Humanoid robot control method and device, electronic equipment and storage medium

Abstract

The application belongs to the technical field of robot control, and discloses a kind of humanoid robot control method, device, electronic equipment and storage medium, the method includes: obtaining the first sensor data and target task information of humanoid robot, based on the first sensor data, combined with the preset instruction action set, symbol field construction is carried out, and target neural symbol field data is obtained, according to target neural symbol field data and preset instruction action set information, directed acyclic task graph data is generated, with directed acyclic task graph data as the basis, based on target neural symbol field data, target task information is carried out motion planning, and the target control trajectory information and target control quantity information of humanoid robot are obtained, and humanoid robot is controlled according to target control trajectory information and target control quantity information;Through the above method, the control precision of humanoid robot is improved.

Description

Technical Field

[0001] This application relates to the technical field of robot control, and more specifically, to a humanoid robot control method, device, electronic device, and storage medium. Background Technology

[0002] With the deepening of the digital transformation of the power system, the operation and maintenance mode of critical infrastructure such as substations is gradually evolving from manual on-site operations to remote intelligent operations. Currently, although inspection and operation robots have been initially applied in substations and other scenarios, the mainstream still relies on track-mounted inspection platforms or remote-controlled operating arms, capable only of performing non-contact tasks such as taking photos and measuring temperatures, or simple switching actions. For complex operations requiring coordinated arm movements, force perception, and precise trajectory control, such as secondary cable insertion / removal and pressure plate tightening, existing systems heavily depend on visual feedback, lacking real-time transmission of tactile and force information. This results in insufficient "presence" for remote operators, making it difficult to perceive the magnitude and direction of contact forces, which can easily lead to equipment damage or operational failure. Furthermore, substations and other similar scenarios present strong electromagnetic interference, varying lighting conditions, and unstructured work surfaces, further exacerbating the uncertainty of perception and control.

[0003] Current mainstream humanoid robot learning methods, such as deep reinforcement learning and dynamic motion primitives, have fundamental limitations in addressing the aforementioned challenges. Deep reinforcement learning strategies generate actions end-to-end, but lack interpretability and may fail unexpectedly when encountering scenarios outside the training distribution, and cannot learn structured lessons from failures. While dynamic motion primitives can generalize trajectory shapes from a single teaching iteration, their generalization ability is limited to geometric shapes and cannot transfer physical interaction patterns. Traditional control methods are designed in Cartesian space, ignoring the inherent curvature and anisotropy of the task surface, leading to decreased motion accuracy and energy efficiency under constrained environments. These problems collectively make it difficult for existing humanoid robots to achieve high-precision, high-reliability control when performing fine operations in complex, unstructured environments, severely restricting the widespread application of humanoid robots in fields such as industrial operation and maintenance.

[0004] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0005] The purpose of this application is to provide a humanoid robot control method, device, electronic device, and storage medium. By using directed acyclic task graph data as a basis and target neural symbol field data, motion planning is performed on the target task information to obtain the target control trajectory information and target control quantity information of the humanoid robot, so as to control the humanoid robot. This solves the problem that existing humanoid robot control methods are difficult to perform high-precision and high-reliability control operations in complex and unstructured environments. It can deeply integrate environmental perception, task understanding, and motion planning, thereby improving the control accuracy of the humanoid robot.

[0006] Firstly, this application provides a humanoid robot control method, including: Acquire first sensor data and target task information of the humanoid robot; Based on the data from the first sensor, and combined with a preset set of command actions, a symbol field is constructed to obtain target neural symbol field data; Based on the target neural symbol field data and the preset instruction action set information, a directed acyclic task graph data of the target neural symbol field data is generated; Based on the directed acyclic task graph data and the target neural symbol field data, motion planning is performed on the target task information to obtain the target control trajectory information of the humanoid robot and the target control quantity information of the target control trajectory information. The humanoid robot is controlled according to the target control trajectory information and the target control quantity information.

[0007] The humanoid robot control method provided in this application can control humanoid robots. By using directed acyclic task graph data as a basis and target neural symbol field data, motion planning is performed on the target task information to obtain the target control trajectory information and target control quantity information of the humanoid robot, thereby controlling the humanoid robot. This solves the problem that existing humanoid robot control methods are difficult to perform high-precision and high-reliability control operations in complex and unstructured environments. It can deeply integrate environmental perception, task understanding and motion planning, thereby improving the control accuracy of humanoid robots.

[0008] Optionally, based on the first sensor data and combined with a preset set of command actions, a symbol field is constructed to obtain target neural symbol field data, including: Based on the data from the first sensor, and combined with a preset set of command actions, a symbol field is constructed to obtain the first initial neural symbol field data. Based on the first initial neural symbol field data, a preset Gaussian distribution function is used to calculate the first confidence level of the first initial neural symbol field data. When the first confidence level is greater than or equal to a preset unconfidence threshold, the first initial neural symbol field data corresponding to the first confidence level is determined as the target neural symbol field data.

[0009] Optionally, after calculating the first confidence level of the first initial neural symbol field data using a preset Gaussian distribution function based on the first initial neural symbol field data, the method further includes: When the first confidence level is less than the preset unconfidence threshold, the current pose of the humanoid robot is adjusted, and a second target neural symbol field data is constructed based on the second sensor data of the humanoid robot after the pose adjustment, so as to determine the target neural symbol field data based on the second target neural symbol field data.

[0010] Optionally, the current pose of the humanoid robot is adjusted, and second target neural symbol field data is constructed based on the second sensor data of the humanoid robot after the pose adjustment, so as to determine the target neural symbol field data based on the second target neural symbol field data, including: Based on the first initial neural symbol field data, the variational free energy of the humanoid robot when performing each candidate sensory action is calculated; the candidate sensory action is the sensory action in the preset candidate sensory action data. Based on the variational free energy, target adjustment operation information is determined from the preset candidate sensing action data; Acquire the second sensor data of the humanoid robot; the second sensor data is the sensor data of the humanoid robot after performing the candidate perception action corresponding to the target adjustment operation information. Construct second target neural symbol field data corresponding to the second sensor data, and calculate based on the second sensor data to obtain the second confidence level of the second target neural symbol field data; When the second confidence level is greater than or equal to the preset unbelievable threshold, the second target neural symbol field data corresponding to the second confidence level is determined as the target neural symbol field data; When the second confidence level is less than a preset unbelievable threshold, new target adjustment operation information is re-determined based on the second target neural symbol field data corresponding to the second confidence level. After the humanoid robot executes the candidate perception action corresponding to the new target adjustment operation information, third target neural symbol field data is constructed so that the third confidence level corresponding to the third target neural symbol field data is greater than or equal to the preset unbelievable threshold, and the third target neural symbol field data is determined as the target neural symbol field data.

[0011] The humanoid robot control method provided in this application can control the humanoid robot and select the optimal pose adjustment action through variational free energy, so that the robot can efficiently acquire high-quality perception data and further improve the accuracy and efficiency of neural symbol field construction.

[0012] Optionally, based on the target neural symbol field data and the preset set of instruction actions, directed acyclic task graph data of the target neural symbol field data is generated, including: Extract each instruction action information from the preset instruction action set information; Based on the target neural symbol field data, determine the task execution completion state transition conditions between each of the instruction action information; Using each instruction action information as a node and the task completion state transition condition as an edge, a directed acyclic task graph data of the target neural symbol field data is generated.

[0013] Optionally, based on the directed acyclic task graph data and the target neural symbol field data, motion planning is performed on the target task information to obtain the target control trajectory information of the humanoid robot and the target control quantity information of the target control trajectory information, including: Based on the directed acyclic task graph data, and using the target neural symbol field data and the target task information, the Riemannian geometric motion planning method is employed to generate the reference trajectory on the Riemannian manifold through integration, thereby obtaining the initial control trajectory information. Based on the initial control trajectory information, a constrained manifold control algorithm is used to calculate the desired acceleration information and desired control torque information of the initial control trajectory information. The model predictive path integral control algorithm is used to optimize the initial control trajectory information, the expected acceleration information and the expected control torque information of the initial control trajectory information to obtain the target control trajectory information and the target control quantity information of the humanoid robot.

