A method for controlling the motion of a humanoid robot with combined text and trajectory constraints

Abstract

This application discloses a humanoid robot motion control method with joint constraints of text and trajectory, belonging to the field of humanoid robot control technology. The method first obtains the text description and motion trajectory corresponding to the preset control action. Through motion synthesis with joint constraints of text and trajectory, a SMPL digital human pose sequence conforming to dual constraints is generated. An optimization strategy combining stochastic gradient descent and momentum methods is used to complete motion redirection, obtaining pose data adapted to the target humanoid robot. A teacher-student model is constructed based on a proximal policy optimization algorithm, using the redirected pose data as prior guidance. Policy distillation is performed through a dataset aggregation algorithm to train the motion control model. Finally, the model is deployed to the robot, and motion control is achieved using the joint target angles output by the model as driving signals. This application provides high-quality prior guidance through dual constraints, effectively improving the semantic consistency, trajectory accuracy, anthropomorphism, and scene robustness of humanoid robot motion, significantly reducing training costs, and possessing excellent engineering application value.

Description

Technical Field

[0001] This application relates to the field of humanoid robot control technology, and in particular to a humanoid robot motion control method with joint constraints of text and trajectory. Background Technology

[0002] 3D motion synthesis, as an important extension of Artificial Intelligence Generated 3D Content (AI-Generated 3D Content) into the field of motion capture, aims to automatically generate 3D human pose sequences that meet preset constraints using AI algorithms. Currently, the application scenarios of motion synthesis technology are mostly concentrated in the field of disembodied artificial intelligence, such as virtual anchor motion generation, film and game group animation production, and AI choreography. Related technical solutions primarily focus on optimizing the visual effects of virtual environments, without fully considering the joint constraints and kinematic differences of humanoid robots in the physical world.

[0003] Currently, the mainstream methods for acquiring training data for humanoid robot motion control include two categories: motion capture and teleoperation. Among them, the RHINO framework constructs a learning architecture by collecting human-object interaction data and teleoperation data, enabling the robot to recognize human intentions and respond to tasks in real time; the ExBody human body control technology utilizes the large-scale human motion capture dataset AMASS (Archive of Motion Capture as Surface Shapes) to provide data support for robot motion control; the Optimus humanoid robot project uses an inertial motion capture system to collect joint motion data of humans performing industrial tasks, and then trains the robot's autonomous operation capabilities after motion redirection; other research uses mixed reality teleoperation assistance frameworks to collect human motion signals and construct enhanced teleoperation modes.

[0004] While motion capture and teleoperation can acquire high-quality human motion data, these methods suffer from drawbacks such as high hardware costs, operational complexity, susceptibility to deployment environment limitations, and limited pose space. In contrast, motion synthesis technology offers advantages such as low cost, high efficiency, independence from external environment and hardware constraints, and a large pose space range, making it a promising supplementary or alternative technology to motion capture in the field of humanoid robot motion control.

[0005] However, the application of motion synthesis technology in the field of humanoid robot motion control still faces key technical bottlenecks. On the one hand, humanoid robots use rotation angles to represent pose, while mainstream motion synthesis technologies use SMPL or its extended version SMPL-X parametric models to represent human pose. The kinematic structures of these two methods differ significantly, making it difficult to directly adapt the generated human pose data to the joint constraints and kinematic structure of humanoid robots. On the other hand, many existing motion synthesis methods lack the precision and controllability of the joint trajectories generated for human motion. For example, using textual language as constraint signals can only macroscopically control the type of action, failing to achieve fine-grained path control; using scenes as constraint signals requires integrating complex factors such as 3D mapping, global path planning, and local obstacle avoidance, making modeling extremely difficult. Summary of the Invention

[0006] The purpose of this application is to provide a humanoid robot motion control method that combines text and trajectory constraints, enabling anthropomorphic motion control that achieves both semantic consistency and trajectory accuracy.

[0007] To achieve the above objectives, this application provides the following solution: A method for motion control of a humanoid robot with joint constraints of text and trajectory includes the following steps: The text descriptions and motion trajectories corresponding to several preset control actions are obtained. Then, a motion synthesis method that combines text and trajectory constraints is used to generate an SMPL digital human pose sequence that simultaneously conforms to the text descriptions and motion trajectories.

[0008] An optimization strategy combining stochastic gradient descent and momentum methods is adopted to perform motion retargeting on the SMPL digital human pose sequence, thereby obtaining robot pose data corresponding to each preset control action.

[0009] A teacher-student model is constructed based on a proximal policy optimization algorithm. A dataset aggregation algorithm is used to distill the motion skills of the teacher model into the student model. The robot pose data corresponding to each preset control action obtained after motion retargeting is used as prior guidance to train a humanoid robot motion control model.

[0010] The text description and motion trajectory corresponding to the current desired control action are obtained. Then, the motion synthesis method and motion relocation method based on the joint constraints of text and trajectory are processed sequentially to obtain the robot pose corresponding to the current desired control action.

[0011] The robot pose corresponding to the current desired control action is input into the humanoid robot motion control model. The joint target angle position output by the humanoid robot motion control model is used as the driving signal to control the movement of the target humanoid robot. The humanoid robot motion control model is deployed to the target humanoid robot.

[0012] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a humanoid robot motion control method based on joint text and trajectory constraints. First, the text description and motion trajectory corresponding to preset control actions are obtained. Then, a motion synthesis method based on joint text and trajectory constraints is used to macroscopically constrain the action type using text and microscopically lock the movement path using trajectory constraints. This provides high-quality action priors with high semantic matching and excellent trajectory accuracy for subsequent control, avoiding the problem of insufficient accuracy in single-constraint control. Subsequently, an optimization strategy combining stochastic gradient descent and momentum methods is employed to perform motion retargeting on the SMPL digital human pose sequence, obtaining robot pose data corresponding to each preset control action. This effectively eliminates the kinematic structural differences between the virtual digital human and the physical robot, accelerates optimization convergence, reduces the risk of joint over-limit and jitter, and ensures the feasibility and control accuracy of pose transfer, solving the core pain point that virtual motion data cannot directly drive the robot. Next, a teacher-student model is constructed based on a proximal policy optimization algorithm, and a dataset aggregation algorithm is used. The motion skills of the teacher model are distilled into the student model, and the robot pose data obtained by retargeting is used as prior guidance to train a humanoid robot motion control model. This overcomes the limitations of sparse rewards and easy pattern collapse in pure reinforcement learning. The high-precision and stable motion capability in full-information scenarios is transferred to a general model that only relies on ontology perception, which greatly improves the model's generalization ability, scene anti-interference ability and the degree of anthropomorphism of motion. Finally, the trained motion control model is deployed to the target humanoid robot. The text description and motion trajectory corresponding to the current expected control action are obtained. After the aforementioned motion synthesis and motion retargeting processing, the corresponding robot pose is obtained and input into the motion control model. The joint target angle position output by the model is used as the driving signal to control the motion of the target humanoid robot. Without relying on high-cost motion capture or teleoperation equipment, anthropomorphic motion control with both semantic consistency and trajectory accuracy can be achieved. It can be flexibly adapted to various task scenarios and has strong engineering practicality. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart illustrating a humanoid robot motion control method based on text and trajectory joint constraints, as provided in one embodiment of this application.

