obtain multiple clustered sub-datasets, and use the multiple clustered sub-datasets to train sub-control models corresponding to each motion type on the basis of the trained full-body control model.

[0035] In the embodiments of the present application, the K-means clustering method can be used to cluster the encoded data into multiple categories. Among them, the K-means clustering algorithm is an iterative clustering analysis algorithm. For example, the data can be divided into K groups, and then K objects are randomly selected as the initial clustering centers; the distance between each object and each seed clustering center is calculated, and each object is assigned to the clustering center closest to it. The clustering centers and the objects assigned to them represent a cluster. Each time a sample is assigned, the clustering centers are recalculated based on the existing objects in the cluster. This process will be repeated continuously until a certain termination condition is met. In the embodiments of the present application, for the selection of the K value, the elbow method can be used to determine the number of clusters, and the trade-off point between using fewer categories and minimizing the sum of squared intra-class distances (i.e., the total squared distance from all points to their respective class centers) is used to determine the K value.

[0036] Please refer to Figure 2 , for example, using the clustering algorithm to structurally partition the encoded AMASS dataset to obtain multiple clustered sub-datasets; in each clustered sub-dataset, fine-tuning is performed based on the basic full-body control model to obtain sub-control models corresponding to each motion type, and each sub-control model can be an expert control strategy for a specific motion type.

[0037] Similar to the foregoing, the embodiments of the present application do not limit the construction and training process of the sub-control model (network), and any model (network) that can be used for controlling various motion types of the humanoid robot belongs to the scope protected by the present application.

[0038] The basic full-body control model has basic motion tracking and execution capabilities; in the embodiments of the present application, taking the basic full-body control model as the initialization point, fine-tuning is performed separately on each clustered sub-dataset to train multiple sub-control models to adapt to their respective specific motion patterns and control preferences.

[0039] Step S130: After deploying the trained sub-control model to the control module of the robot, collect the execution trajectory sequence of the robot by controlling the movement of the robot.

[0040] Please refer to Figure 2 , continuing with the above example, deploy the trained sub-control model to the robot, control the movement of the robot, and collect the execution trajectory sequence of the robot (i.e., the motion trajectory).

[0041] Step S140: Use the collected execution trajectory sequence to train the action increment model corresponding to the sub-control model, so as to adjust the corresponding sub-control model through the trained action increment model.

[0042] Please refer toFigure 2 , following the above example, for each sub-control model, the corresponding delta action model is further trained.

[0043] In a preferred embodiment of the present application, training the delta action model corresponding to the sub-control model may include: step S141, constructing a simulator; step S142, re-executing the corresponding motion in the simulator, and training the delta action model based on the collected execution trajectory sequence.

[0044] In an embodiment of the present application, the delta action model can be defined as:

[0045] wherein, represents the corrected action Δa t output according to the current state s t and the action a t . This corrected action is added to the collected real action to reduce the deviation between the simulation and the real dynamics.

[0046] Preferably, the dynamics of the simulator can be modified to:

[0047] wherein, f sim () is the simulator dynamics, and s t+1 represents the next state.

[0048] Further preferably, step S142 may include: initializing the first state collected based on the collected execution trajectory sequence; calculating a reward signal to minimize the difference between the simulated next state and the collected next state; and optimizing the delta action model using the PPO algorithm.

[0049] An embodiment of the present application can use reinforcement learning to train the delta action model, which may include: initializing the (first) state collected by the robot based on the real environment in the simulator; calculating a reward signal to minimize the difference between the simulated next state s and the next state of the real trajectory t+1 ; using the PPO algorithm to optimize the delta action learning to correct the Δa that can correct the difference between the simulation and the reality. t .

[0050] In a preferred embodiment of the present application, adjusting the corresponding sub-control model through the trained delta action model may include: constructing a new simulator; and in the new simulator, fixing the parameters of the delta action model and adjusting the corresponding sub-control model.

[0051] For the trained action increment model, the embodiments of the present application can reconstruct the emulator. And keep the parameters of the action increment model fixed during subsequent training. Fine-tune the pre-trained sub-control model in the new emulator. This stage can be iterated until the sub-control model converges.

[0052] In the embodiments of the present application, the simulation environment is, for example:

[0053] Step S150: Use the knowledge distillation technology to fuse the adjusted sub-control model to obtain an optimized full-body control model.