[0014] The humanoid robot control method provided in this application can control humanoid robots. By combining Riemannian geometric motion planning with constrained manifold control and model predictive path integral control, it fully considers the inherent geometric characteristics and constraints of the task surface and achieves high-precision and high-efficiency motion planning.

[0015] Optionally, after controlling the humanoid robot according to the target control trajectory information and the target control quantity information, the method further includes: When the humanoid robot fails to execute the target task information, it uses counterfactual reasoning to calculate trajectory correction parameters based on the target neural symbol field data. These trajectory correction parameters are used by the humanoid robot to correct the generated initial control trajectory information in the next generation.

[0016] Secondly, this application provides a humanoid robot control device, comprising: The acquisition module is used to acquire the first sensor data and target task information of the humanoid robot; The construction module is used to construct a symbol field based on the data from the first sensor and combined with a preset set of instruction actions to obtain target neural symbol field data; The generation module is used to generate directed acyclic task graph data of the target neural symbol field data based on the target neural symbol field data and the preset instruction action set information; The planning module is used to perform motion planning on the target task information based on the directed acyclic task graph data and the target neural symbol field data to obtain the target control trajectory information of the humanoid robot and the target control quantity information of the target control trajectory information. The control module is used to control the humanoid robot according to the target control trajectory information and the target control quantity information.

[0017] This humanoid robot control device uses directed acyclic task graph data and target neural symbol field data to perform motion planning on target task information, thereby obtaining target control trajectory information and target control quantity information for the humanoid robot. This addresses the problem that existing humanoid robot control methods struggle to perform high-precision and high-reliability control operations in complex and unstructured environments. It deeply integrates environmental perception, task understanding, and motion planning, improving the control accuracy of the humanoid robot.

[0018] Thirdly, this application provides an electronic device including a processor and a memory, wherein the memory stores a computer program executable by the processor, and when the processor executes the computer program, it performs the steps in the humanoid robot control method described above.

[0019] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the humanoid robot control method described above.

[0020] Beneficial effects: The humanoid robot control method, device, electronic device, and storage medium provided in this application, based on directed acyclic task graph data and target neural symbol field data, perform motion planning on target task information to obtain target control trajectory information and target control quantity information of the humanoid robot, thereby controlling the humanoid robot. This solves the problem that existing humanoid robot control methods are difficult to perform high-precision and high-reliability control operations in complex and unstructured environments. It can deeply integrate environmental perception, task understanding, and motion planning, thereby improving the control accuracy of the humanoid robot. Attached Figure Description

[0021] Figure 1 A flowchart of a humanoid robot control method provided in an embodiment of this application.

[0022] Figure 2 This is a schematic diagram of the humanoid robot control device provided in the embodiments of this application.

[0023] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0024] Labeling Explanation: 1. Acquisition Module; 2. Construction Module; 3. Generation Module; 4. Planning Module; 5. Control Module; 301. Processor; 302. Memory; 303. Communication Bus. Detailed Implementation

[0025] 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 a part of the embodiments of this application, and not all of the embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0026] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0027] Please refer to Figure 1 , Figure 1 This application discloses a humanoid robot control method in some embodiments, used to control a humanoid robot, comprising the following steps: Step S101: Obtain the first sensor data and target task information of the humanoid robot; Step S102: Based on the data from the first sensor and combined with the preset set of command actions, a symbol field is constructed to obtain the target neural symbol field data; Step S103: Generate directed acyclic task graph data of the target neural symbol field data based on the target neural symbol field data and the preset instruction action set information; Step S104: Based on the directed acyclic task graph data and the target neural symbol field data, motion planning is performed on the target task information to obtain the target control trajectory information and the target control quantity information of the humanoid robot. Step S105: Control the humanoid robot according to the target control trajectory information and target control quantity information.

[0028] This humanoid robot control method uses directed acyclic task graph data and target neural symbol field data to perform motion planning on target task information, thereby obtaining the target control trajectory information and target control quantity information of the humanoid robot. This method addresses the problem that existing humanoid robot control methods struggle to perform high-precision and high-reliability control operations in complex and unstructured environments. It deeply integrates environmental perception, task understanding, and motion planning, thus improving the control accuracy of the humanoid robot.

[0029] Specifically, in step S101, the first sensor data and target task information of the humanoid robot are acquired. The first sensor data is sensor data acquired through various sensors installed on the humanoid robot, including image data of the surrounding environment and its own state information. The image data of the surrounding environment is image data of the humanoid robot within a certain range, containing spatial feature information of the humanoid robot's surrounding environment. The self-state information refers to various state information describing the humanoid robot's body, such as joint angles, joint torques, end effector pose, tactile sensor data, force sensor data, inertial measurement unit (IMU) data, etc., which collectively describe the current kinematics, dynamics, and interaction state of the humanoid robot with the environment. The target task information refers to the specific task description that the humanoid robot needs to complete, such as "tightening a screw" or "inserting a cable," which includes the target state of the task, constraints, and the expected method of completion.

[0030] Specifically, in step S102, based on the first sensor data and combined with a preset set of command actions, a symbol field is constructed to obtain target neural symbol field data, including: Based on the data from the first sensor, combined with the preset set of command actions, a symbol field is constructed to obtain the first initial neural symbol field data; Based on the first initial neural symbol field data, a preset Gaussian distribution function is used to calculate the first confidence level of the first initial neural symbol field data. When the first confidence level is greater than or equal to the preset unconfidence threshold, the first initial neural symbol field data corresponding to the first confidence level is determined as the target neural symbol field data.

[0031] In step S102, after initially integrating the first sensor data (such as environmental feature information and self-state information) perceived by the humanoid robot and the preset set of command actions, the continuous perceptual information is fused with discrete symbolic knowledge to form a preliminary understanding of the current environment and task state, generating a preliminary, unverified neural symbolic representation. The preset set of command actions refers to a predefined series of basic actions or operational primitives. The humanoid robot can combine these actions to complete complex tasks, such as atomic operations like "grasping," "moving," and "rotating."

[0032] The neural symbol field data F (including target neural symbol field data, first initial neural symbol field data, subsequent second initial neural symbol field data, and third initial neural symbol field data) includes the operational symbol field. Geometric symbolic field and physical symbol field .

[0033] Operation symbol field This represents any point in a given scene, and outputs the degree (membership) at which that point is suitable for performing a specific operation (such as "press", "rotate", "grip"). The mathematical expression for the operation symbol field is as follows: ; in, Let be the symbolic field of operations, representing the suitability of performing operation i at position point (three-dimensional spatial coordinates) x. , indicating the operation symbol field The suitability of performing operation i on any point x in the matrix is ​​a real number between 0 and 1. It is a three-dimensional space; The function represents the S-shaped growth curve; K represents the total number of bit support vectors. This is the k-th support vector; The weights of the k-th support vector; Let be the class label of the k-th support vector with respect to operation i, indicating whether the k-th support vector is suitable for performing operation i. ,or ,when When, it means that the k-th support vector is not suitable for performing operation i. When, it means that the k-th support vector is suitable for performing operation i; For Gaussian kernel functions (Radial Basis Function kernel, RBF kernel); This refers to the operation of extracting deep features from environmental feature data using a CNN encoder; This represents the position x corresponding to the k-th support vector; To extract depth feature data for any location point x from environmental feature data; To extract the depth feature data of the k-th support vector at any location point x from environmental feature data; Representing deep feature data With deep feature data The result of the Gaussian kernel function calculation between; b is the bias term in the SVM classifier, also known as the intercept.

[0034] Geometric symbol field This represents the local geometric properties of operation i at any point x, including geometric features related to operation i such as normal vector, curvature, and contact mode. Geometric symbol field. The specific mathematical expression is as follows: ; in, Let x be a geometric symbol field, representing the local geometric properties at any point x relative to operation i. , indicating that the local geometric property of operation i at any point x is a d-dimensional real vector. It is a d-dimensional real vector; It is a multilayer perceptron network; The location code corresponding to location point x; To obtain the output result, an arbitrary location point x and its corresponding location code are input into a multilayer perceptron network.

[0035] Physical symbol field The physical interaction parameters at location x include parameters such as local stiffness, damping, and friction cone, representing the physical properties required for contact with location x when performing operation i. Physical symbol field The specific mathematical expression is as follows: ; in, Let be a physical symbol field, representing the physical properties required for operation i to come into contact with position point x; The function is the Gaussian process regression function. Let x be the half-angle of the friction cone at position x, used to describe the coefficient of friction at position x; Location point ; Let position point x and position point The Gaussian kernel function calculation results between them.