[0015] Figure 2This is a schematic diagram comparing the motion synthesis results with and without trajectory constraints in a humanoid robot motion control method with joint text and trajectory constraints provided in an embodiment of this application.

[0016] Figure 3 This is a schematic diagram of the mapping of 15 key joints from an SMPL digital human to a humanoid robot in a text and trajectory joint constraint humanoid robot motion control method provided in an embodiment of this application.

[0017] Figure 4 This is a schematic diagram illustrating the motion redirection process from an SMPL digital human to a humanoid robot in a text and trajectory joint constraint humanoid robot motion control method provided in an embodiment of this application.

[0018] Figure 5 This is a flowchart illustrating the task formalization and state-action-reward mapping based on Markov decision processes in a humanoid robot motion control method with joint constraints of text and trajectory provided in an embodiment of this application.

[0019] Figure 6 This is a flowchart illustrating the teacher policy training based on the PPO algorithm in a humanoid robot motion control method with joint constraints of text and trajectory provided in an embodiment of this application.

[0020] Figure 7 This is a schematic diagram illustrating the network selection and initialization process of a student strategy in a humanoid robot motion control method with joint constraints of text and trajectory provided in an embodiment of this application.

[0021] Figure 8 This is a schematic diagram illustrating the design process of the pattern mask and sparse mask in a humanoid robot motion control method with joint constraints of text and trajectory provided in an embodiment of this application.

[0022] Figure 9 This is a schematic diagram illustrating the process of student policy training based on the Dagger algorithm in a humanoid robot motion control method with joint constraints of text and trajectory provided in an embodiment of this application.

[0023] Figure 10 This is a schematic diagram illustrating the operation of a humanoid robot motion control method with combined text and trajectory constraints provided in an embodiment of this application, under different training steps.

[0024] Figure 11 This is a schematic diagram illustrating the training and evaluation data statistics in a humanoid robot motion control method with joint constraints of text and trajectory provided in an embodiment of this application.

[0025] Figure 12This diagram illustrates how the values ​​of different evaluation metrics change with the number of training iterations of the motion controller in a humanoid robot motion control method that combines text and trajectory constraints, as provided in an embodiment of this application.

[0026] Figure 13 This is a schematic diagram illustrating the process and brief experimental results of a humanoid robot motion control method based on text and trajectory joint constraints provided in an embodiment of this application, which implements motion training based on specified text and trajectory. Detailed Implementation

[0027] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0028] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] This application provides a method for motion control of a humanoid robot with joint constraints of text and trajectory. In one exemplary embodiment, such as... Figure 1 As shown, it includes the following steps: S1. Obtain the text descriptions and motion trajectories corresponding to several preset control actions, and generate an SMPL digital human pose sequence that simultaneously conforms to the text descriptions and motion trajectories through a motion synthesis method that combines text and trajectory constraints.

[0030] In an exemplary embodiment, when generating the SMPL digital human pose sequence in step S1, motion type semantic features are extracted from the text description, and a movement path based on three-dimensional spatial coordinates is extracted from the motion trajectory. The generated SMPL digital human pose sequence simultaneously satisfies the constraints of motion type semantic features and movement path. Motion type semantic features define what actions the target humanoid robot should take, and movement path defines the path along which the target humanoid robot should take actions.

[0031] S2. An optimization strategy combining stochastic gradient descent and momentum methods is adopted to perform motion redirection processing on the SMPL digital human pose sequence to obtain robot pose data corresponding to each preset control action.

[0032] As an optional implementation, the motion redirection process in step S2 includes four sub-steps in sequence: shape parameter adaptation, joint mapping, optimization solution, and safety filtering; shape parameter adaptation is: optimizing the shape parameters of the SMPL digital human. βThe limb proportions of the SMPL digital human structure are matched with those of the target humanoid robot. The joint mapping is to select several key joints to establish a one-to-one correspondence between the SMPL digital human and the target humanoid robot. The optimization solution is to construct a function with the objective of minimizing the difference in joint positions between the SMPL digital human and the target humanoid robot. The safety filtering is to remove motion sequences that have joint over-limits or abnormal ground contact.

[0033] S3. Construct a teacher-student model based on the proximal policy optimization algorithm, and use a dataset aggregation algorithm to distill the motion skills of the teacher model into the student model. Using the robot pose data corresponding to each preset control action obtained after motion redirection as prior guidance, train the humanoid robot motion control model. In this embodiment, step S3 specifically includes the following steps: S31. Construct a feasible motion dataset using the robot pose data corresponding to each preset control action obtained after motion redirection.

[0034] S32. Construct a teacher-student model architecture based on the proximal policy optimization algorithm, and train the teacher model using a feasible motion dataset as the training basis.

[0035] Specifically, the ontological perception state of the teacher model's input data includes the rigid body position, posture, linear velocity, and angular velocity of the target humanoid robot over several consecutive historical steps. The target state consists of the robot's reference pose data corresponding to the preset control action obtained after motion retargeting and the difference between the current pose data and the reference pose data. A proximal policy optimization algorithm is used to optimize a multi-dimensional composite reward function that includes penalty terms, regularization terms, and task reward terms, thereby learning a high-precision motion control strategy that meets the preset control action pose requirements. The output data of the teacher model serves as the reference motion pose for the student model to learn.

