Humanoid robot control method and system

Abstract

The invention discloses a humanoid robot control method and system, the system comprises a sensor module, a sensing module, a state representation module, a decision module, a motion control module, an actuator module and a feedback module, and the method comprises the following steps: step 1, data acquisition; step 2, environment perception; step 3, state representation; step 4, action selection; step 5, motion control; step 6, performing feedback and adjustment; according to the method, the hybrid model combining the ViT and the CNN is adopted for environment perception, the environment perception precision is improved, the error recognition rate is reduced, motion planning is carried out through reinforcement learning, the robot can autonomously learn the optimal motion strategy, decision making and adjustment can be autonomously carried out according to the environment change and the task target, and the robot experience is improved. The problem that a traditional motion planning algorithm is high in calculation complexity is solved, the robot motion planning efficiency, the autonomous decision-making ability and the environment adaptability are improved, and safe, efficient and autonomous motion of the humanoid robot in a complex dynamic environment is achieved.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and specifically to a control method and system for a humanoid robot. Background Art

[0002] Due to its human-like form and potential adaptability, humanoid robots have broad application prospects in fields such as service, medical, and industry. In the existing technology, control methods adopted by humanoid robots, such as PID control and model-based control, perform well in simple environments. However, in complex, dynamic, and unstructured environments, their performance is often limited. With the rise of deep learning technology, a new path has been opened up for robot control. Control methods based on deep learning have the ability to learn complex non-linear relationships and enable robots to better adapt to environmental changes. However, the existing robot control methods based on deep learning still have the following disadvantages: 1. Insufficient environmental perception accuracy: The existing technology generally uses convolutional neural networks to process environmental information. However, in the face of complex scenes, it is easily affected by factors such as occlusion and lighting changes, resulting in a decrease in object recognition accuracy; 2. Low motion planning efficiency: Robot motion planning algorithms adopted by the existing technology, such as RRT and PRM, have high computational complexity, resulting in long operation time-consuming and being difficult to meet the real-time requirements of humanoid robots; 3. Lack of autonomous decision-making ability: Existing robot control systems often rely on artificially designed rules and strategies and lack the ability of autonomous learning and adaptation. In the face of new and unpreset scenarios, robots are often at a loss. Summary of the Invention

[0003] The purpose of the present invention is to provide a control method and system for a humanoid robot to solve the problems raised in the above background art.

[0004] To achieve the above purpose, the present invention provides the following technical solution: A control system for a humanoid robot, including a sensor module, a perception module, a state representation module, a decision-making module, a motion control module, an actuator module, and a feedback module. The sensor module respectively establishes data connections with the perception module, the state representation module, and the feedback module. The perception module establishes a data connection with the state representation module. The decision-making module respectively establishes data connections with the state representation module, the motion control module, and the feedback module. The motion control module establishes a data connection with the actuator module.

[0005] Preferably, the sensor module is used to collect information about the environment around the robot and the robot's own state information. The perception module is used to process the information about the environment around the robot to obtain environmental feature information. The state representation module is used to fuse the environmental feature information, the robot's own state information, and the task objective to obtain a state representation. The decision-making module is used to select the optimal action according to the state representation. The motion control module is used to convert the action instruction into a specific joint control instruction. The actuator module is used to execute the joint control instruction. The feedback module is used to feedback error information to the decision-making module.

[0006] Preferably, the sensor module includes a camera, a lidar, an inertial measurement unit, a joint angle sensor, and a force and torque sensor. The camera is used to obtain visual information. The lidar is used to obtain distance information. The inertial measurement unit is used to obtain the robot's attitude and acceleration. The joint angle sensor is used to obtain the current angles of the robot's various joints. The force and torque sensor is used to measure the force and torque received by the joints.

[0007] Preferably, the perception module includes a local feature extraction unit, a global feature extraction unit, a feature fusion unit, and a target detection and recognition unit. The local feature extraction unit is used to extract the local features of the image and output a feature map. The global feature extraction unit takes the feature map as input and outputs the global features of the image. The feature fusion unit is used to fuse the global features and local features of the image and output a fused feature. The target detection and recognition unit takes the fused feature as input and outputs environmental feature information.