[0036] In step S102, a preset Gaussian distribution function is used to quantitatively evaluate the reliability or accuracy of the neural symbol field data (first initial neural symbol field data) to obtain the confidence level of the neural symbol field data (first confidence level). The confidence level can assess the consistency between the neural symbol field data and the original perceptual data, the completeness of the symbol representation, and the degree of conflict between different information sources, providing an objective basis for subsequent decision-making to determine whether the currently constructed neural symbol field data is sufficient to support reliable motion planning.

[0037] The preset Gaussian distribution function is specifically as follows: ; in, To determine the confidence level of neural symbol field data F, environmental state variables from the neural symbol field data F can be input. (Mainly includes geometric symbol fields such as object pose and contact state) and physical symbol field Calculations are performed using the data in the data. It is a Gaussian distribution function; The mean; Let be the covariance matrix.

[0038] The system determines whether the first confidence level is greater than or equal to a preset unconfidence threshold to quantify the gap between the environmental and state information reflected by the current initial neural symbol field data and the minimum reliability required by the system. The preset unconfidence threshold is a pre-defined standard used to determine whether the neural symbol field data is sufficiently reliable for subsequent use; it can be set according to actual needs, such as by obtaining it experimentally.

[0039] When the first confidence level is greater than or equal to the preset unconfidence threshold, the initial perception result (first initial neural symbol field data) is considered to be reliable enough. The first initial neural symbol field data corresponding to the first confidence level is determined as the target neural symbol field data, so that the first initial neural symbol field data can be directly used for subsequent processing to improve efficiency.

[0040] Specifically, in step S102, after calculating the first confidence level of the first initial neural symbol field data using a preset Gaussian distribution function based on the first initial neural symbol field data, the following steps are also included: When the first confidence level is less than the preset unconfidence threshold, the current pose of the humanoid robot is adjusted, and the second target neural symbol field data is constructed based on the second sensor data of the humanoid robot after the pose adjustment, so as to determine the target neural symbol field data based on the second target neural symbol field data.

[0041] In step S102, when the first confidence level is less than a preset unconfidence threshold, it indicates that the reliability of the first initial neural symbol field data is insufficient. By actively changing the humanoid robot's perception perspective or position, new environmental and state data are acquired, thereby constructing new neural symbol field data. Its confidence level is continuously evaluated until the requirement is met, thus determining the obtained target neural symbol field data. "Adjusting the current pose of the humanoid robot" involves having the humanoid robot perform a series of preset perception actions, such as head rotation, body tilting forward, or lateral movement, to obtain more comprehensive environmental information or eliminate occlusion. In this way, the humanoid robot can acquire new second sensor data, which is used to construct the second target neural symbol field data. Subsequently, the second confidence level of the second target neural symbol field data is calculated and compared again with the unconfidence threshold until the confidence level after pose adjustment is greater than or equal to the unconfidence threshold, thereby ensuring that the determined target neural symbol field data is reliable.

[0042] Specifically, in step S102, the current pose of the humanoid robot is adjusted, and based on the second sensor data of the humanoid robot after the pose adjustment, second target neural symbol field data is constructed to determine target neural symbol field data, including: Based on the initial neural symbol field data, the variational free energy of the humanoid robot when performing each candidate sensory action is calculated; the candidate sensory actions are the sensory actions in the preset candidate sensory action data. Based on variational free energy, target adjustment operation information is determined from preset candidate sensory action data; Acquire the second sensor data of the humanoid robot; the second sensor data is the sensor data of the humanoid robot after performing the candidate perception action corresponding to the target adjustment operation information. Construct second target neural symbol field data corresponding to the second sensor data, and calculate the second confidence level of the second target neural symbol field data based on the second sensor data; When the second confidence level is greater than or equal to the preset unconfidence threshold, the second target neural symbol field data corresponding to the second confidence level is determined as the target neural symbol field data; When the second confidence level is less than the preset unbelievable threshold, new target adjustment operation information is re-determined based on the second target neural symbol field data corresponding to the second confidence level. After the humanoid robot executes the candidate perception action corresponding to the new target adjustment operation information, third target neural symbol field data is constructed so that the third confidence level corresponding to the third target neural symbol field data is greater than or equal to the preset unbelievable threshold, and the third target neural symbol field data is determined as the target neural symbol field data.

[0043] In step S102, based on the first initial neural symbol field data, the variational free energy of the humanoid robot when performing each candidate sensory action from the preset candidate sensory action data is calculated. This is to evaluate the potential effect of each potential sensory action on reducing environmental uncertainty and improving the confidence of the neural symbol field data, as well as to measure the difference between prediction and actual observation and the complexity of the model. The smaller the value of the variational free energy, the more accurate the understanding of the environment and the lower the uncertainty. The preset candidate sensory action data may include various pose adjustment actions, such as head rotation, body translation, arm extension, etc., which are designed to acquire environmental information from different angles or distances.

[0044] The specific formula for calculating the variational free energy is as follows: ; in, Candidate sensory actions The variational free energy; For humanoid robots to respond to environmental state variables The approximate posterior distribution; For the approximate posterior distribution The mathematical expectation; For humanoid robots in environmental state variables Execute candidate sensing actions under the condition The subsequent observation matrix (likelihood) is provided by the neural symbol field; The observation matrix obtained after the action is performed (i.e., environmental feature data and multi-source state data after the candidate sensing action is performed). It is the relative entropy, also known as the Kullback-Leibler divergence (KL divergence), used to measure the difference between two distributions; Humanoid robots' response to environmental state variables The prior distribution; For humanoid robots in environmental state variables Execute candidate sensing actions under the condition Approximate distribution after; For candidate sensory actions Prior action preferences (encoding of candidate perceived actions) (innate tendency) For humanoid robots to respond to environmental state variables Approximate posterior distribution Compared with humanoid robots in response to environmental state variables The relative entropy between prior distributions; For humanoid robots in environmental state variables Execute candidate sensing actions under the condition Approximate distribution after Regarding candidate sensory actions The relative entropy between prior action preferences; This indicates the expected surprise, used to encourage the selection of actions with a high probability of observation, i.e., reducing uncertainty; This represents a complex cost term for confidence levels, used to penalize deviations between the current confidence level and the prior confidence level. This represents the cost of action complexity, and the deviation between the distribution of penalized actions and prior action preferences.

[0045] The minimum value of the variational free energy is determined, and the candidate sensing action corresponding to the minimum value is selected from the preset candidate sensing action data as the current optimal adjustment operation. This proactively selects one or more sensing actions that can minimize its uncertainty to the environment. The specific pose adjustment command corresponding to the current optimal adjustment operation is determined as the target adjustment operation information.

[0046] After the humanoid robot executes the candidate sensory action corresponding to the selected target adjustment operation information, it will acquire new sensor data, namely second sensor data. This new sensor data will be used to update the humanoid robot's understanding of the environment.

[0047] The neural symbol field data, i.e. the second target neural symbol field data, is recalculated based on the new sensing data (second sensor data), and the confidence of the second target neural symbol field data is reassessed, i.e. the second confidence (the specific calculation method is as shown in the confidence calculation steps above).

[0048] If, after a pose adjustment, the credibility of the newly constructed neural symbol field data reaches the preset standard (greater than or equal to the unbelievability threshold), then the data is considered reliable and can be directly used as the final target neural symbol field data.

[0049] If a single adjustment fails to obtain sufficiently reliable neural symbol field data, the robot will use the currently unreliable neural symbol field data and, following the calculation steps described above, redetermine new target adjustment operation information. After the humanoid robot performs the new adjustment action, it will re-evaluate whether the new confidence level is greater than or equal to the unreliability threshold. This forms an iterative loop until the humanoid robot, through a series of intelligent pose adjustments, finally obtains neural symbol field data with a confidence level greater than or equal to the unreliability threshold, and then identifies this neural symbol field data as the target neural symbol field data.