[0036] The input to the student model includes proprioceptive data and the target state processed by double masking. The proprioceptive data includes the joint positions, joint linear velocities, base angular velocities, and gravity vectors of the target humanoid robot's historical action sequence with a preset number of steps. The target state is processed by pattern masking and sparse masking to form multi-mode control commands. When performing policy distillation using a dataset aggregation algorithm, the student model generates motion trajectories in the simulation environment, and the optimal reference action at the corresponding moment is obtained through the teacher model. The L2 norm loss of the student model's output action and the teacher model's output action is used as the optimization objective. The transfer of motor skills to the student model is completed through supervised learning.

[0037] The student model uses the same 3-layer fully connected neural network structure as the teacher model, with hidden layer dimensions of 512, 256, and 128 respectively. The hidden layers use the ReLU activation function, and the output layer uses the Tanh activation function. To further reduce the computational load on the humanoid robot, the student model's network structure can be simplified based on the teacher model's network structure, reducing the number of model parameters and computational overhead while ensuring core performance.

[0038] S33. The motion control strategy of the teacher model is distilled into the student model using a dataset aggregation algorithm to train the humanoid robot motion control model. In this embodiment, step S33 specifically includes the following steps: S331. Initialize the preset total number of rounds, maximum time step per round, and policy update interval for distillation training. Load the pre-trained teacher model and the initialized student model, and establish a global supervised experience cache.

[0039] S332. Generate a combination of pattern mask and sparse mask, based on the reference motion pose output by the teacher model, and obtain the target state data of the student model for the corresponding training round by processing the pattern mask and sparse mask.

[0040] S333. In each time step of a single round, the observation vector of the student model is constructed based on the real-time ontology perception data of the target humanoid robot at the corresponding time step and the target state at the corresponding time step.

[0041] S334. Input the observation vector into the student model to output the target angle action of the corresponding controllable joint, and input the action into the simulation environment to drive the humanoid robot to perform the motion.

[0042] S335. Simultaneously obtain the optimal reference action for the corresponding time step through the teacher model, and encapsulate the observation vector, target state and optimal reference action as a single sample and store them in a temporary cache.

[0043] S336. After a single round of training, all single samples in the temporary cache are merged into the global supervised experience cache.

[0044] S337. When the number of samples in the global supervised experience cache reaches the preset update threshold or the preset number of training batches are completed, the network parameters of the student model are updated through supervised learning, with the L2 norm loss of the student model's output action and the teacher model's output action as the optimization objective.

[0045] Determine whether the iteration termination condition has been met. If so, output the trained student model as the humanoid robot motion control model; otherwise, jump to step S332 until the iteration termination condition is met. As an optional implementation, the iteration termination condition is that a preset total number of rounds has been completed or the student model has converged and can stably reproduce the motion trajectory corresponding to the preset control action.

[0046] S4. Obtain the text description and motion trajectory corresponding to the current desired control action, and process them sequentially through the motion synthesis method and motion relocation method based on the joint constraints of text and trajectory to obtain the robot pose corresponding to the current desired control action.

[0047] S5. Input the robot pose corresponding to the desired control action into the humanoid robot motion control model. Use the joint target angle position output by the humanoid robot motion control model as the driving signal to control the movement of the target humanoid robot. The humanoid robot motion control model is deployed to the target humanoid robot. The humanoid robot motion control model supports pose tracking mode, joint angle tracking mode, and root motion stabilization control mode. It adapts to different task scenarios by dynamically configuring mode masks and sparse masks.

[0048] The humanoid robot motion control method based on text and trajectory joint constraints provided in the above embodiments of this application can be divided into four parts: SMPL digital human pose motion synthesis with text and trajectory joint constraints → motion redirection from SMPL digital human pose to humanoid robot pose → training of humanoid robot motion control model based on PPO algorithm and teacher-student model → Sim-to-Sim deployment and verification. Supplementary explanations are provided below for each part: For the first part, "SMPL Digital Human Pose Motion Synthesis with Joint Text and Trajectory Constraints," this step requires considering which text and trajectory joint constraint motion synthesis method to use, based on the common basic movements (such as walking, turning, and linear movement) and movement trajectories that the humanoid robot needs to be trained on. Existing text and trajectory joint constraint motion synthesis methods are generally based on supervised learning. The generalization ability of these methods highly depends on the size and annotation accuracy of the training dataset. KIT-ML and HumanML3D are currently the most widely used public datasets for motion synthesis. KIT-ML contains 3911 motion sequences and 6278 fine-grained text descriptions, covering basic daily movements; HumanML3D contains 14616 motion sequence fragments and 44970 annotations, covering scenarios such as daily activities and sports movements. The motion poses of both datasets use SMPL digital human pose representation, which can meet the model training requirements of text and trajectory joint constraint motion synthesis methods.

[0049] As a benchmark, this application first explores the feasibility of applying the scene-constrained motion synthesis method CHOIS (Controllable Human-Object Interaction Synthesis) to humanoid robot control. Experiments show that CHOIS cannot accurately constrain the pose's movement path, and is prone to collisions and clipping between the generated digital human limbs and surrounding objects. Figure 2 The unconstrained scene shown in part (a) is the object-picking SMPL digital human pose sequence generated by CHOIS, where the digital human's limbs can be seen penetrating the table and cabinet.

[0050] Although current motion synthesis methods are still predominantly driven by single-modal signals, a number of methods have adopted multi-modal signal-driven approaches. Guided Motion Diffusion (GMD) and TLControl, both using text and trajectory joint constraints, are representative of this type. GMD's overall framework is based on a diffusion model, embedding text semantic encoding and trajectory geometric features as dual-guided conditions into the diffusion sampling process. It uses a scene occupancy mesh modeling algorithm to detect the distance between the limb and the surrounding environment in real time, suppressing clipping behavior. TLControl constructs a trajectory and text collaborative network, designing a "coarse prediction of the full joint trajectory - fine optimization of trajectory details" process to achieve deep alignment of the two different modal constraints. While both GMD and TLControl can generate poses represented by SMPL for subsequent adaptation to motion retargeting steps, TLControl's generated SMPL digital human poses significantly outperform GMD in trajectory accuracy (as shown in Table 1, where arrows indicate the direction of index optimization). Therefore, this embodiment selects TLControl as the motion synthesis method. Figure 2 As shown in part (b), after introducing trajectory constraints, the trajectory limits the movement path of the SMPL digital human from the initial position to the cabinet, which not only completes the object retrieval action, but also avoids collisions and clipping with the table and cabinet.