[0008] Preferably, the motion control module includes a high-level motion planning unit and a low-level motion control unit. The high-level motion planning unit is used to convert the abstract action instruction into a specific motion trajectory. The low-level motion control unit is used to generate the joint control instructions of the robot according to the specific motion trajectory.

[0009] A control method for a humanoid robot includes Step 1: data collection; Step 2: environmental perception; Step 3: state representation; Step 4: action selection; Step 5: motion control; Step 6: feedback and adjustment.

[0010] In Step 1 above, the sensor module collects information about the environment around the robot and the robot's own state information, and sends the information about the environment around the robot to the perception module and the robot's own state information to the state representation module.

[0011] In Step 2 above, the perception module processes the information about the environment around the robot, extracts environmental feature information, and sends the environmental feature information to the state representation module.

[0012] In step 3 above, the state representation module fuses the environmental feature information, the robot's own state information, and the task objective to form the state representation of the robot and sends it to the decision-making module;

[0013] In step 4 above, based on the state representation, the decision-making module uses the policy network trained by the reinforcement learning algorithm to select the optimal action and sends it to the motion control module;

[0014] In step 5 above, the motion control module converts the selected action into joint control instructions and sends them to the actuator module. The actuator module executes the joint control instructions to drive the joints of the robot to move and achieve the desired motion trajectory;

[0015] In step 6 above, the feedback module obtains the actual motion result and environmental change information of the robot, compares them with the desired motion result, and then feeds back the error information to the decision-making module. The decision-making module optimizes the control strategy based on the error information.

[0016] Preferably, in step 2 above, the perception module uses a ViT combined with a CNN hybrid model to process the environmental information around the robot. Specifically, the local feature extraction unit uses the CNN algorithm to extract the local features of the image, outputs the feature map to the global feature extraction unit and the feature fusion unit. The global feature extraction unit uses the ViT algorithm to divide the feature map into multiple patches, then maps these patches into vectors, and inputs them into the Transformer encoder. The Transformer encoder uses the self-attention mechanism to process these vectors to obtain the global feature representation of the image and sends it to the feature fusion unit. The feature fusion unit fuses the global feature and the local feature of the image, outputs the fused feature to the target detection and recognition unit, and the target detection and recognition unit outputs the environmental feature information.

[0017] Preferably, the local feature extraction unit uses the ResNet50 model and is pre-trained using the ImageNet dataset. The global feature extraction unit uses the ViT model, the patch size is 16, the number of encoder layers is 12, the hidden size is 768, the feature fusion unit uses the concatenate fusion model, and the feature fusion unit uses the Softmax classifier.

[0018] Preferably, in step 3 above, the environmental feature information includes the position information of objects, obstacles, and targets, and the robot's own state information includes the position, attitude, speed, joint angle, and joint angular velocity of the robot.

[0019] Preferably, in step 4 above, the reinforcement learning algorithm used by the decision-making module is the improved DQN algorithm.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention uses a hybrid model combining ViT and CNN for environmental perception, improving the environmental perception accuracy and reducing the misrecognition rate. It uses reinforcement learning for motion planning, enabling the robot to autonomously learn the optimal motion strategy, make decisions and adjustments autonomously according to environmental changes and task objectives, solve the problem of high computational complexity of traditional motion planning algorithms, improve the robot's motion planning efficiency, autonomous decision-making ability and environmental adaptability, and achieve safe, efficient and autonomous motion of the humanoid robot in a complex dynamic environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is the system structure block diagram of the present invention;

[0022] Figure 2 is the sensor module structure block diagram of the present invention;

[0023] Figure 3 is the perception module structure block diagram of the present invention;

[0024] Figure 4 is the motion control module structure block diagram of the present invention;

[0025] Figure 5 is the method flow chart of the present invention;

[0026] Figure 6 is the schematic diagram of the ViT+CNN hybrid model structure of the perception module of the present invention;

[0027] Figure 7 is the architecture diagram of the multi-head attention mechanism of the perception module of the present invention;

[0028] Figure 8 is the algorithm flow chart of the decision module of the present invention

[0029] Figure 9 is the architecture diagram of the motion control module of the present invention.