[0050] Suppose a humanoid robot is given the task of "grabbing a water glass on a table". In the initial perception stage, due to the water glass being partially obscured by other objects or poor lighting conditions, the first sensor data acquired by the humanoid robot is insufficient to construct a first initial neural symbol field data with high confidence (i.e., the first confidence of the first initial neural symbol field data is greater than or equal to the unconfidence threshold).

[0051] At this point, the humanoid robot enters the pose adjustment phase. Specifically, the humanoid robot calculates the variational free energy that might be generated after executing a series of preset candidate sensory actions (e.g., turning the head 15 degrees to the left, turning the head 15 degrees to the right, moving the body forward 10 centimeters, moving the body backward 10 centimeters, etc.). Assuming the calculation results show that the variational free energy corresponding to the action of "turning the head 15 degrees to the left" is the smallest, this means that this action is most likely to reveal complete information about the water cup and reduce environmental uncertainty. Therefore, the humanoid robot identifies "turning the head 15 degrees to the left" as the target adjustment operation information and executes it.

[0052] After performing this action, the humanoid robot reacquires data from the second sensor and constructs a second target neural symbol field based on this new data (which can be updated using an extended Kalman filter). It then calculates its second confidence level. If the second confidence level is greater than or equal to a preset unconfidence threshold, it indicates that sufficiently reliable target neural symbol field data has been obtained, and this data will be determined as the final target neural symbol field data for subsequent task planning. Conversely, if the second confidence level is still insufficient, the humanoid robot will recalculate the variational free energy of new candidate perceptual actions based on its current state and repeat the above process until it obtains neural symbol field data that meets the requirements. Through this iterative optimization, the humanoid robot can efficiently find the optimal perceptual pose, ensuring that its understanding of the environment is accurate and reliable.

[0053] Specifically, in step S103, directed acyclic task graph data of the target neural symbol field is generated based on the target neural symbol field data and the preset set of instruction actions, including: Extract individual instruction / action information from the preset set of instruction / action information; Based on the target neural symbol field data, determine the task execution completion state transition conditions between each instruction action information; Using each instruction action information as a node and the task completion state transition condition as an edge, a directed acyclic task graph data of the target neural symbol field data is generated.

[0054] In step S103, predefined atomic-level actions (atomic-level operations, such as grasping, placing, and moving) that the humanoid robot can perform are identified and separated from a set or library. These instruction action information are the basic units that constitute complex tasks, and their extraction is the basis for constructing a task graph.

[0055] By leveraging the rich environmental perception and state understanding capabilities provided by target neural symbol field data, we can determine which conditions will be met and which subsequent instructions / actions can be triggered after the execution of a command / action. For example, after a "grab" action is completed, its state transition conditions might include "the object has been grasped"; while the prerequisites for a "place" action might include "the hand is holding an object" and "the target location is free." These conditions define the logical dependencies and execution order between command / action information.

[0056] Using each instruction / action information as a node and the task completion state transition conditions as edges, a directed acyclic task graph (DAG) of the target neural symbol field data is generated. A DAG is a special graph structure where edges are directional and contain no cycles. Specifically, each extracted instruction / action information is considered a node in the graph, and the task completion state transition conditions between these instructions / actions are considered directed edges connecting these nodes. These directed edges represent the order and logical dependencies of task execution, ensuring the reasonable decomposition and orderly execution of the task.

[0057] Thus, a complex task is broken down into a series of subtasks with a clear logical order and dependencies, allowing the overall structure and execution flow of the task to be clearly presented. This structured representation provides a solid foundation for subsequent motion planning.

[0058] Specifically, in step S104, based on the directed acyclic task graph data and the target neural symbol field data, motion planning is performed on the target task information to obtain the target control trajectory information and target control quantity information of the humanoid robot, including: Based on the directed acyclic task graph data, and using the target neural symbol field data and target task information, the Riemann geometric motion planning method is adopted to generate the reference trajectory on the Riemann manifold through integration, thereby obtaining the initial control trajectory information. Based on the initial control trajectory information, the constrained manifold control algorithm is used to calculate and obtain the desired acceleration information and desired control torque information of the initial control trajectory information. The model predictive path integral control algorithm is used to optimize the initial control trajectory information, the expected acceleration information and the expected control torque information of the initial control trajectory information to obtain the target control trajectory information and the target control quantity information of the humanoid robot.

[0059] In step S104, based on the directed acyclic task graph data and utilizing the properties of Riemannian geometry, an initial path satisfying the task objective and environmental constraints is generated through an integration process within the humanoid robot's configuration space or a more generalized state space, thus obtaining initial control trajectory information. Riemannian manifolds can naturally represent the nonlinear kinematics and dynamics of the humanoid robot, making the generated trajectory geometrically more reasonable and continuous. The initial control trajectory information is the direct output of this integration process, providing a foundation for subsequent refined control.

[0060] The initial control trajectory information can be obtained through the generation of geometric dynamic motion primitives. The specific calculation formula for the geometric dynamic motion primitives is as follows: ; ; in, This is a time constant used to control the duration of motion; Let M be the position of the end effector of the humanoid robot on the Riemannian manifold (i.e., the system state). For tangential space velocity, that is, at position The velocity vector at that location; For tangent space velocity The first derivative with respect to time; For position The first derivative with respect to time; It is the first positive gain constant. It is the second positive gain constant. and Used to determine the linear part of the system; The target state is the attractor provided by the neural symbol field. , For the arg function, Indicates that the operation symbol field is made The set of positions that achieves the maximum value; This is a non-linear forcing term used to encode the trajectory shape, which can be learned from teaching data. , Let be the j-th Gaussian kernel function with respect to the phase variable s. This represents the total number of Gaussian kernel functions. Here, is the learnable weight parameter corresponding to the j-th Gaussian kernel function, and s is the phase variable, i.e., the phase, which decays from 1 to 0 over time and is used to modulate the nonlinear forcing term. Changes over time; For the field attractor coupling term, , This is the coupling strength coefficient; The third positive gain constant controls the rate at which the phase variable s decays from its initial value of 1 to 0.

[0061] By using geometric dynamic motion primitives and existing numerical integration methods (such as fourth-order Runge-Kutta), a reference trajectory is generated to obtain the initial control trajectory information. and the initial control speed information corresponding to the initial control trajectory information. .

[0062] The motion of a humanoid robot is controlled on a Riemannian manifold M, while considering its physical constraints. The constrained manifold control algorithm projects the humanoid robot's dynamic equations onto the Riemannian manifold, thereby calculating the desired acceleration and desired control torque information required to achieve the initial control trajectory while satisfying the manifold constraints. This ensures that the generated control quantities are physically realizable and can effectively drive the humanoid robot to move along the reference trajectory.

[0063] In practical applications, the current end effector pose of each end effector of a humanoid robot can be obtained in real time. and current speed Based on the current end pose and current speed Calculate the geometric errors of pose and velocity separately. The specific formula for calculating the geometric error is as follows: ; ; in, This refers to the pose geometry error; This is the front end pose; For velocity geometric error; For initial control speed information, i.e. ; From the front end pose To the initial control trajectory information Parallel transport corresponding to the initial target location.

[0064] The desired acceleration information is calculated based on the geometric error using a preset formula for calculating desired acceleration. The specific formula for calculating desired acceleration is as follows: ; in, For desired acceleration information; For initial control trajectory information The corresponding initial control acceleration; This is the proportional gain coefficient matrix; This is the differential gain coefficient matrix; For Christofel's notation, used to compensate for manifold curvature; The desired control torque information is calculated using a preset formula. The specific formula for calculating the desired control torque is as follows: ; in, For the desired control torque information; The Jacobian matrix for humanoid robots is used to map joint velocities to end-effector velocities. This is the Riemann metric matrix, used to determine the orientation sensitivity of the task space; For the Coriolis and centripetal force matrices; This is the frictional torque, used for torque compensation; The external contact torque is estimated by the force sensor.

[0065] Utilizing the framework of model predictive control and incorporating the concept of path integral, the initially generated trajectory and control variable are iteratively optimized to obtain target control trajectory information and target control variable information. Specifically, the model predictive path integral control algorithm evaluates multiple randomly sampled paths in the prediction time domain and selects the optimal path based on a cost function. This generates smoother, more robust, and optimal target control trajectory and target control variable information, taking into account the humanoid robot's dynamics model and future state predictions. This optimization process effectively handles nonlinear systems and uncertainties, improving the execution accuracy and stability of the trajectory.