[0051] Table 1 Evaluation results of GMD and TLControl on the two datasets.

[0052] For the second part, "Motion Redirection from SMPL Digital Human Pose to Humanoid Robot Pose", the core of this step is to adapt the SMPL digital human pose sequence generated by the joint constraints of text and trajectory to the hardware structure of the target humanoid robot through three steps: "offline calibration - online migration - simulation verification". This solves the differences in kinematics and geometry between the SMPL digital human and the humanoid robot, and outputs a feasible pose that conforms to the robot's hardware constraints.

[0053] This embodiment uses Yushu H1 as an example, and the motion redirection process is as follows: Figure 3 As shown. SMPL Digital Human ( Figure 3 (b) is based on the human skeletal tree. Figure 3 The design of (a) section matches the joint freedom and limb proportions of the real human body, including multiple joints such as the head, spine, and limbs. Figure 3 The text indicates that 15 key joints involved in the motion redirection process in this embodiment are marked; while the humanoid robot ( Figure 3 Due to hardware limitations, the number of joints, degrees of freedom (e.g., the shoulder joint of the Yushu H1 has 2 degrees of freedom), and range of motion (e.g., hip joint rotation ±90°) of the digital human are significantly different from those of a digital human. If the pose of a digital human is directly adopted, problems such as joint over-limit and motion instability can easily occur. For example, if the large-angle arm raising movement of a digital human exceeds the range of freedom of the shoulder joint of a humanoid robot, it will cause mechanical structure jamming.

[0054] This application employs an improved motion redirection method, mainly divided into three stages: an offline stage to unify the morphological benchmarks of the digital human and robot, an online stage to optimize joint angles to adapt to hardware constraints, and a simulation stage to verify the physical feasibility of the motion. Figure 4 As shown, the detailed process is as follows: 1) Offline Phase (Morphological and Model Calibration): ... Figure 3 Based on the digital human skeleton tree in Part (b), scaling parameters were adjusted to make the limb length and joint spacing of the digital human more consistent with the mechanical structure of the Yushu H1 humanoid robot, reducing the fundamental geometric differences. The digital human pose was mapped to the humanoid robot pose based on a custom robot standing posture. The optimal shape parameters were obtained by using a combination of stochastic gradient descent (SGD) and momentum optimization algorithms. β With the morphologically calibrated SMPL digital human parameter file, offline benchmark establishment was completed. The optimization algorithm used in this application has advantages over the Adam optimization algorithm used in traditional motion retargeting methods. The performance comparison of the two optimization algorithms is shown in Table 2.

[0055] Table 2 Performance comparison of Adam algorithm and SGD+Momentum algorithm for shape optimization

[0056] 2) Online Subprocess (Motion Redirection): Input text and trajectory into TLControl to obtain the SMPL digitizer pose sequence. Map the global position and orientation of the digitizer's root node (pelvis) to the coordinates and orientation of the humanoid robot's root node, ensuring consistent spatial reference orientation for motion. Based on Figure 3The key joint mapping relationships shown are analyzed. Gradient descent algorithms are used to optimize joint angles, ensuring that the limb motion trajectory of the SMPL matches the robot's hardware structure. For example, the 3-DOF shoulder joint motion of the SMPL is converted into a feasible 2-DOF shoulder joint pose for the robot. Angles exceeding the robot's joint range of motion are filtered to ensure the pose conforms to the robot's hardware constraints. The robot's foot joint angles are adjusted to ensure stable contact between the feet and the ground during walking, preventing imbalance. Finally, the optimal shape parameters and joint angle data for motion retargeting are summarized.

[0057] 3) Simulation phase (feasibility verification): Load the optimal shape parameters and joint angle data for completing motion redirection, input them into the Isaac Gym simulation platform, and verify the physical stability of the motion (such as no joint jamming and no center of gravity shift).

[0058] The improved motion relocation method achieves accurate transfer of SMPL digital human pose to robot pose. Compared with the traditional motion relocation method, its advantages are: 1) The offline calibration stage helps to reduce the morphological differences between the digital human and the robot, reducing the optimization complexity of the online transfer stage; 2) Joint angle optimization and limit link can make the hardware compliance rate of the relocated pose reach 100%; 3) Ground contact correction helps to ensure the stability of walking, turning and other actions.

[0059] Considering the limited amount of data generated by the motion synthesis method, this embodiment redirects both the data generated by the motion synthesis method and the data provided by the AMASS dataset to provide training support for the reinforcement learning of the motion control model of the humanoid robot.

[0060] The core of Part Three, "Training of Humanoid Robot Motion Control Model Based on PPO Algorithm and Teacher-Student Model," is to transfer the anthropomorphic coordination characteristics of human motion to robot control strategies. Traditional methods have two significant limitations: first, they rely on motion capture data, which is costly to collect and has limited scene coverage; second, a single control strategy is difficult to adapt to multiple types of action tasks. This embodiment constructs a full-process training framework based on the PPO algorithm and teacher-student model, characterized by "high-precision teacher guidance, lightweight student distillation, and multi-task adaptation." The method flow is as follows: Figures 5 to 9 As shown, each point will be explained below.

[0061] like Figure 5As shown, the core value of pre-input and task formalization lies in transforming the abstract business requirement of "robots accurately replicating human movements and possessing generalization and adaptation capabilities" into a standard mathematical model that can be quantified and solved by reinforcement learning algorithms. This process is divided into two stages: pre-input definition and task Markov (Markov Decision Process, MDP) formalization, as explained in detail below: (1) Pre-input definition Pre-input is a prerequisite for MDP (Multi-Device Propagation), and its core task requirement is to enable humanoid robots to reproduce human movements, possessing multi-mode tracking and anti-occlusion generalization capabilities. The abstract task requirements are transformed into pre-input content, specifically including the following three categories: 1) Two quantifiable sub-objectives: First, motion reproduction accuracy, requiring the minimization of deviations between joint angles, poses, root node motion parameters, etc., and human reference motion. Second, motion stability, requiring the robot to avoid falls where the non-foot contact points touch the ground during movement, ensuring the continuity of trajectory execution. 2) Two constraints: First, robot hardware constraints, namely, the robot's joint configuration must be compatible with 19 controllable joints, the drive must use a proportional-derivative controller, receiving joint target angle commands, and the perception capability must be based on collected body position, posture, velocity, and other state data. Second, environmental constraints, covering common characteristics of simulation and real environments (such as gravity and friction) and environmental interference characteristics (such as sensor noise and signal delay). 3) Training data and evaluation benchmarks: Considering the limited amount of data generated by motion synthesis methods, the motion data used for training in this embodiment includes motion synthesis data processed through retargeting and data provided by the AMASS dataset. The evaluation criteria include stability indicators (such as survival rate) and motion accuracy indicators (such as joint errors).

