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

Abstract

The application provides a full-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 full-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 full-body control model by using a plurality of clustering sub-data sets; after the trained sub-control model is deployed to a control module of the robot, an execution trajectory sequence of the robot is collected by controlling the motion of the robot; an action increment model corresponding to the sub-control model is trained by using the collected execution trajectory sequence, so as to adjust the corresponding sub-control model; and the adjusted sub-control model is fused by using a knowledge distillation technology, so as to obtain an optimized full-body control model. The embodiment of the application provides a training framework of “generalist-specialist-generalist”, which is used for learning of a general full-body control strategy and effectively alleviates the interference problem between different motion types.

Description

TECHNICAL FIELD

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

[0002] A humanoid robot has the morphological advantage of a human body structure and can perform various complex daily tasks such as grabbing, carrying, walking and crossing irregular terrains. However, the full-body control problem of a humanoid robot is still a major challenge in current robot control research. The main reasons are that the hardware of a humanoid robot is limited, the action space dimension is high, and the control coupling is complex, and so on, 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 precision of the control strategy.

[0003] Related technologies solve the above problems based on dynamics modeling and optimal control technology, but are difficult to adapt to changes in complex task scenarios. Reinforcement learning combined with sim-to-real technology shows good prospects in controlling complex full-body actions. However, such methods are usually difficult to balance the needs of different action types. SUMMARY

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

[0005] According to a first aspect of the application, a full-body control method for a robot is provided, the full-body control method comprising: acquiring an AMASS dataset, and training a basic full-body control model by using the AMASS dataset; structurally dividing the AMASS dataset by using a clustering algorithm to obtain a plurality of clustering sub-datasets, and training a sub-control model corresponding to each motion type on the basis of the trained basic full-body control model by using the plurality of clustering sub-datasets; after deploying the trained sub-control model to a control module of the robot, collecting an execution trajectory sequence of the robot by controlling the motion of the robot; training an action increment model corresponding to the sub-control model by using the collected execution trajectory sequence, so as to adjust the corresponding sub-control model by using the trained action increment model; fusing the adjusted sub-control model by using a knowledge distillation technology to obtain an optimized full-body control model; and controlling the full body of the robot by using the optimized full-body control model.

[0006] In an optional implementation, after the AMASS dataset is acquired, the whole body control method further comprises: acquiring a HumanML3D dataset; and based on motion features and action text descriptions, using a self-encoder and a self-attention mechanism network to encode the AMASS dataset and the HumanML3D dataset to complete training of the basic whole body control model.

[0007] In an optional implementation, encoding the AMASS dataset comprises: obtaining an SMPL format action sequence in the AMASS dataset; and using a self-attention mechanism network structure to encode the action sequence, including converting joint axis angles and root coordinates into three-dimensional coordinates of joint nodes, deleting redundant joint nodes, and introducing movement speed of corresponding parts relative to a world coordinate system based on the SMPL format action sequence.

[0008] In an optional implementation, encoding the HumanML3D dataset comprises: using a BERT model to serialize the HumanML3D dataset to obtain a text sequence; and using a self-attention mechanism network structure to encode the text sequence.

[0009] In an optional implementation, the training of the action increment model corresponding to the sub-control model comprises: constructing a simulator. In the simulator, the corresponding motion is re-executed, and based on the collected execution trajectory sequence, the action increment model is trained, including: based on the collected execution trajectory sequence, initializing the first state collected; calculating a reward signal to minimize the difference between the next state simulated and the next state collected; and using a PPO algorithm to optimize the action increment model.