[0066] In practical applications, an initial path replanning is performed at a preset update time (which can be set according to actual needs) to obtain multiple reference trajectories (i.e., initial control trajectory information). The running cost of each reference trajectory is then calculated. The specific formula for calculating the running cost is as follows: ; in, The cost of running the w-th reference trajectory; This is the w-th reference trajectory; This represents the penalty cost at the end of the timeframe; Represents the state cost term; Indicates the control cost item; is the weighting coefficient (regularization parameter) of the control cost, used to balance tracking error and control energy; r is the weight matrix of the control cost; Preset update time The optimal control quantity, i.e., the target control quantity information, is calculated using a preset target control quantity calculation formula. The specific target control quantity calculation formula is as follows: ; in, For target control quantity information; This is an inverse temperature parameter used to control the smoothness of the weighted softmax. The larger the value, the more uniform the weight distribution. The smaller the value, the more likely it is to choose the trajectory with the lowest cost; K is the total number of reference trajectories; For the first Reference trajectory; For the first The cost of running a reference trajectory; The control sequence (including acceleration or control torque) corresponding to the w-th reference trajectory.

[0067] Specifically, in step S105, the humanoid robot is controlled according to the target control trajectory information and target control quantity information. For example, the target control quantity information (such as control torque or acceleration) can be directly sent to the lower-level controller of the humanoid robot to drive each joint of the humanoid robot to move according to the planned trajectory. Furthermore, during the movement, the humanoid robot continuously acquires new environmental feature data and multi-source state data, and updates the target neural symbol field data based on this real-time data. Simultaneously, the aforementioned Riemannian geometric motion planning method can be used to periodically optimize the target control trajectory information and target control quantity information, thereby achieving closed-loop control, adaptation to environmental changes, and real-time optimization of the target control trajectory.

[0068] Specifically, after controlling the humanoid robot according to the target control trajectory information and target control quantity information, the process also includes: When the humanoid robot fails to execute the target task information, the trajectory correction parameters are calculated based on the target neural symbol field data using the counterfactual reasoning method. The trajectory correction parameters are used by the humanoid robot to correct the generated initial control trajectory information in the next generation.

[0069] After a humanoid robot performs a target task, if the result is deemed a failure, a counterfactual reasoning method is activated for adjustment. Counterfactual reasoning is a technique that learns and improves its decision-making process by simulating "what would have happened if different actions had been taken." It can analyze the root causes of task failure and infer alternative actions or adjustment strategies to avoid failure in the same or similar situations. The trajectory correction parameters are a key output calculated by the counterfactual reasoning method. Specifically, they can be adjustments to the initial control trajectory information in certain dimensions, such as a set of values ​​for fine-tuning waypoints, velocity, acceleration, or joint torques. These parameters aim to guide the humanoid robot to generate more optimized initial control trajectory information when encountering similar tasks or environments in the future, thereby improving the success rate of the task. The trajectory correction parameters can be a vector, a matrix, or a function, depending on the output of the counterfactual reasoning model and the need for trajectory correction.

[0070] In practical applications, based on the principles of counterfactual reasoning, a structural causal model can be pre-constructed. Specifically, the structural causal model is as follows: ; ; ; in, Insertion depth (an internal variable of the structural causal model, which can be manipulated); A causal relationship related to insertion depth; The pose of the target object; The gripping force of the humanoid robot's hand during the insertion action; To describe the causal relationship between insertion depth and target object pose and gripping force, it represents the theoretically achievable insertion depth under given target object pose and gripper gripping force. These are random factors that affect the insertion depth but are not observed (i.e., external variables of the structural causal model, which cannot be interfered with), such as uneven friction at the contact surface, minute deformation, sensor noise, etc. The lateral force (an internal variable in the structural causal model, which can be manipulated); For causal relationships related to lateral forces, To describe the causal relationship between the lateral force and the target object's pose and insertion depth, it represents the theoretically achievable lateral force given the target object's pose and insertion depth. These are random factors that affect lateral forces but are not observed (i.e., external variables of the structural causal model, which cannot be interfered with). This serves as a marker of task success (an internal variable of the structural causal model, which can be manipulated). An indicator function that outputs 1 when the condition inside the parentheses is true, and 0 otherwise; The insertion depth threshold; This is the lateral force threshold.

[0071] When the task fails, a three-step reasoning mechanism is triggered: (1) Extrapolation mechanism: Based on observed facts (e.g., insertion depth of only 0.3 cm), calculate the external variables of the structural causal model (e.g., random factors). The posterior distribution of ).

[0072] (2) Intervention mechanism: By changing the action variable, such as in the structural causal model Adjusted to (For example, changing from 10N to 15N), delete in the structural causal model The relevant description, and replace it with Related descriptions.

[0073] (3) Prediction mechanism: New result parameters (insertion depth, lateral force and mission success indicator) are calculated in the new structural causal model.

[0074] By repeatedly executing the three-step reasoning mechanism, new task success indicators are repeatedly obtained. exist Under the given conditions, the action variable that minimizes the number of variables is denoted as the counterfactual optimal action. The updated structural causal model can then be used for subsequent counterfactual reasoning.

[0075] After determining the counterfactual optimal action, the correction influence parameters of the trajectory correction parameters are defined, and the state vector at time t is defined. ,in, The contact force at time t (measured by a six-dimensional force sensor). The insertion depth at time t (e.g., the distance a cable is inserted into a socket). Define the action at time t. , Let be the clamping force applied to the gripper at time t. Define the counterfactual optimal action. , The action variable is the one that minimizes the variable. Define the value estimation function. It is obtained by online fitting using a small neural network and is used to evaluate the quality of the next state after the action is performed.

[0076] Taking into account the influence parameters of the above correction terms, the correction terms are calculated. The specific formula for calculating the correction terms is as follows: ; in, For trajectory correction parameters, set the trajectory correction parameters Input nonlinear forced term In this process, it is used by humanoid robots to generate initial control trajectory information in the next iteration, and to correct the generated initial control trajectory information; Let be the mathematical expectation of the posterior distribution of the external variables in the structural causal model; Given an observation matrix (i.e., environmental feature data and multi-source state data), this represents the posterior distribution of the external variables in a structural causal model. For external variables in structural causal models; For the specific values ​​of the external variables in the external variables of the structural causal model; Indicates from the posterior distribution An external variable that samples a specific value; Let be the gradient of the value function with respect to actions, representing the direction and sensitivity of the impact of changes in actions on long-term returns.

[0077] Track correction parameters Input nonlinear forced term In the middle, we get: ; in, This is the corrected nonlinear forcing term; The original nonlinear forcing term is the term before correction. , The weighting coefficients for trajectory correction parameters.

[0078] As can be seen from the above, this humanoid robot control method acquires the first sensor data and target task information of the humanoid robot. Based on the first sensor data and combined with the preset set of command actions, a symbol field is constructed to obtain target neural symbol field data. According to the target neural symbol field data and the preset set of command actions, a directed acyclic task graph data of the target neural symbol field data is generated. Based on the directed acyclic task graph data and the target neural symbol field data, motion planning is performed on the target task information to obtain the target control trajectory information and target control quantity information of the humanoid robot. The humanoid robot is controlled according to the target control trajectory information and target control quantity information. Thus, by using the directed acyclic task graph data as a basis and the target neural symbol field data to perform motion planning on the target task information, the target control trajectory information and target control quantity information of the humanoid robot are obtained, thereby controlling the humanoid robot. This solves the problem that existing humanoid robot control methods are difficult to perform high-precision and high-reliability control operations in complex and unstructured environments. It can deeply integrate environmental perception, task understanding and motion planning, and improve the control accuracy of humanoid robots.

[0079] refer to Figure 2This application provides a humanoid robot control device for controlling a humanoid robot, comprising: Module 1 is used to acquire the first sensor data and target task information of the humanoid robot; Construction module 2 is used to construct a symbol field based on the data from the first sensor and combined with a preset set of command actions to obtain target neural symbol field data; The generation module 3 is used to generate directed acyclic task graph data of the target neural symbol field data based on the target neural symbol field data and the preset instruction action set information. Planning module 4 is used to perform motion planning on the target task information based on the directed acyclic task graph data and the target neural symbol field data, to obtain the target control trajectory information and target control variable information of the humanoid robot. Control module 5 is used to control the humanoid robot according to the target control trajectory information and target control quantity information.