[0062] (2) Task MDP This is mainly achieved by constructing the state space (S), action space (A), state transition probabilities (P), reward function (R), and discount factor (S). The five key elements complete the mathematical abstraction of the task, and their specific meanings are as follows: ① State space (S): A quantitative representation of the decision input The state space S is defined as the concatenated set of the robot's proprioceptive state and target motion state at time t, that is: Here, the semicolon indicates the concatenation of vector dimensions, and the ontology-aware state. It is the real-time motion data collected by the robot's own sensors during training, representing the target's motion state. This refers to the human reference motion (the training data itself) that the robot needs to reproduce. The design goal is to enable the robot to receive all the necessary information to predict future states and optimal actions without relying on historical data.

[0063] ② Action Space (A): Quantization definition of control output The design of motion space A directly matches the robot's hardware constraints, using a 19-dimensional vector. This indicates that each bit in the vector corresponds to the target angle of the 19 controllable joints of the humanoid robot (provided by training data). The design goal is to enable the output joint target angles to be directly input into a proportional-derivative (PD) controller, which then converts the angle commands into motor drive currents, achieving a closed-loop motion control.

[0064] ③ State transition probability (P): Implicit modeling of environmental dynamics State transition probability This describes the state of the robot at time t. And perform the action Then, the state transitions to the next moment. The probability distribution is essentially a model of the dynamic interaction between the robot and the environment, covering the influence of actions on the robot's state and the random characteristics of environmental disturbances (such as friction fluctuations and sensor noise).

[0065] ④ Reward Function (R): A quantitative benchmark for learning orientation reward function It is a direct quantitative representation of the task objective, and its core function is to provide the agent with clear learning guidance. That is, by quantifying the fit between the current agent state and the target state, it outputs the corresponding reward value, guiding the agent to learn the optimal action that maximizes the reward. The reward function adopts a composite design of three major terms, as shown in Table 3.

[0066] Table 3. Reward Function Design and Weighting of Teacher Strategies

[0067] ⑤ Discount factor ( ): Balancing and adjusting short-term and long-term goals Discount factor The parameter used to adjust the agent's emphasis on current and future rewards is a key parameter for balancing short-term motion accuracy and long-term motion stability. This embodiment uses... for This means that when making decisions, the agent must fully consider the reward feedback for multiple steps in the future, avoid sacrificing the stability of subsequent movements in pursuit of high precision of single-step joint angles, and ensure that the learned strategy has the ability to optimize the global trajectory.

[0068] ⑥ Termination condition: Boundary constraints of training rounds The termination condition is used to define the end time in reinforcement learning training. The design rule is that if a non-foot part touches the ground and falls, the training is terminated. That is, after each action is executed, the robot's contact sensor detects the ground contact status. If a non-foot part (such as the torso, arm, or head) touches the ground, it is determined to be a fall, the current round is terminated immediately, and the final reward is set to zero. If no fall occurs and the current round has completed the preset time steps (t=T, where T is the preset maximum time steps), the training is terminated normally, and the complete trajectory is stored in the training cache.

[0069] like Figure 6 As shown, Oracle strategy ( This is the teacher strategy within the entire framework. The training objective is to enable the robot to accurately reproduce the movements of the entire human skeleton under unconstrained conditions of full-information observation and maskless action space, through the PPO algorithm, thus providing high-quality action demonstrations for subsequent student strategy distillation. The complete training process is as follows: 1) Data preparation phase: From the feasible humanoid dataset that has undergone motion retargeting. Randomly sample a reference motion trajectory data This serves as the target movement benchmark for a single round of training.

[0070] 2) Training cycle phase: ① Initialize the current round count (Episode) and time step count (t), perform state observation, and obtain the teacher's strategy state at the current moment. ,in The ontological perception state of the teacher's strategy includes the robot's position. ,angle linear velocity angular velocity and the action of the previous moment Wait for complete ontology information, The target state for the teacher's strategy includes the inter-frame angle difference. Location difference Speed ​​difference angular velocity difference and absolute reference angle Reference position wait.

[0071] ② Based on the observed state, the teacher's strategy outputs the action at the current moment through probability sampling. (19-dimensional joint target angle) The proportional-differential controller of the robot converts the motion input into motor current, driving the robot to perform motion, and obtains two feedback information from the simulation environment: one is the reward at the current moment. Secondly, the state at the next moment. .

[0072] ③ Perform a fall detection. If no fall occurs, check if the current time step t has reached the preset maximum time step T. If not, increment time step t by 1 and continue the closed-loop interaction. If the maximum time step has been reached, transfer the valid data information (including the data at each time step). , , , Store in the experience cache.

[0073] 3) After the data is stored in the experience cache, it is determined whether the data accumulated in the cache has reached the preset update threshold K steps. If it has, the policy update process of the PPO algorithm is triggered, so that the teacher's policy gradually outputs the optimal movement that can perfectly reproduce human movement. a .

[0074] 4) Termination of training: After the strategy is updated, determine whether the training termination condition is met, that is, the strategy can stably reproduce the movement of the human skeleton, no falls occur in multiple consecutive rounds, and the movement reproduction error is lower than the preset threshold, and output the final teacher strategy.

[0075] like Figure 7 As shown, the goal of student policy initialization is to build an initial student model that can inherit the teacher's basic policy capabilities and is adaptable to subsequent multi-task distillation training. The specific process and details are as follows: 1) Input initialization: The input is the already trained teacher strategy. The basic information includes the structure type of the teacher strategy network, the dimensions of the hidden layers, the selection of activation functions, and the dimensions of action outputs.

[0076] 2) Design the student policy network structure, including two schemes: One is completely consistent with the teacher policy network, using a 3-layer fully connected neural network with hidden layer dimensions of 512, 256, and 128 respectively. The second is a lightweight network design, which reduces the number of model parameters and computational overhead by simplifying the hidden layer dimensions or reducing the number of network layers while ensuring core performance. Both schemes use the same activation function: ReLU activation for the hidden layers and Tanh activation for the output layer.