[0010] In an optional implementation, the action increment model trained adjusts the corresponding sub-control model, comprising: 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 optional implementation, the fusion of the adjusted sub-control models to obtain the optimized whole-body control model comprises: taking each adjusted sub-control model as a teacher model, and constructing a student model and a dynamic gating network. The weight distribution of each teacher model is generated by using the dynamic gating network. Within a preset time window, the following operations are performed synchronously: driving each teacher model and the student model to generate a predicted control sequence and a state trajectory; fusing the outputs of each teacher model according to the weight distribution to generate a reference control sequence and a reference state trajectory; and calculating the spatiotemporal joint loss of the output of the student model and the reference control sequence. The student model is deployed in a simulation environment to perform closed-loop control, to calculate a performance loss by using collected trajectory data, and to iteratively optimize the student model in combination with the spatiotemporal joint loss; and the optimized student model is taken as the optimized whole-body control model.

[0012] According to a second aspect of the present application, a control device is provided, which comprises 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 described above.

[0013] According to a third aspect of the present application, a machine-readable storage medium is provided, which stores instructions for causing a machine to execute the whole-body control method described above.

[0014] According to a fourth aspect of the present application, a robot is provided, which comprises the control device described above.

[0015] By the technical solution, the full-body control method for the robot provided in the embodiment of the present application comprises: acquiring an AMASS dataset, and training a basic full-body control model by using the AMASS dataset; structurally dividing the AMASS dataset by using a clustering algorithm to obtain a plurality of clustering sub-datasets, and training a sub-control model corresponding to each motion type on the basis of the basic full-body control model by using the plurality of clustering sub-datasets; after deploying the trained sub-control model to a control module of the robot, collecting an execution trajectory sequence of the robot by controlling the robot to move; training an action increment model corresponding to the sub-control model by using the collected execution trajectory sequence, so as to adjust the corresponding sub-control model by using the trained action increment model; fusing the adjusted sub-control model by using a knowledge distillation technology to obtain an optimized full-body control model; and controlling the robot by using the optimized full-body control model. The embodiment of the present application provides a training framework of “generalist-expert-generalist” (i.e., “basic full-body control model-multiple sub-control modules-optimized full-body control model”), which is used for learning of a general full-body control strategy, and effectively alleviates the interference problem between different motion types. The embodiment of the present application further provides a clustering method based on an autoencoder (AE), which combines action features of each part and action text descriptions to structurally divide the dataset, so as to better train an expert full-body control strategy. By clustering the action data, the corresponding action increment model is trained for each motion subset. Since the same type of action has consistency in the dynamic characteristics, such expert training can significantly improve the fitting accuracy of the action increment model, and more efficiently correct the difference between simulation and real environment. Compared with a general action increment model, the action increment model provided in 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 by expert knowledge distillation to obtain a general full-body control strategy capable of processing various complex and agile motions.

[0016] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art from the following description, or can be learned by practice of the present application. The objects and other advantages of the present application can be realized and attained by the structure and process particularly pointed out in the description and the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the following will briefly introduce the drawings needed to be used in the embodiment or related art description. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without any creative effort.

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

[0019] Figure 2 is a flowchart of a full-body control method provided by an example embodiment of the present application. DETAILED DESCRIPTION

[0020] In order 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 described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0021] As described above, reinforcement learning combined with sim-to-real technology shows good prospects in controlling complex full-body movements. However, such methods are usually difficult to balance the needs of different types of movements. Different types of movements have significant differences in control focus. For example, jumping movements pay more attention to high-torque control, and smooth walking pays more attention to balance and continuity. Direct training of a unified strategy is likely to cause gradient conflicts between tasks, resulting in performance degradation.

[0022] To this end, an embodiment of the present application provides a full-body control method for a robot, which optimizes from two directions of model structure and data structure to provide a general full-body control strategy for humanoid robots facing various agile movements.

[0023] Figure 1 is a flowchart of a full-body control method for a robot provided by an example embodiment of the present application, which can include the following steps:

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

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

[0026] In the preferred embodiments of the present application, after obtaining the AMASS dataset, the full-body control method can further include: obtaining a HumanML3D dataset; and based on the motion features and action text descriptions, using a self-encoder and a self-attention mechanism network to encode the AMASS dataset and the HumanML3D dataset.