[0030] In the figure: 1, sensor module; 11, camera; 12, lidar; 13, inertial measurement unit; 14, joint angle sensor; 15, force and torque sensor; 2, perception module; 21, local feature extraction unit; 22, global feature extraction unit; 23, feature fusion unit; 24, target detection and recognition unit; 3, state representation module; 4, decision module; 5, motion control module; 51, high-level motion planning unit; 52, low-level motion control unit; 6, actuator module; 7, feedback module. DETAILED DESCRIPTION OF THE INVENTION

[0031] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0032] Please refer to the attached Figure 1 - attached Figure 4 , an embodiment provided by the present invention: A control system for a humanoid robot, including a sensor module 1, a perception module 2, a state representation module 3, a decision-making module 4, a motion control module 5, an actuator module 6, and a feedback module 7. The sensor module 1 establishes data connections with the perception module 2, the state representation module 3, and the feedback module 7 respectively. The perception module 2 establishes a data connection with the state representation module 3. The decision-making module 4 establishes data connections with the state representation module 3, the motion control module 5, and the feedback module 7 respectively. The motion control module 5 establishes a data connection with the actuator module 6. The sensor module 1 is used to collect information about the robot's surrounding environment and the robot's own state information. The perception module 2 is used to process the information about the robot's surrounding environment to obtain environmental feature information. The state representation module 3 is used to fuse the environmental feature information, the robot's own state information, and the task objective to obtain a state representation. The decision-making module 4 is used to select the optimal action according to the state representation. The motion control module 5 is used to convert the action instruction into a specific joint control instruction. The actuator module 6 is used to execute the joint control instruction. The feedback module 7 is used to feedback error information to the decision-making module 4. The sensor module 1 includes a camera 11, a lidar 12, an inertial measurement unit 13, a joint angle sensor 14, and a force and torque sensor 15. The camera 11 is used to obtain visual information. The lidar 12 is used to obtain distance information. The inertial measurement unit 13 is used to obtain the robot's posture and acceleration. The joint angle sensor 14 is used to obtain the current angles of the robot's various joints. The force and torque sensor 15 is used to measure the force and torque received by the joints. The perception module 2 includes a local feature extraction unit 21, a global feature extraction unit 22, a feature fusion unit 23, and a target detection and recognition unit 24. The local feature extraction unit 21 is used to extract local features of the image and output a feature map. The global feature extraction unit 22 takes the feature map as input and outputs the global features of the image. The feature fusion unit 23 is used to fuse the global features and local features of the image and output fused features. The target detection and recognition unit 24 takes the fused features as input and outputs environmental feature information. The motion control module 5 includes a high-level motion planning unit 51 and a low-level motion control unit 52. The high-level motion planning unit 51 is used to convert the abstract action instruction into a specific motion trajectory. The low-level motion control unit 52 is used to generate the joint control instructions of the robot according to the specific motion trajectory.

[0033] Please refer to the attached Figure 5 - Attachment Figure 9 , an embodiment provided by the present invention: a control method for a humanoid robot, including Step 1, data acquisition; Step 2, environmental perception; Step 3, state representation; Step 4, action selection; Step 5, motion control; Step 6, feedback and adjustment;

[0034] Among them, in the above Step 1, the sensor module 1 collects the environmental information around the robot and the robot's own state information, and sends the environmental information around the robot to the perception module 2, and sends the robot's own state information to the state representation module 3;

[0035] Among them, in the above Step 2, the perception module 2 processes the environmental information around the robot, extracts the environmental feature information, and sends the environmental feature information to the state representation module 3; among them, the perception module 2 uses a hybrid model of ViT combined with CNN to process the environmental information around the robot. Specifically: the local feature extraction unit 21 uses the CNN algorithm to extract the local features of the image, outputs the feature map to the global feature extraction unit 22 and the feature fusion unit 23. The global feature extraction unit 22 uses the ViT algorithm to divide the feature map into multiple patches, then maps these patches into vectors, and inputs them into the Transformer encoder. The Transformer encoder uses the self-attention mechanism to process these vectors to obtain the global feature representation of the image, and sends it to the feature fusion unit 23. The feature fusion unit 23 fuses the global feature and the local feature of the image, outputs the fused feature to the target detection and recognition unit 24, and the target detection and recognition unit 24 outputs the environmental feature information; the local feature extraction unit 21 uses the ResNet50 model and is pre-trained using the ImageNet dataset. The global feature extraction unit 22 uses the ViT model, the patch size is 16, the number of encoder layers is 12, the hidden size is 768, the feature fusion unit 23 uses the concatenate fusion model, and the feature fusion unit 23 uses the Softmax classifier;