[0077] 3) Constructing a student observation space: stitching together 25 historical action frames to form a "dynamic observation window" containing time dimension information to enhance students' ability to perceive and generalize strategies.

[0078] like Figure 8 As shown, the command mask engine employs a dual-mask (pattern mask + sparse mask) mechanism to generate diverse sub-tasks for student policy distillation training, helping to address the pain points of traditional reinforcement learning strategies, such as single-task adaptability and poor generalization. The specific process is as follows: 1) Input requirements: This includes two levels. First, it generates diverse task modes, including pose tracking, joint angle tracking, and root motion stabilization control. Second, it generates sparse observation constraints to simulate sensor occlusion, information loss, and other situations in real-world scenarios, thereby enhancing the robustness of student strategies.

[0079] 2) Generate pattern mask : It supports independent decision-making between the upper and lower body layers. By masking the motion pose, joint angle, and root motion state obtained from the input motion data used for training, it corresponds to three control modes, specifically including pose tracking mode, joint angle tracking mode, and root motion stabilization control mode.

[0080] 3) Generate sparse mask : By randomly discarding some joint information from the input motion data used for training, the system simulates real-world scenarios such as sensor occlusion and data transmission loss, thereby improving the generalization ability of students' strategies.

[0081] like Figure 9 As shown, the Dagger algorithm is used to train the general student strategy. In this embodiment, the total number of rounds M is preset to be 1000. The specific process is as follows: 1) Initialization parameters: including total number of rounds M, maximum time step per round T, and policy update interval K.

[0082] 2) Generate a mask and construct the target state: Call the command mask engine to randomly generate a set of mask combinations ( + ). The reference pose generated by the teacher model ( For the upper body target, The target state of the student's strategy is obtained by performing pattern masking and sparse masking on the lower-level body target. : .

[0083] in, This is the Hadama product operation.

[0084] 3) Start the loop (t=1 to T): Based on the robot's state at the current time step t and historical data, construct the observation vector of the student's policy. , The ontology-aware state of the student's strategy is specifically inputted as 25 frames of historical stitched data, including joint positions from time t-25 to time t. Joint velocity , base angular velocity ,gravity And the action history from t-25 to t-1 , The target state of the student policy is defined. The observation vector is then input into the student policy. The action at the current time step is output through network forward propagation calculation. The output action Inputting data into the robot's PD controller, the robot simulation environment calculates and returns the state for the next moment based on the physics engine. and the reward for the current time step The teacher's strategy, based on the current time step, generates the optimal action for the current moment. The student observations, student goals, and teacher actions at the current time step are encapsulated as supervised samples to guide students in learning the mapping relationship between the input observation vector and the output optimal action.

[0085] 4) Integrate samples and statistically analyze core metrics: After the current round's time step interaction ends, merge all supervised samples in the temporary cache of this round into the global supervised experience cache. Statistically analyze the core metrics of this round, including the mean square error of student and teacher actions, robot motion error, etc.

[0086] 5) Update strategy: When the number of samples in the global cache reaches the preset threshold or K rounds of effective training are completed, the student policy network parameters are updated using supervised learning and the student policy learning continues.

[0087] 6) Termination Loop: When the preset total number of rounds M is completed or the student policy converges, and the motion reproduction can be stably completed in the three modes of the task, the distillation loop ends, and the trained general student policy is output. The student model takes into account the humanoid robot's pose and control mode based on axis angles, outputs the optimal motion sequence (driving current), and drives the humanoid robot.

[0088] The experimental results of the training process are as follows Figure 10 As shown, reinforcement learning training was completed using the Isaac Lab virtual platform, with the number of humanoid robots in the environment, num, set to 10. Figure 10 Part (a) shows the simulation results of the motion control model based on reinforcement learning training iteration 1000 times. All robots are almost standing still, indicating that the controller training is trapped in a local optimum. The controller would rather choose to "stay still" than explore new actions. In essence, it avoids the punishment caused by falling and collision by giving up the task movement reward. Figure 10 Part (b) shows the simulation results of the motion control model based on reinforcement learning training iterations of 5000 times. The robot is still almost standing still. Figure 10Part (c) shows the simulation results of the motion control model based on 10,000 reinforcement learning training iterations. The humanoid robot can basically move, but it falls down most of the time after taking its first step. It wasn't until after 50,000 reinforcement learning training iterations that, as... Figure 10 As shown in part (d), the humanoid robot based on the motion control model can walk normally and perform specified actions, but a few robots still exhibit abnormal behavior. After reinforcement learning training iterates to 100,000 times, as... Figure 10 As shown in section (e), all humanoid robots can move normally and perform specified actions based on the motion control model.

[0089] Training evaluation data statistics as follows Figure 11 As shown, Figure 11 Intermediate survival rate ( Survive This refers to the percentage of robots that maintain stable movement without falling; a higher value indicates better movement stability. The true value represents the average error of joint positions. This refers to the average deviation between the actual position and the true reference position of a robot joint, measuring the accuracy of joint position prediction; a smaller value is better. (Joint position average error) This refers to the average deviation of joint positions after attitude alignment, aiming to reduce global rotational and translational interference and more accurately assess joint attitude fit; acceleration error Velocity error refers to the deviation between the actual acceleration of a robot joint or end effector and the target acceleration, reflecting the smoothness of motion; This refers to the deviation between the actual speed and the target speed of a robot's joints, measuring the real-time performance of motion trajectory tracking; upper limb joint angle deviation. This refers to the angular deviation of the robot's upper limb (shoulder, elbow, wrist, etc.) joints, used to assess the accuracy of upper limb motion control; lower limb joint angular deviation... This refers to the angular deviation of the robot's lower limb (hip, knee, ankle, etc.) joints, which is the core factor in evaluating the robot's walking, running, jumping, and other motor performance; root node velocity error. This refers to the velocity deviation of the robot's torso root node, measuring the overall coordination of its motion; root node height error. This refers to the vertical height deviation of the robot's torso root node, used to assess center of gravity stability during walking; root node roll angle error. This refers to the roll angle deviation (left and right tilt) of the robot's torso root node, reflecting the torso balance during movement; root node pitch angle error This refers to the pitch angle deviation (forward and backward tilt) of the robot's torso root node, which measures the accuracy of torso posture control during walking / climbing; root node yaw angle error. It refers to the yaw angle deviation (left and right turning) of the robot's trunk root node, which is used to evaluate the accuracy of the turning motion.