[0027] Among them, in addition to the action sequences expressed in SMPL format in 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 AMASS dataset.

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

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

[0030] For reference Figure 2 For example, based on the motion features of the leg joint nodes and the action text descriptions, the embodiments of the present application use a self-encoder (AE) and a self-attention mechanism network to encode the AMASS dataset and the HumanML3D dataset. For example, using a self-attention mechanism network structure to encode the action, for example, converting joint axis angles and root coordinates into three-dimensional coordinates of the joint nodes based on forward dynamics, deleting redundant joint nodes, and introducing the moving speed of the foot relative to the world coordinate system. For text data, serialization is performed through a BERT model, and is also processed through a self-attention mechanism network structure to obtain a hidden space representation corresponding to the action modality.

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

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

[0033] As described above, unlike the method of directly using the autoencoder to reconstruct the complete trajectory, the embodiment of the present application introduces text annotation, thereby aligning the action at the semantic level to ensure that actions with different styles but consistent semantics are closer in the hidden space. For example, the walking action can be straight forward or walking in a circle. If only relying on the autoencoder training, it is difficult to reflect the similarity of the two, and through text semantic alignment, this problem can be effectively overcome. The embodiment of the present application effectively improves the aggregation of semantically similar actions in the hidden space by introducing text alignment. Further, the embodiment of the present application does not directly use the action sequence expressed in the SMPL format for reconstruction. Since SMPL mainly represents joint angles and base coordinate transformation, this representation method is difficult to intuitively reflect the action features of interest. Therefore, the embodiment of the present application converts the joint axis angle and root coordinates into three-dimensional coordinates of the joint node based on the forward dynamics; deletes redundant joint nodes, and introduces the movement 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 part of the key joint nodes, foot speed, root node speed and rotation information, thereby focusing on the core action features.

[0034] Step S120: structurally divide the AMASS dataset by using a clustering algorithm to obtain a plurality of clustered sub-datasets, and train a sub-control model corresponding to each motion type based on the fully trained whole-body control model and the plurality of clustered sub-datasets.

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

[0036] For example, the encoded AMASS dataset is divided into multiple cluster sub-datasets using a clustering algorithm; in each cluster sub-dataset, a fine-tuning is performed based on the basic whole-body control model to obtain a sub-control model corresponding to each motion type, and each sub-control model can be a specialized control strategy for a specific motion type. Figure 2

[0037] Similarly, the embodiments of the present application do not limit the construction and training process of the sub-control model (network), and the model (network) for controlling each motion type of the humanoid robot belongs to the scope of protection of the present application.

[0038] The basic whole-body control model has basic motion tracking and execution capabilities; in the embodiments of the present application, the basic whole-body control model is used as an initialization point, and a fine-tuning is performed on each cluster sub-dataset to train multiple sub-control models to adapt to specific motion patterns and control preferences.

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

[0040] For example, the trained sub-control model is deployed to the robot, the motion of the robot is controlled, and the execution trajectory sequence (i.e., the motion trajectory) of the robot is collected. Figure 2

[0041] Step S140: The action increment model corresponding to the sub-control model is trained using the collected execution trajectory sequence to adjust the corresponding sub-control model through the trained action increment model.

[0042] For example, the trained sub-control model is deployed to the robot, the motion of the robot is controlled, and the execution trajectory sequence (i.e., the motion trajectory) of the robot is collected.​​Figure 2 Further to the above example, for each sub-control model, a corresponding delta action model is further trained.

[0043] In the preferred embodiment of the present application, training the delta action model corresponding to the sub-control model can include: step S141, constructing a simulator; and 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 the embodiment of the present application, the delta action model can be defined as:

[0045] wherein, represents the correction action Δa t outputted according to the current state s t and the action a t . The positive 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 as:

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

[0048] Further preferably, step S142 can include: initializing the collected first state 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] The embodiment of the present application can train the delta action model using reinforcement learning, which can include: initializing the robot in the simulator based on the (first) state collected by the real environment; calculating a reward signal to minimize the difference between the simulated next state s t+1 and the real trajectory next state ; and optimizing the delta action using the PPO algorithm to learn Δa t that can correct the simulation and real difference.