[0080] This humanoid robot control device uses directed acyclic task graph data and target neural symbol field data to perform motion planning on target task information, thereby obtaining target control trajectory information and target control quantity information for the humanoid robot. This addresses the problem that existing humanoid robot control methods struggle to perform high-precision and high-reliability control operations in complex and unstructured environments. It deeply integrates environmental perception, task understanding, and motion planning, improving the control accuracy of the humanoid robot.

[0081] Specifically, during execution, module 1 acquires the first sensor data and target task information of the humanoid robot. The first sensor data is sensor data acquired through various sensors installed on the humanoid robot, including image data of the surrounding environment and its own state information. The image data of the surrounding environment is image data of the humanoid robot within a certain range, containing spatial feature information of the surrounding environment. The self-state information refers to various state information describing the humanoid robot's body, such as joint angles, joint torques, end effector pose, tactile sensor data, force sensor data, inertial measurement unit (IMU) data, etc., which collectively describe the current kinematics, dynamics, and interaction state with the environment of the humanoid robot. The target task information refers to the specific task description that the humanoid robot needs to complete, such as "tightening a screw" or "inserting a cable," which includes the target state of the task, constraints, and the expected method of completion.

[0082] Specifically, module 2 constructs a symbol field based on the data from the first sensor and combined with a preset set of command actions. When the target neural symbol field data is obtained, the following is executed: Based on the data from the first sensor, combined with the preset set of command actions, a symbol field is constructed to obtain the first initial neural symbol field data; Based on the first initial neural symbol field data, a preset Gaussian distribution function is used to calculate the first confidence level of the first initial neural symbol field data. When the first confidence level is greater than or equal to the preset unconfidence threshold, the first initial neural symbol field data corresponding to the first confidence level is determined as the target neural symbol field data.

[0083] When module 2 is executed, after initially integrating the first sensor data perceived by the humanoid robot (such as environmental feature information and its own state information) and the preset set of command actions, it fuses continuous perceptual information with discrete symbolic knowledge to form a preliminary understanding of the current environment and task state, generating a preliminary, unverified neural symbolic representation. The preset set of command actions refers to a predefined series of basic actions or operational primitives. The humanoid robot can combine these actions to complete complex tasks, such as atomic operations like "grasping," "moving," and "rotating."

[0084] The neural symbol field data F (including target neural symbol field data, first initial neural symbol field data, subsequent second initial neural symbol field data, and third initial neural symbol field data) includes the operational symbol field. Geometric symbolic field and physical symbol field .

[0085] Operation symbol field This represents any point in a given scene, and outputs the degree (membership) at which that point is suitable for performing a specific operation (such as "press", "rotate", "grip"). The mathematical expression for the operation symbol field is as follows: ; in, Let be the symbolic field of operations, representing the suitability of performing operation i at position point (three-dimensional spatial coordinates) x. , indicating the operation symbol field The suitability of performing operation i on any point x in the matrix is ​​a real number between 0 and 1. It is a three-dimensional space; The function represents the S-shaped growth curve; K represents the total number of bit support vectors. This is the k-th support vector; The weights of the k-th support vector; Let be the class label of the k-th support vector with respect to operation i, indicating whether the k-th support vector is suitable for performing operation i. ,or ,when When, it means that the k-th support vector is not suitable for performing operation i. When, it means that the k-th support vector is suitable for performing operation i; For Gaussian kernel functions (Radial Basis Function kernel, RBF kernel); This refers to the operation of extracting deep features from environmental feature data using a CNN encoder; This represents the position x corresponding to the k-th support vector; To extract depth feature data for any location point x from environmental feature data; To extract the depth feature data of the k-th support vector at any location point x from environmental feature data; Representing deep feature data With deep feature data The result of the Gaussian kernel function calculation between; b is the bias term in the SVM classifier, also known as the intercept.

[0086] Geometric symbol field This represents the local geometric properties of operation i at any point x, including geometric features related to operation i such as normal vector, curvature, and contact mode. Geometric symbol field. The specific mathematical expression is as follows: ; in, Let x be a geometric symbol field, representing the local geometric properties at any point x relative to operation i. , indicating that the local geometric property of operation i at any point x is a d-dimensional real vector. It is a d-dimensional real vector; It is a multilayer perceptron network; The location code corresponding to location point x; To obtain the output result, an arbitrary location point x and its corresponding location code are input into a multilayer perceptron network.

[0087] Physical symbol field The physical interaction parameters at location x include parameters such as local stiffness, damping, and friction cone, representing the physical properties required for contact with location x when performing operation i. Physical symbol field The specific mathematical expression is as follows: ; in, Let be a physical symbol field, representing the physical properties required for operation i to come into contact with position point x; The function is the Gaussian process regression function. Let x be the half-angle of the friction cone at position x, used to describe the coefficient of friction at position x; Location point ; Let position point x and position point The Gaussian kernel function calculation results between them.

[0088] During execution, module 2 employs a preset Gaussian distribution function to quantitatively evaluate the reliability or accuracy of the neural symbol field data (first initial neural symbol field data), obtaining the confidence level of the neural symbol field data (first confidence level). The confidence level assesses the consistency between the neural symbol field data and the original perceptual data, the completeness of the symbol representation, and the degree of conflict between different information sources, providing an objective basis for subsequent decision-making to determine whether the currently constructed neural symbol field data is sufficient to support reliable motion planning.

[0089] The preset Gaussian distribution function is specifically as follows: ; in, To determine the confidence level of neural symbol field data F, environmental state variables from the neural symbol field data F can be input. (Mainly includes geometric symbol fields such as object pose and contact state) and physical symbol field Calculations are performed using the data in the data. It is a Gaussian distribution function; The mean; Let be the covariance matrix.

[0090] The system determines whether the first confidence level is greater than or equal to a preset unconfidence threshold to quantify the gap between the environmental and state information reflected by the current initial neural symbol field data and the minimum reliability required by the system. The preset unconfidence threshold is a pre-defined standard used to determine whether the neural symbol field data is sufficiently reliable for subsequent use; it can be set according to actual needs, such as by obtaining it experimentally.

[0091] When the first confidence level is greater than or equal to the preset unconfidence threshold, the initial perception result (first initial neural symbol field data) is considered to be reliable enough. The first initial neural symbol field data corresponding to the first confidence level is determined as the target neural symbol field data, so that the first initial neural symbol field data can be directly used for subsequent processing to improve efficiency.

[0092] Specifically, after the construction module 2 calculates the first confidence level of the first initial neural symbol field data using a preset Gaussian distribution function based on the first initial neural symbol field data, it also performs the following: When the first confidence level is less than the preset unconfidence threshold, the current pose of the humanoid robot is adjusted, and the second target neural symbol field data is constructed based on the second sensor data of the humanoid robot after the pose adjustment, so as to determine the target neural symbol field data based on the second target neural symbol field data.

[0093] When module 2 is executed, if the first confidence level is less than a preset unconfidence threshold, it indicates that the reliability of the first initial neural symbol field data is insufficient. By actively changing the humanoid robot's perception perspective or position, new environmental and state data are acquired, thereby constructing new neural symbol field data. Its confidence level is continuously evaluated until the requirement is met, thus determining the obtained target neural symbol field data. "Adjusting the current pose of the humanoid robot" involves causing the humanoid robot to perform a series of preset perception actions, such as head rotation, body tilting forward, or lateral movement, to obtain more comprehensive environmental information or eliminate occlusion. In this way, the humanoid robot can acquire new second sensor data, which is used to construct the second target neural symbol field data. Subsequently, the second confidence level of the second target neural symbol field data is calculated and compared again with the unconfidence threshold until the confidence level after pose adjustment is greater than or equal to the unconfidence threshold, thereby ensuring that the determined target neural symbol field data is reliable.