[0090] According to the training evaluation data, as the number of reinforcement learning training iterations increases, the task completion rate of the robot based on the motion control model rapidly increases from 0.12% after 500 iterations to 80.46% after 100,000 iterations. After 5,000 iterations, it enters a rapid growth phase, and after 30,000 iterations, it tends to stabilize in the range of 77% to 81%. This shows that imitation reinforcement learning effectively improves the robot's ability to complete target motion tasks, and the performance tends to saturate in the later iterations, with the task completion rate reaching a high level.

[0091] The global joint position error decreased from 176.480 in the initial stage to 109.950 after 100,000 iterations, and the partial alignment position error decreased from 68.066 to 52.702, showing a significant downward trend overall. The rate of decrease accelerated after 5,000 iterations, and the fluctuations converged in the later stage. This indicates that the alignment accuracy between the robot's motion trajectory and the target motion continued to improve, and the trajectory fitting effect of imitation learning was continuously optimized with the increase of the number of iterations.

[0092] The acceleration error decreased from 13.754 initially to 4.768 after 100,000 iterations, and the velocity error decreased from 22.360 to 8.629. After 5,000 iterations, the velocity error decreased significantly and stabilized in the low range. This indicates that the smoothness of the robot's motion and the accuracy of its speed control have been significantly improved, and the stability of the motion process has been continuously enhanced with training iterations.

[0093] The upper limb joint error decreased from 0.246 to 0.154, and the lower limb joint error decreased from 0.356 to 0.258, showing an overall downward trend. The improvement in the lower limb joint error was particularly significant, indicating that the motion coordination and accuracy of the robot's upper and lower limb joints continued to improve, and the effect of imitation learning on the refined control of joint motion gradually became apparent.

[0094] The root node velocity error decreased from 0.095 to 0.055, the height error decreased from 0.70 to 0.318, and the roll, pitch, and yaw attitude errors also decreased from the initial 0.116, 0.137, and 0.196 to 0.068, 0.072, and 0.192, respectively. This indicates that the motion stability and attitude control accuracy of the robot's root node have been greatly improved, providing a more stable foundation for the overall motion.

[0095] Based on the data above, it can be observed that in the first 5000 iterations, the various indicators fluctuated significantly but began to decline overall, while the completion rate rose rapidly, indicating that the model quickly mastered the basic motion patterns. From 5000 to 30000 iterations, both the completion rate and error indicators showed significant improvement, and the model gradually mastered precise motion control strategies, resulting in a significant increase in stability. After 30000 iterations, the various indicators tended to stabilize, the completion rate remained at around 80%, the error indicators fluctuated and converged, and the model performance was close to optimal, with only a small room for optimization.

[0096] like Figure 12 The figure shown illustrates how the values ​​of different evaluation indicators change with the number of iterations. Figure 12 Part (a) is a graph of the training process data, showing the survival rate versus the main error. Figure 12 Part (b) is a data graph of the training process for acceleration and velocity error. Figure 12 Part (c) is a data graph of the training process for the joint angle deviation of the upper and lower body. Figure 12 Part (d) is a training process data diagram of the root node related error. It can be seen that all indicators converge around 10,000 iterations.

[0097] For Part Four, "Sim-to-Sim Deployment and Verification," this part can be implemented using the motion control of the Unitree H1 robot. The proposed motion control method is deployed and verified using the Isaac Sim simulation platform, such as... Figure 13 As shown in the figure. Experimental results show that the method proposed in this application can significantly improve the text semantic consistency, trajectory tracking accuracy, and robustness to real-world scenarios of humanoid robot motion, while also exhibiting good anthropomorphism and stability of motion.

[0098] Compared with existing technologies, this application has the following beneficial effects: 1. This application adopts a motion synthesis technology with joint constraints of text and trajectory. Text controls the motion type at the macro level, while trajectory controls the movement path at the micro level. The two constraints complement each other, providing high-quality prior guidance for subsequent model training, effectively strengthening the consistency between humanoid robot motion and text control semantics, and significantly improving the control accuracy of motion trajectory. 2. This application optimizes the motion retargeting method, adopting an optimization strategy combining stochastic gradient descent and momentum methods, along with a complete process of shape parameter adaptation, joint mapping, optimization solution, and safety filtering. This effectively reduces the kinematic structural differences between SMPL digital humans and humanoid robots, reduces joint jitter and over-limit risks, improves the physical feasibility and control accuracy of pose transfer, and enables the pose data generated by motion synthesis to be directly adapted to the hardware constraints of humanoid robots. 3. This application constructs a teacher-student model based on a proximal policy optimization algorithm, and performs policy distillation through a dataset aggregation algorithm. This transfers the high-precision motion skills of the teacher model, which possesses privileged information, to the lightweight student model. Combined with a dual-mask mechanism, multi-mode control is achieved, effectively improving the generalization ability, anti-interference ability, and robustness to real-world scenarios of the humanoid robot's motion control model, while ensuring the anthropomorphism and stability of the motion. 4. The method in this application does not rely on high-cost motion capture or teleoperation equipment, significantly reducing the training cost of the humanoid robot's motion control model and improving training efficiency. It can also be integrated with existing global path planning and local obstacle avoidance algorithms, possessing strong scenario adaptability and engineering application value.

[0099] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0100] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for human-shaped robot motion control with text and trajectory joint constraints, characterized in that, include: Obtain text descriptions and motion trajectories corresponding to several preset control actions, and generate SMPL digital human pose sequences that simultaneously conform to the text descriptions and motion trajectories through a motion synthesis method that combines text and trajectory constraints. An optimization strategy combining stochastic gradient descent and momentum methods is adopted to perform motion redirection processing on the SMPL digital human pose sequence to obtain robot pose data corresponding to each preset control action. A teacher-student model is constructed based on the proximal policy optimization algorithm. The motion skills of the teacher model are distilled into the student model using a dataset aggregation algorithm. The robot pose data corresponding to each preset control action obtained after motion redirection are used as prior guidance to train the humanoid robot motion control model. The text description and motion trajectory corresponding to the current desired control action are obtained, and then processed by the motion synthesis method and motion relocation method based on the joint constraints of text and trajectory to obtain the robot pose corresponding to the current desired control action. The robot pose corresponding to the current desired control action is input into the humanoid robot motion control model, and the joint target angle position output by the humanoid robot motion control model is used as the driving signal to control the movement of the target humanoid robot; the humanoid robot motion control model is deployed to the target humanoid robot.