[0054] In a preferred embodiment of the present application, step S150 may include: step S151, take each adjusted control model as a teacher model, and construct a student model and a dynamic gating network. Step S152, use the dynamic gating network to generate the weight distribution of each teacher model. Step S153, synchronously perform the following operations within a preset time window: drive each teacher model and the student model to generate a predicted control sequence and a state trajectory; fuse the outputs of each teacher model according to the weight distribution to generate a reference control sequence and a reference state trajectory; and calculate the spatio-temporal joint loss between the output of the student model and the reference control sequence. Step S154, deploy the student model in the simulation environment to perform closed-loop control, calculate the performance loss through the collected trajectory data, and jointly optimize the student model with the spatio-temporal joint loss; step S155, take the optimized student model as the optimized full-body control model.

[0055] Continuing with the above example, take multiple pre-trained and independently adjusted sub-control models as teacher models; construct a lightweight student control model and a dynamic gating network. In the dynamic expert fusion stage: based on the current state, task objectives, and environmental context, etc., use the dynamic gating network to generate the real-time weight distribution of each teacher model. In the spatio-temporal awareness distillation stage: the following operations can be synchronously performed within a preset time window: 1) drive each teacher model and the student model to generate a predicted control sequence and a state trajectory; 2) fuse the outputs of the teacher models according to the real-time weight distribution to generate a reference control sequence and a reference state trajectory; 3) calculate the spatio-temporal joint loss between the output of the student model and the reference sequence, which may include control action deviation, state trajectory deviation, and hidden layer feature alignment deviation. In the closed-loop refinement stage: the student model can be deployed in the simulation environment to perform closed-loop control, collect actual trajectory data and calculate the performance loss, and jointly optimize the student model with the spatio-temporal joint loss. Take the optimized student model as the optimized full-body control model.

[0056] In the embodiments of the present application, the distillation loss function can be defined as: L distill =E {s～D} [KL(p general (a|s)||p expert,k(s)(a|s))]。

[0057] Among them, E {s～D} represents taking the expectation of the state s in the dataset D. KL is the Kullback-Leibler divergence. p expert,k(s) is the sub-control module corresponding to the state s. p general represents the basic whole-body control model. The action output by the a controller is used to control each joint of the robot.

[0058] Step S160: Use the optimized whole-body control model to perform whole-body control on the robot.

[0059] In the embodiment of the present application, the optimized whole-body control model is deployed in the control module of the robot for whole-body control of the robot.

[0060] Accordingly, the whole-body control method for a robot provided by the embodiment of the present application includes: obtaining the AMASS dataset and training the basic whole-body control model using the AMASS dataset; using a clustering algorithm to structurally partition the AMASS dataset to obtain multiple clustered sub-datasets, and using the multiple clustered sub-datasets to train sub-control models corresponding to each motion type on the basis of the basic whole-body control model; after deploying the trained sub-control models to the control module of the robot, collecting the execution trajectory sequence of the robot by controlling the robot's movement; using the collected execution trajectory sequence to train the action increment model corresponding to the sub-control model to adjust the corresponding sub-control model through the trained action increment model; using the knowledge distillation technology to fuse the adjusted sub-control models to obtain an optimized whole-body control model; and using the optimized whole-body control model to perform whole-body control on the robot. The embodiment of the present application provides a training framework of "general-expert-general" (i.e., "basic whole-body control model - multiple sub-control modules - optimized whole-body control model") for learning general whole-body control strategies, effectively alleviating the interference problem between different motion types; the embodiment of the present application also provides a clustering method based on an autoencoder (AE), which combines the action characteristics of each part and the action text description to structurally partition the dataset, so as to better train the expert whole-body control strategy. By clustering the action data, for each motion subset, the corresponding action increment model is trained separately. Since the same type of action has consistency in dynamic characteristics, this expertized training can significantly improve the fitting accuracy of the action increment model and more efficiently correct the difference between the simulation and the real environment. Compared with the general action increment model, the action increment model provided by the embodiment of the present application has significant advantages in action compensation accuracy, training convergence speed, and final control performance. In the general strategy training stage, the embodiment of the present application fuses the knowledge of each expert strategy through expert knowledge distillation to obtain a general whole-body control strategy that can handle various complex and agile motions.

[0061] Furthermore, the training process of the expert control strategy (multiple sub-control modules) provided by the embodiments of the present application has two stages. In the first stage, after training the basic full-body control model (i.e., the general full-body control strategy) on the AMASS dataset, in each clustered sub-dataset, fine-tuning is performed based on the general full-body control strategy to obtain an expert control strategy for a specific type of movement. In the second stage, each expert control strategy is deployed to the robot, and the corresponding real-world movement trajectories are collected; for each expert control strategy, the corresponding action increment model is further trained.