[0094] Specifically, when module 2 adjusts the current pose of the humanoid robot and constructs second target neural symbol field data based on the second sensor data after the pose adjustment, and determines the target neural symbol field data based on the second target neural symbol field data, it executes the following: Based on the initial neural symbol field data, the variational free energy of the humanoid robot when performing each candidate sensory action is calculated; the candidate sensory actions are the sensory actions in the preset candidate sensory action data. Based on variational free energy, target adjustment operation information is determined from preset candidate sensory action data; Acquire the second sensor data of the humanoid robot; the second sensor data is the sensor data of the humanoid robot after performing the candidate perception action corresponding to the target adjustment operation information. Construct second target neural symbol field data corresponding to the second sensor data, and calculate the second confidence level of the second target neural symbol field data based on the second sensor data; When the second confidence level is greater than or equal to the preset unconfidence threshold, the second target neural symbol field data corresponding to the second confidence level is determined as the target neural symbol field data; When the second confidence level is less than the preset unbelievable threshold, new target adjustment operation information is re-determined based on the second target neural symbol field data corresponding to the second confidence level. After the humanoid robot executes the candidate perception action corresponding to the new target adjustment operation information, third target neural symbol field data is constructed so that the third confidence level corresponding to the third target neural symbol field data is greater than or equal to the preset unbelievable threshold, and the third target neural symbol field data is determined as the target neural symbol field data.

[0095] During execution, module 2 calculates the variational free energy of the humanoid robot when performing each candidate sensory action from the preset candidate sensory action data, based on the first initial neural symbol field data. This is to evaluate the potential effect of each potential sensory action on reducing environmental uncertainty and improving the confidence of the neural symbol field data, as well as to measure the difference between prediction and actual observation and the complexity of the model. The smaller the value of the variational free energy, the more accurate the understanding of the environment and the lower the uncertainty. The preset candidate sensory action data can include various pose adjustment actions, such as head rotation, body translation, and arm extension, designed to acquire environmental information from different angles or distances.

[0096] The specific formula for calculating the variational free energy is as follows: ; in, Candidate sensory actions The variational free energy; For humanoid robots to respond to environmental state variables The approximate posterior distribution; For the approximate posterior distribution The mathematical expectation; For humanoid robots in environmental state variables Execute candidate sensing actions under the condition The subsequent observation matrix (likelihood) is provided by the neural symbol field; The observation matrix obtained after the action is performed (i.e., environmental feature data and multi-source state data after the candidate sensing action is performed). It is the relative entropy, also known as the Kullback-Leibler divergence (KL divergence), used to measure the difference between two distributions; Humanoid robots' response to environmental state variables The prior distribution; For humanoid robots in environmental state variables Execute candidate sensing actions under the condition Approximate distribution after; For candidate sensory actions Prior action preferences (encoding of candidate perceived actions) (innate tendency) For humanoid robots to respond to environmental state variables Approximate posterior distribution Compared with humanoid robots in response to environmental state variables The relative entropy between prior distributions; For humanoid robots in environmental state variables Execute candidate sensing actions under the condition Approximate distribution after Regarding candidate sensory actions The relative entropy between prior action preferences; This indicates the expected surprise, used to encourage the selection of actions with a high probability of observation, i.e., reducing uncertainty; This represents a complex cost term for confidence levels, used to penalize deviations between the current confidence level and the prior confidence level. This represents the cost of action complexity, and the deviation between the distribution of penalized actions and prior action preferences.

[0097] The minimum value of the variational free energy is determined, and the candidate sensing action corresponding to the minimum value is selected from the preset candidate sensing action data as the current optimal adjustment operation. This proactively selects one or more sensing actions that can minimize its uncertainty to the environment. The specific pose adjustment command corresponding to the current optimal adjustment operation is determined as the target adjustment operation information.

[0098] After the humanoid robot executes the candidate sensory action corresponding to the selected target adjustment operation information, it will acquire new sensor data, namely second sensor data. This new sensor data will be used to update the humanoid robot's understanding of the environment.

[0099] The neural symbol field data, i.e. the second target neural symbol field data, is recalculated based on the new sensing data (second sensor data), and the confidence of the second target neural symbol field data is reassessed, i.e. the second confidence (the specific calculation method is as shown in the confidence calculation steps above).

[0100] If, after a pose adjustment, the credibility of the newly constructed neural symbol field data reaches the preset standard (greater than or equal to the unbelievability threshold), then the data is considered reliable and can be directly used as the final target neural symbol field data.

[0101] If a single adjustment fails to obtain sufficiently reliable neural symbol field data, the robot will use the currently unreliable neural symbol field data and, following the calculation steps described above, redetermine new target adjustment operation information. After the humanoid robot performs the new adjustment action, it will re-evaluate whether the new confidence level is greater than or equal to the unreliability threshold. This forms an iterative loop until the humanoid robot, through a series of intelligent pose adjustments, finally obtains neural symbol field data with a confidence level greater than or equal to the unreliability threshold, and then identifies this neural symbol field data as the target neural symbol field data.

[0102] Suppose a humanoid robot is given the task of "grabbing a water glass on a table". In the initial perception stage, due to the water glass being partially obscured by other objects or poor lighting conditions, the first sensor data acquired by the humanoid robot is insufficient to construct a first initial neural symbol field data with high confidence (i.e., the first confidence of the first initial neural symbol field data is greater than or equal to the unconfidence threshold).

[0103] At this point, the humanoid robot enters the pose adjustment phase. Specifically, the humanoid robot calculates the variational free energy that might be generated after executing a series of preset candidate sensory actions (e.g., turning the head 15 degrees to the left, turning the head 15 degrees to the right, moving the body forward 10 centimeters, moving the body backward 10 centimeters, etc.). Assuming the calculation results show that the variational free energy corresponding to the action of "turning the head 15 degrees to the left" is the smallest, this means that this action is most likely to reveal complete information about the water cup and reduce environmental uncertainty. Therefore, the humanoid robot identifies "turning the head 15 degrees to the left" as the target adjustment operation information and executes it.

[0104] After performing this action, the humanoid robot reacquires data from the second sensor and constructs a second target neural symbol field based on this new data (which can be updated using an extended Kalman filter). It then calculates its second confidence level. If the second confidence level is greater than or equal to a preset unconfidence threshold, it indicates that sufficiently reliable target neural symbol field data has been obtained, and this data will be determined as the final target neural symbol field data for subsequent task planning. Conversely, if the second confidence level is still insufficient, the humanoid robot will recalculate the variational free energy of new candidate perceptual actions based on its current state and repeat the above process until it obtains neural symbol field data that meets the requirements. Through this iterative optimization, the humanoid robot can efficiently find the optimal perceptual pose, ensuring that its understanding of the environment is accurate and reliable.

[0105] Specifically, when generating directed acyclic task graph data of the target neural symbol field based on the target neural symbol field data and the preset set of instruction actions, generation module 3 executes: Extract individual instruction / action information from the preset set of instruction / action information; Based on the target neural symbol field data, determine the task execution completion state transition conditions between each instruction action information; Using each instruction action information as a node and the task completion state transition condition as an edge, a directed acyclic task graph data of the target neural symbol field data is generated.

[0106] During execution, generation module 3 identifies and separates predefined atomic-level actions (atomic-level operations, such as grasping, placing, and moving) that can be performed by the humanoid robot from a set or library. These instruction action information are the basic units that constitute complex tasks, and their extraction is the foundation for constructing task graphs.

[0107] By leveraging the rich environmental perception and state understanding capabilities provided by target neural symbol field data, we can determine which conditions will be met and which subsequent instructions / actions can be triggered after the execution of a command / action. For example, after a "grab" action is completed, its state transition conditions might include "the object has been grasped"; while the prerequisites for a "place" action might include "the hand is holding an object" and "the target location is free." These conditions define the logical dependencies and execution order between command / action information.

[0108] Using each instruction / action information as a node and the task completion state transition conditions as edges, a directed acyclic task graph (DAG) of the target neural symbol field data is generated. A DAG is a special graph structure where edges are directional and contain no cycles. Specifically, each extracted instruction / action information is considered a node in the graph, and the task completion state transition conditions between these instructions / actions are considered directed edges connecting these nodes. These directed edges represent the order and logical dependencies of task execution, ensuring the reasonable decomposition and orderly execution of the task.

[0109] Thus, a complex task is broken down into a series of subtasks with a clear logical order and dependencies, allowing the overall structure and execution flow of the task to be clearly presented. This structured representation provides a solid foundation for subsequent motion planning.