[0036] Among them, in the above Step 3, the state representation module 3 fuses the environmental feature information, the robot's own state information, and the task target to form the state representation of the robot, and sends it to the decision-making module 4; among them, the environmental feature information includes the position information of objects, obstacles, and targets, and the robot's own state information includes the position, attitude, speed, joint angle, and joint angular velocity of the robot;

[0037] Among them, in the above Step 4, the decision-making module 4 selects the optimal action based on the state representation using the policy network trained by the improved DQN algorithm, and sends it to the motion control module 5;

[0038] Among them, in the above step 5, the motion control module 5 converts the selected action into joint control instructions and sends them to the actuator module 6. The actuator module 6 executes the joint control instructions to drive the joints of the robot to move and achieve the desired motion trajectory;

[0039] Among them, in the above step 6, the feedback module 7 obtains the actual motion results and environmental change information of the robot, compares them with the desired motion results, and then feeds back the error information to the decision-making module 4. The decision-making module 4 optimizes the control strategy based on the error information.

[0040] Based on the above, the advantages of the present invention are as follows. When the invention is in use, the sensor module 1 collects the surrounding environment information of the robot and the robot's own state information, sends the surrounding environment information of the robot to the perception module 2, and sends the robot's own state information to the state representation module 3. The sensor module 1 obtains visual information through the camera 11, obtains distance information through the lidar 12, obtains the attitude and acceleration of the robot through the inertial measurement unit 13, obtains the current angles of each joint of the robot through the joint angle sensor 14, and measures the forces and torques received by the joints through the force and torque sensor 15; the perception module 2 obtains the surrounding environment information of the robot, extracts the local features of the image through the local feature extraction unit 21, outputs the feature map, the global feature extraction unit 22 takes the feature map as input, outputs the global features of the image, the feature fusion unit 23 fuses the global features and local features of the image, outputs the fused features, the target detection and recognition unit 24 takes the fused features as input, outputs the environmental feature information, and sends the environmental feature information to the state representation module 3; the state representation module 3 fuses the environmental feature information, the robot's own state information, and the task objective to form the state representation of the robot, and sends it to the decision-making module 4. The policy network trained by using the improved DQN algorithm selects the optimal action and sends the selected action to the motion control module 5. The motion control module 5 converts the abstract action instruction into a specific motion trajectory through the high-level motion planning unit 51, generates the control instructions for each joint of the robot according to the specific motion trajectory through the low-level motion control unit 52, and sends the control instructions for each joint to the actuator module 6 for execution. The feedback module 7 receives the output of the sensor module 1, compares it with the expected motion result, and then feeds back the error information to the decision-making module 4. The decision-making module 4 optimizes the control strategy accordingly to improve the control accuracy and robustness; in the perception module 2, the ViT algorithm adopted by the global feature extraction unit 22 introduces a spatial encoding mechanism, embeds the spatial position information of the input data into the feature representation, and enhances the model's ability to model spatial relationships. The specific method can be as follows: absolute position encoding: directly encodes the coordinate information of pixels or points as a position vector, and adds or concatenates it with the feature vector; relative position encoding: calculates the relative position relationship between pixels or points and encodes it as a position vector; learnable position encoding: sets the position encoding as a learnable parameter and automatically learns the optimal position representation through network training; based on the self-attention mechanism of the Transformer, the