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

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

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

[0053] Step S150: using the knowledge distillation technology, fusing the adjusted sub-control model to obtain the optimized whole-body control model.

[0054] In the preferred embodiment of the present application, step S150 can include: step S151, taking each adjusted control model as a teacher model, and constructing a student model and a dynamic gating network. Step S152, using the dynamic gating network to generate the weight distribution of each teacher model. Step S153, synchronously performing the following operations within a preset time window: driving each teacher model and student model to generate a predicted control sequence and a state trajectory; fusing the outputs of each teacher model according to the weight distribution to generate a reference control sequence and a reference state trajectory; and calculating the spatio-temporal joint loss of the output of the student model and the reference control sequence. Step S154, deploying the student model in the simulation environment to perform closed-loop control, calculating the performance loss through the collected trajectory data, and iteratively optimizing the student model jointly with the spatio-temporal joint loss; step S155, taking the optimized student model as the optimized whole-body control model.

[0055] Taking the above example, multiple pre-trained and independently adjusted sub-control models are taken as teacher models; a lightweight student control model and a dynamic gating network are constructed. In the dynamic expert fusion stage: based on the current state, task target and environment context, etc., the dynamic gating network is used to generate the real-time weight distribution of each teacher model. In the spatio-temporal perception distillation stage: the following operations can be synchronously performed within a preset time window: 1) driving each teacher model and student model to generate a predicted control sequence and a state trajectory; 2) fusing 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) calculating the spatio-temporal joint loss of the output of the student model and the reference sequence, which can include control action deviation, state trajectory deviation and hidden layer feature alignment deviation. In the closed-loop refining stage: the student model can be deployed in the simulation environment to perform closed-loop control, collect actual trajectory data and calculate performance loss, and iteratively optimize the student model jointly with the spatio-temporal joint loss. The optimized student model is taken as the optimized whole-body control model.

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

[0057] where E {s～D} denotes taking expectation over states s in dataset D, KL is the Kullback-Leibler divergence, p expert,k(s) is the sub-control module corresponding to state s, p general denotes the base whole-body control model, a is the action output by the controller, which is used to control the joints of the robot.

[0058] Step S160: controlling the robot by using the optimized whole-body control model.

[0059] In the embodiments of the present application, the optimized whole-body control model is deployed in the control module of the robot, and is used to control the robot.

[0060] Accordingly, the whole-body control method for the robot provided in the embodiments of the present application comprises: acquiring an AMASS dataset, and training a base whole-body control model by using the AMASS dataset; structurally dividing the AMASS dataset by using a clustering algorithm to obtain a plurality of clustering sub-datasets, and training a sub-control model corresponding to each motion type on the basis of the base whole-body control model by using the plurality of clustering sub-datasets; after the trained sub-control model is deployed to the control module of the robot, collecting an execution trajectory sequence of the robot by controlling the motion of the robot; training an action increment model corresponding to the sub-control model by using the collected execution trajectory sequence, so as to adjust the corresponding sub-control model by using the trained action increment model; fusing the adjusted sub-control model by using a knowledge distillation technology, so as to obtain an optimized whole-body control model; and controlling the robot by using the optimized whole-body control model. The embodiments of the present application provide a “generalist-specialist-generalist” (i.e., “base whole-body control model-multiple sub-control modules-optimized whole-body control model”) training framework, which is used for learning of a general whole-body control strategy, and effectively alleviates 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, so as to structurally divide the dataset, thereby better training a specialist whole-body control strategy. By clustering the action data, an action increment model corresponding to each motion subset is trained respectively. Since the same type of action has consistency in the kinetic characteristics, such specialist training can significantly improve the fitting accuracy of the action increment model, and more efficiently correct the difference between simulation and real environment. Compared with a general action increment model, the action increment model provided in the embodiments 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 embodiments of the present application fuse the knowledge of each specialist strategy by specialist knowledge distillation, so as to obtain a general whole-body control strategy which can handle multiple complex and agile motions.