2. The text and trajectory joint constraint humanoid robot motion control method according to claim 1, characterized in that, When generating the SMPL digital human pose sequence, motion type semantic features are extracted from the text description, and a movement path based on three-dimensional spatial coordinates is extracted from the motion trajectory. The generated SMPL digital human pose sequence simultaneously satisfies the constraints of the motion type semantic features and the movement path. The motion type semantic features define what action the target humanoid robot should take, and the movement path defines the path along which the target humanoid robot takes the action.

3. The humanoid robot motion control method based on text and trajectory joint constraints according to claim 1, characterized in that, The motion reorientation processing sequentially comprises four sub-steps of shape parameter adaptation, joint mapping, optimization solving and safety filtering; the shape parameter adaptation is to optimize shape parameters of the SMPL digital person β match the SMPL digital person structure with the limb proportion of the target humanoid robot; the joint mapping is to select a plurality of key joints to establish a one-to-one correspondence between the SMPL digital person and the target humanoid robot; the optimization solving is to construct a function with the minimum joint position difference between the SMPL digital person and the target humanoid robot as the target; and the safety filtering is to remove the motion sequence with joint over-limit or abnormal ground contact.

4. The humanoid robot motion control method based on text and trajectory joint constraints according to claim 1, characterized in that, A teacher-student model is constructed based on a proximal policy optimization algorithm. A dataset aggregation algorithm is used to distill the motion skills of the teacher model into the student model. Using the robot pose data corresponding to each preset control action obtained after motion redirection as prior guidance, a humanoid robot motion control model is trained, specifically including: A feasible motion dataset is constructed using the robot pose data corresponding to each preset control action obtained after motion redirection. A teacher-student model architecture is constructed based on the proximal policy optimization algorithm, and the teacher model is trained using a feasible motion dataset as the training basis. A dataset aggregation algorithm is used to distill the motion control strategy of the teacher model into the student model, thereby training a motion control model for the humanoid robot.

5. The humanoid robot motion control method based on joint text and trajectory constraints according to claim 4, characterized in that, The ontological perception state of the input data of the teacher model includes the rigid body position, posture, linear velocity, and angular velocity of the target humanoid robot over several consecutive historical steps. The target state is the robot reference pose data corresponding to the preset control action obtained after motion retargeting and the difference between the current pose data and the robot reference pose data. A multi-dimensional composite reward function containing penalty terms, regularization terms, and task reward terms is optimized using a proximal policy optimization algorithm to learn a high-precision motion control strategy that meets the preset control action pose requirements. The output data of the teacher model is used as a reference motion pose for the student model to learn.

6. The humanoid robot motion control method based on text and trajectory joint constraints according to claim 5, characterized in that, The input to the student model includes proprioceptive data and a target state processed by double masking. The proprioceptive data includes the joint positions, joint linear velocities, base angular velocities, and gravity vectors of the target humanoid robot's historical action sequence with a preset number of steps. The target state is processed by pattern masking and sparse masking to form multi-mode control commands. When performing policy distillation using a dataset aggregation algorithm, the student model generates motion trajectories in the simulation environment, and the optimal reference action at the corresponding moment is obtained simultaneously through the teacher model. The L2 norm loss between the student model's output action and the teacher model's output action is used as the optimization objective, and the transfer of motor skills to the student model is completed through supervised learning.

7. The humanoid robot motion control method based on joint text and trajectory constraints according to claim 6, characterized in that, The motion control strategy of the teacher model is distilled into the student model using a dataset aggregation algorithm, and a humanoid robot motion control model is trained, which specifically includes: Initialize the preset total number of rounds, maximum time step per round, and policy update interval for distillation training; load the pre-trained teacher model and the initialized student model; and establish a global supervised experience cache. The combination of the generated pattern mask and sparse mask is based on the reference motion pose output by the teacher model. After processing by the pattern mask and sparse mask, the target state data of the student model for the corresponding training round is obtained. In each time step of a single round, the observation vector of the student model is constructed based on the real-time ontology perception data of the target humanoid robot at the corresponding time step and the target state at the corresponding time step. The observation vector is input into the student model to output the target angle action of the corresponding controllable joint, and the action is input into the simulation environment to drive the humanoid robot to perform the movement. The optimal reference action for the corresponding time step is obtained synchronously through the teacher model, and the observation vector, target state and optimal reference action are encapsulated as a single sample and stored in a temporary cache. After a single training round, all single samples in the temporary cache are merged into the global supervised experience cache. When the number of samples in the global supervised experience cache reaches the preset update threshold or the preset number of training batches are completed, the network parameters of the student model are updated through supervised learning with the L2 norm loss of the student model output action and the optimal reference action output by the teacher model as the optimization objective. If the iteration termination condition is not met, the process jumps to the step "Generate a combination of pattern mask and sparse mask, and based on the reference motion pose output by the teacher model, process the pattern mask and sparse mask to obtain the target state of the student model for the corresponding round of training", until the iteration termination condition is met, and the trained student model is output as the humanoid robot motion control model; the iteration termination condition is to complete the preset total number of rounds or the student model converges and can stably reproduce the motion trajectory corresponding to the preset control action.

8. The humanoid robot motion control method based on joint text and trajectory constraints according to claim 6, characterized in that, The network structure of the student model is the same as that of the teacher model, consisting of a 3-layer fully connected neural network with hidden layer dimensions of 512, 256, and 128 respectively. The hidden layers use the ReLU activation function, and the output layer uses the Tanh activation function.

9. The humanoid robot motion control method based on joint text and trajectory constraints according to claim 6, characterized in that, The network structure of the student model is based on the network structure of the teacher model. It adopts a lightweight network design to simplify the network structure of the model while ensuring core performance, thereby reducing the number of model parameters and computational overhead.

10. The humanoid robot motion control method based on joint text and trajectory constraints according to claim 6, characterized in that, The humanoid robot motion control model supports pose tracking mode, joint angle tracking mode, and root motion stabilization control mode, and adapts to different task scenarios by dynamically configuring mode masks and sparse masks.

Images