global feature extraction unit 22 introduces a bidirectional cross-attention mechanism to capture the dependency relationships between features from both the horizontal and vertical directions respectively, and enhances the model's ability to understand global context information. The specific implementation method can be: horizontal cross-attention: divides the feature map by rows, calculates the attention weights between different rows, and realizes information interaction in the horizontal direction; vertical cross-attention: divides the feature map by columns, calculates the attention weights between different columns, and realizes information interaction in the vertical direction;Alternate horizontal and vertical cross-attention: Alternate horizontal and vertical cross-attention in different layers of the model to achieve more comprehensive information interaction; The global feature extraction unit 22 fuses the outputs of spatial perception and bidirectional cross-attention to obtain the final feature representation. The specific implementation method can be: performing weighted summation on the outputs of different modules, concatenating the outputs of different modules, or using the attention mechanism to perform weighted fusion on the outputs of different modules; In the improved DQN algorithm adopted by the decision-making module 4, the state space includes the robot's own state information and environmental feature information, the action space includes discrete actions and continuous actions, and the reward function includes positive rewards and negative rewards. The positive rewards include proximity-to-goal rewards, reaching-goal rewards, and smooth movement rewards, and the negative rewards include obstacle collision penalties, out-of-bounds penalties, falling penalties, time penalties, and energy consumption penalties; The decision-making module 4 can use the potential field method or distance rewards to accelerate learning. The potential field method sets an attractive potential field near the target position and a repulsive potential field near the obstacle, and uses the potential field value as part of the reward. The distance reward gives different degrees of rewards according to the distance between the robot and the target; The improved DQN algorithm can perform prioritized experience replay, sample experiences according to the importance of the experiences, such as TD error, to improve learning efficiency; It can adopt multi-step learning, use multi-step returns to update the Q value, and reduce bias; It can adopt a noisy network, add noise to the weights of the neural network to enhance the exploration ability; It can adopt distributed reinforcement learning, use multiple agents to train in parallel to accelerate the learning process; It can adopt imitation learning, combine expert data to accelerate the learning process; It can adopt curriculum learning, start training from simple tasks, and gradually increase the difficulty of the tasks; It can combine traditional control methods, use the output of DQN as the input of the high-level planner, and still use traditional PID or model predictive control at the bottom layer to control the joints of the robot; The specific process of training the policy network using the improved DQN algorithm is as follows: a) Initialize the parameters of the Q-Network and Target Network, and initialize the experience buffer; b) Obtain the current state state; c) Select an action: Select an action according to the exploration strategy, such as ε-greedy; d) Execute the action: Execute the selected action in the simulation environment or real environment to obtain the next state next_state, reward reward, and whether it is over done; e) Store the experience: Store state, action, reward, next_state, and done in the experience buffer; f) Sample and train: Randomly sample a batch of experiences from the experience buffer; g) Calculate the target Q value: Calculate the target Q value using the Target Network; h) Calculate the loss: Calculate the mean squared error MSE or Huber loss between the current Q value and the target Q value; i) Update the Q-Network: Use the optimizer to update the parameters of the Q-Network according to the loss;j) Update the Target Network: Regularly copy the parameters of the Q-Network to the Target Network; k) State update: Set the next_state as the current state, return to step b), and start the next loop; In the motion control module 5, use the inverse kinematics algorithm to convert the position and orientation of the end effector into joint angles, which can be implemented using methods such as analytical solution and numerical solution, use spline curves to generate smooth motion trajectories, and use a PID controller to control the joint angles of the robot.;