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

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

[0063] The embodiments of the present application also provide a machine-readable storage medium, which stores instructions for causing a machine to execute the whole-body control method described above.

[0064] The embodiments of the present application also provide a robot, which can comprise the control device described above.

[0065] It should be noted that the control device, the machine-readable storage medium, and the robot described above can implement the whole-body control method provided by the embodiments described above, and the specific implementation manner can be referred to the description of the whole-body control method in the embodiments described above, which will not be described here.

[0066] It can be understood that the circuit structure, name and parameter described in the above embodiments are only examples. Those skilled in the art can also easily think of combinations and adjustments of the structural features of the above multiple embodiments according to the use needs, and the concept of the present application should not be limited 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 skilled in the art should understand that they can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A whole body control method for a robot, characterized by, The whole body control method comprises: obtaining an AMASS dataset and training a basic whole body control model using the AMASS dataset; structuring and dividing the AMASS dataset using a clustering algorithm to obtain a plurality of clustering sub-datasets, and training a sub-control model corresponding to each motion type based on the trained basic whole body control model using the plurality of clustering sub-datasets; after deploying the trained sub-control model to the control module of the robot, collecting an execution trajectory sequence of the robot by controlling the motion 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; fusing the adjusted sub-control model using a knowledge distillation technology to obtain an optimized whole body control model; and controlling the robot using the optimized whole body control model, the fusing the adjusted sub-control model using a knowledge distillation technology to obtain an optimized whole body control model comprises: taking each adjusted sub-control model as a teacher model, and constructing a student model and a dynamic gating network; generating a weight distribution of each teacher model using the dynamic gating network; synchronously performing the following operations within a preset time window: driving each teacher model and the student model to generate a predicted control sequence and a state trajectory; fusing the outputs of each teacher model according to the weight distribution to generate a reference control sequence and a reference state trajectory; and calculating the spatiotemporal joint loss of the output of the student model and the reference control sequence; deploying the student model in a simulation environment to perform closed-loop control, calculating the performance loss through the collected trajectory data, and iteratively optimizing the student model in combination with the spatiotemporal joint loss; and taking the optimized student model as the optimized whole body control model.

2. The whole-body control method according to claim 1, characterized by, After obtaining the AMASS dataset, the whole body control method further comprises: obtaining a HumanML3D dataset; and based on the motion features and action text description, encoding the AMASS dataset and the HumanML3D dataset using a self-encoder and a self-attention mechanism network to complete the training of the basic whole body control model.

3. The whole-body control method according to claim 2, characterized by, Encoding the AMASS dataset comprises: obtaining an action sequence in SMPL format 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 nodes, deleting redundant joint nodes, and introducing movement velocities of corresponding parts relative to the world coordinate system based on the action sequence in SMPL format.

4. The whole-body control method according to claim 2, characterized by, Encoding the HumanML3D dataset comprises: 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.

5. The global control method according to claim 1, characterized in that, The training of the action increment model corresponding to the sub-control model comprises: constructing a simulator; and re-perform the corresponding motion in the simulator, and train the action increment model based on the collected execution trajectory sequence, including: initializing the collected first state 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 a PPO algorithm.

6. The global control method according to claim 5, characterized in that, the trained action increment model adjusts the corresponding sub-control model, including: constructing a new simulator; and in the new simulator, fixing the parameters of the action increment model and adjusting the corresponding sub-control model.

7. A control device characterized by comprising: 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 1-6.

8. A machine-readable storage medium, characterized in that, The machine readable storage medium has instructions stored thereon, which cause the machine to perform the whole body control method according to any one of claims 1-6.

9. A robot, characterized in that The robot includes the control device of claim 7.

Images