[0062] The embodiments of the present application also provide a control device, which includes: a memory, a processor, and a computer program stored on the memory and executable on the processor. The processor executes the computer program to implement the full-body control method according to the above.

[0063] The embodiments of the present application also provide a machine-readable storage medium, on which instructions are stored, and these instructions cause the machine to execute the above full-body control method.

[0064] The embodiments of the present application also provide a robot, which may include the above control device.

[0065] It should be noted that the above control device, machine-readable storage medium, and robot can implement the full-body control method provided by the above embodiments. The specific implementation manners can refer to the description of the full-body control method in the above embodiments and will not be elaborated here.

[0066] It can be understood that the circuit structures, names, and parameters described in the above embodiments are only examples. Those skilled in the art can also easily combine and adjust the structural features of the above multiple embodiments according to the usage needs, and should not limit the concept of the present application to the specific details of the above examples.

[0067] Although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A full-body control method for a robot, characterized in that, The described full-body control method includes: Obtain the AMASS dataset and use the AMASS dataset to train a basic full-body control model; Use a clustering algorithm to structurally partition the AMASS dataset to obtain multiple clustered sub-datasets, and use the multiple clustered sub-datasets to train sub-control models corresponding to each motion type based on the trained basic full-body control model; After deploying the trained sub-control models to the control module of the robot, collect the execution trajectory sequence of the robot by controlling the movement of the robot; Use the collected execution trajectory sequence to train the action increment model corresponding to the sub-control model, so as to adjust the corresponding sub-control model through the trained action increment model; Use knowledge distillation technology to fuse the adjusted sub-control models to obtain an optimized full-body control model; and Use the optimized full-body control model to perform full-body control on the robot.

2. The whole body control method according to claim 1, wherein, After obtaining the AMASS dataset, the full-body control method further includes: Obtain the HumanML3D dataset; and Based on motion features and action text descriptions, use an autoencoder and a self-attention mechanism network to encode the AMASS dataset and the HumanML3D dataset to complete the training of the basic full-body control model.

3. The whole body control method according to claim 2, characterized in that Encoding the AMASS dataset includes: Obtain the action sequence in SMPL format in the AMASS dataset; and Use the self-attention mechanism network structure to encode the action sequence, including converting joint axis angles and root coordinates to three-dimensional coordinates of joint points based on the SMPL format action sequence, deleting redundant joint points, and introducing the moving speed of the corresponding part relative to the world coordinate system.

4. The whole body control method according to claim 2, characterized in that Encoding the HumanML3D dataset includes: Use the BERT model to serialize the HumanML3D dataset to obtain a text sequence; and Use the self-attention mechanism network structure to encode the text sequence.

5. The whole body control method according to claim 1, characterized in that Training the action increment model corresponding to the sub-control model includes: Construct a simulator; and Re-execute the corresponding motion in the simulator and train the action increment model based on the collected execution trajectory sequence, including: Initialize the first state collected based on the collected execution trajectory sequence; Calculate a reward signal to minimize the difference between the simulated next state and the collected next state; and Use the PPO algorithm to optimize the action increment model.

6. The whole body control method according to claim 5, characterized in that Adjusting the corresponding sub-control model through the trained action increment model includes: Construct a new simulator; and In the new simulator, fix the parameters of the action increment model and adjust the corresponding sub-control model.

7. The whole body control method according to claim 1, wherein Using knowledge distillation technology to fuse the adjusted sub-control models to obtain an optimized full-body control model includes: Use each adjusted sub-control model as a teacher model, and construct a student model and a dynamic gating network; Use the dynamic gating network to generate the weight distribution of each teacher model; Synchronously perform the following operations within a preset time window: Drive each teacher model and the student model to generate a predictive control sequence and a state trajectory; Fuse the outputs of each teacher model according to the weight distribution to generate a reference control sequence and a reference state trajectory; and Calculate the spatio-temporal joint loss between the output of the student model and the reference control sequence; Deploy the student model in a simulation environment to perform closed-loop control, calculate the performance loss based on the collected trajectory data, and jointly optimize the student model with the spatio-temporal joint loss; and Use the optimized student model as the optimized whole-body control model.

8. A control device, characterized in that, The control device includes: a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor executes the computer program to implement the whole-body control method according to any one of claims 1-7.

9. A machine-readable storage medium, characterized in that, Instructions are stored on the machine-readable storage medium, and the instructions cause the machine to execute the whole-body control method according to any one of claims 1-7.

10. A robot, characterized in that, The robot includes the control device according to claim 8.

Images