[0041] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be embraced within the present invention. Any reference signs in the claims should not be construed as limiting the claims involved.

Claims

1. A control system for a humanoid robot, comprising a sensor module (1), a perception module (2), a state representation module (3), a decision-making module (4), a motion control module (5), an actuator module (6) and a feedback module (7), characterized in that: The sensor module (1) establishes data connections with the perception module (2), the state representation module (3), and the feedback module (7) respectively. The perception module (2) establishes a data connection with the state representation module (3). The decision-making module (4) establishes data connections with the state representation module (3), the motion control module (5), and the feedback module (7) respectively. The motion control module (5) establishes a data connection with the actuator module (6).

2. The control system of a humanoid robot according to claim 1, characterized in that: The sensor module (1) is used to collect information about the environment around the robot and the state of the robot itself. The perception module (2) is used to process the information about the environment around the robot to obtain environmental feature information. The state representation module (3) is used to fuse the environmental feature information, the state information of the robot itself, and the task objective to obtain a state representation. The decision-making module (4) is used to select the optimal action according to the state representation. The motion control module (5) is used to convert the action instruction into a specific joint control instruction. The actuator module (6) is used to execute the joint control instruction. The feedback module (7) is used to feedback error information to the decision-making module (4).

3. The control system of a humanoid robot according to claim 2, characterized in that: The sensor module (1) includes a camera (11), a lidar (12), an inertial measurement unit (13), a joint angle sensor (14), and a force and torque sensor (15). The camera (11) is used to obtain visual information. The lidar (12) is used to obtain distance information. The inertial measurement unit (13) is used to obtain the pose and acceleration of the robot. The joint angle sensor (14) is used to obtain the current angles of the various joints of the robot. The force and torque sensor (15) is used to measure the force and torque received by the joints.

4. The control system of a humanoid robot according to claim 2, characterized in that: The perception module (2) includes a local feature extraction unit (21), a global feature extraction unit (22), a feature fusion unit (23), and a target detection and recognition unit (24). The local feature extraction unit (21) is used to extract local features of the image and output a feature map. The global feature extraction unit (22) takes the feature map as input and outputs the global features of the image. The feature fusion unit (23) is used to fuse the global features and local features of the image and output fused features. The target detection and recognition unit (24) takes the fused features as input and outputs environmental feature information.

5. The control system of a humanoid robot according to claim 2, characterized in that: The motion control module (5) includes a high-level motion planning unit (51) and a low-level motion control unit (52). The high-level motion planning unit (51) is used to convert the abstract action instruction into a specific motion trajectory. The low-level motion control unit (52) is used to generate the joint control instructions of the robot according to the specific motion trajectory.

6. A control method for a humanoid robot, including Step 1, data acquisition; Step 2, environmental perception; Step 3, state representation; Step 4, action selection; Step 5, motion control; Step 6, feedback and adjustment; characterized in that: In the above Step 1, the sensor module (1) collects information about the environment around the robot and the state of the robot itself, and sends the information about the environment around the robot to the perception module (2), and sends the information about the state of the robot itself to the state representation module (3); In the above step 2, the perception module (2) processes the environmental information around the robot, extracts the environmental feature information, and sends the environmental feature information to the state representation module (3); In the above step 3, the state representation module (3) fuses the environmental feature information, the robot's own state information, and the task objective to form the state representation of the robot, and sends it to the decision-making module (4); In the above step 4, the decision-making module (4) selects the optimal action based on the state representation using the policy network trained by the reinforcement learning algorithm, and sends it to the motion control module (5); In the above step 5, the motion control module (5) converts the selected action into joint control instructions and sends them to the actuator module (6). The actuator module (6) executes the joint control instructions to drive the joints of the robot to move and achieve the desired motion trajectory; In the above step 6, the feedback module (7) obtains the actual motion result and environmental change information of the robot, compares them with the desired motion result, and then feeds back the error information to the decision-making module (4). The decision-making module (4) optimizes the control strategy based on the error information.

7. A control method for a humanoid robot according to claim 6, characterized in that: In the above step 2, the perception module (2) uses a ViT combined with CNN hybrid model to process the environmental information around the robot. Specifically: the local feature extraction unit (21) uses the CNN algorithm to extract the local features of the image, outputs the feature map to the global feature extraction unit (22) and the feature fusion unit (23). The global feature extraction unit (22) uses the ViT algorithm to divide the feature map into multiple patches, then maps these patches into vectors, and inputs them into the Transformer encoder. The Transformer encoder uses the self-attention mechanism to process these vectors to obtain the global feature representation of the image, and sends it to the feature fusion unit (23). The feature fusion unit (23) fuses the global feature and local feature of the image, outputs the fused feature to the target detection and recognition unit (24), and the target detection and recognition unit (24) outputs the environmental feature information.

8. A control method for a humanoid robot according to claim 7, characterized in that: The local feature extraction unit (21) uses the ResNet50 model and is pre-trained using the ImageNet dataset. The global feature extraction unit (22) uses the ViT model, the patch size is 16, the number of encoder layers is 12, the hidden size is 768, the feature fusion unit (23) uses the concatenate fusion model, and the feature fusion unit (23) uses the Softmax classifier.

9. A control method for a humanoid robot according to claim 6, characterized in that: In the above step 3, the environmental feature information includes the position information of objects, obstacles, and targets, and the robot's own state information includes the position, attitude, speed, joint angles, and joint angular velocities of the robot.

10. A control method for a humanoid robot according to claim 6, characterized in that: In the above step 4, the reinforcement learning algorithm used by the decision-making module (4) is the improved DQN algorithm.

Images