[0110] Specifically, when planning module 4 performs motion planning on the target task information based on the directed acyclic task graph data and the target neural symbol field data to obtain the target control trajectory information and the target control quantity information of the humanoid robot, it executes: Based on the directed acyclic task graph data, and using the target neural symbol field data and target task information, the Riemann geometric motion planning method is adopted to generate the reference trajectory on the Riemann manifold through integration, thereby obtaining the initial control trajectory information. Based on the initial control trajectory information, the constrained manifold control algorithm is used to calculate and obtain the desired acceleration information and desired control torque information of the initial control trajectory information. The model predictive path integral control algorithm is used to optimize the initial control trajectory information, the expected acceleration information and the expected control torque information of the initial control trajectory information to obtain the target control trajectory information and the target control quantity information of the humanoid robot.

[0111] During execution, planning module 4, based on the directed acyclic task graph data, utilizes the characteristics of Riemannian geometry to generate an initial path that satisfies the task objective and environmental constraints through an integration process within the humanoid robot's configuration space or a more generalized state space, thus obtaining initial control trajectory information. Riemannian manifolds naturally represent the nonlinear kinematics and dynamics of the humanoid robot, making the generated trajectory geometrically more reasonable and continuous. The initial control trajectory information is the direct output of this integration process, providing a foundation for subsequent refined control.

[0112] The initial control trajectory information can be obtained through the generation of geometric dynamic motion primitives. The specific calculation formula for the geometric dynamic motion primitives is as follows: ; ; in, This is a time constant used to control the duration of motion; Let M be the position of the end effector of the humanoid robot on the Riemannian manifold (i.e., the system state). For tangential space velocity, that is, at position The velocity vector at that location; For tangent space velocity The first derivative with respect to time; For position The first derivative with respect to time; It is the first positive gain constant. It is the second positive gain constant. and Used to determine the linear part of the system; The target state is the attractor provided by the neural symbol field. , For the arg function, Indicates that the operation symbol field is made The set of positions that achieves the maximum value; This is a non-linear forcing term used to encode the trajectory shape, which can be learned from teaching data. , Let be the j-th Gaussian kernel function with respect to the phase variable s. This represents the total number of Gaussian kernel functions. Here, is the learnable weight parameter corresponding to the j-th Gaussian kernel function, and s is the phase variable, i.e., the phase, which decays from 1 to 0 over time and is used to modulate the nonlinear forcing term. Changes over time; For the field attractor coupling term, , This is the coupling strength coefficient; The third positive gain constant controls the rate at which the phase variable s decays from its initial value of 1 to 0.

[0113] By using geometric dynamic motion primitives and existing numerical integration methods (such as fourth-order Runge-Kutta), a reference trajectory is generated to obtain the initial control trajectory information. and the initial control speed information corresponding to the initial control trajectory information. .

[0114] The motion of a humanoid robot is controlled on a Riemannian manifold M, while considering its physical constraints. The constrained manifold control algorithm projects the humanoid robot's dynamic equations onto the Riemannian manifold, thereby calculating the desired acceleration and desired control torque information required to achieve the initial control trajectory while satisfying the manifold constraints. This ensures that the generated control quantities are physically realizable and can effectively drive the humanoid robot to move along the reference trajectory.

[0115] In practical applications, the current end effector pose of each end effector of a humanoid robot can be obtained in real time. and current speed Based on the current end pose and current speed Calculate the geometric errors of pose and velocity separately. The specific formula for calculating the geometric error is as follows: ; ; in, This refers to the pose geometry error; This is the front end pose; For velocity geometric error; For initial control speed information, i.e. ; From the front end pose To the initial control trajectory information Parallel transport corresponding to the initial target location.

[0116] The desired acceleration information is calculated based on the geometric error using a preset formula for calculating desired acceleration. The specific formula for calculating desired acceleration is as follows: ; in, For desired acceleration information; For initial control trajectory information The corresponding initial control acceleration; This is the proportional gain coefficient matrix; This is the differential gain coefficient matrix; For Christofel's notation, used to compensate for manifold curvature; The desired control torque information is calculated using a preset formula. The specific formula for calculating the desired control torque is as follows: ; in, For the desired control torque information; The Jacobian matrix for humanoid robots is used to map joint velocities to end-effector velocities. This is the Riemann metric matrix, used to determine the orientation sensitivity of the task space; For the Coriolis and centripetal force matrices; This is the frictional torque, used for torque compensation; The external contact torque is estimated by the force sensor.

[0117] Utilizing the framework of model predictive control and incorporating the concept of path integral, the initially generated trajectory and control variable are iteratively optimized to obtain target control trajectory information and target control variable information. Specifically, the model predictive path integral control algorithm evaluates multiple randomly sampled paths in the prediction time domain and selects the optimal path based on a cost function. This generates smoother, more robust, and optimal target control trajectory and target control variable information, taking into account the humanoid robot's dynamics model and future state predictions. This optimization process effectively handles nonlinear systems and uncertainties, improving the execution accuracy and stability of the trajectory.

[0118] In practical applications, an initial path replanning is performed at a preset update time (which can be set according to actual needs) to obtain multiple reference trajectories (i.e., initial control trajectory information). The running cost of each reference trajectory is then calculated. The specific formula for calculating the running cost is as follows: ; in, The cost of running the w-th reference trajectory; This is the w-th reference trajectory; This represents the penalty cost at the end of the timeframe; Represents the state cost term; Indicates the control cost item; is the weighting coefficient (regularization parameter) of the control cost, used to balance tracking error and control energy; r is the weight matrix of the control cost; Preset update time The optimal control quantity, i.e., the target control quantity information, is calculated using a preset target control quantity calculation formula. The specific target control quantity calculation formula is as follows: ; in, For target control quantity information; This is an inverse temperature parameter used to control the smoothness of the weighted softmax. The larger the value, the more uniform the weight distribution. The smaller the value, the more likely it is to choose the trajectory with the lowest cost; K is the total number of reference trajectories; For the first Reference trajectory; For the first The cost of running a reference trajectory; The control sequence (including acceleration or control torque) corresponding to the w-th reference trajectory.

[0119] Specifically, during execution, control module 5 controls the humanoid robot according to the target control trajectory information and target control quantity information. For example, the target control quantity information (such as control torque or acceleration) can be directly sent to the humanoid robot's underlying controller, driving each joint of the humanoid robot to move according to the planned trajectory. Furthermore, during the movement, the humanoid robot continuously acquires new environmental feature data and multi-source state data, and updates the target neural symbol field data based on this real-time data. Simultaneously, the aforementioned Riemannian geometric motion planning method can be used to periodically optimize the target control trajectory information and target control quantity information, thereby achieving closed-loop control, adaptation to environmental changes, and real-time optimization of the target control trajectory.

[0120] Specifically, the humanoid robot control device also includes: The correction module is used to calculate trajectory correction parameters based on the target neural symbol field data when the humanoid robot fails to execute the target task information. The trajectory correction parameters are used to correct the initial control trajectory information generated by the humanoid robot in the next generation.

[0121] After a humanoid robot performs a target task, if the result is deemed a failure, the correction module initiates a counterfactual reasoning method for adjustment. Counterfactual reasoning is a technique that learns and improves its decision-making process by simulating "what would have happened if different actions had been taken." It can analyze the root causes of task failure and infer alternative actions or adjustment strategies to avoid failure in the same or similar situations. The trajectory correction parameters are a key output calculated by the counterfactual reasoning method. Specifically, they can be adjustments to the initial control trajectory information in certain dimensions, such as a set of values ​​for fine-tuning waypoints, velocity, acceleration, or joint torques. These parameters aim to guide the humanoid robot to generate more optimized initial control trajectory information when encountering similar tasks or environments in the future, thereby improving the success rate of the task. The trajectory correction parameters can be a vector, a matrix, or a function, depending on the output of the counterfactual reasoning model and the need for trajectory correction.

[0122] In practical applications, based on the principles of counterfactual reasoning, a structural causal model can be pre-constructed. Specifically, the structural causal model is as follows: ; ; ; in, Insertion depth (an internal variable of the structural causal model, which can be manipulated); A causal relationship related to insertion depth; The pose of the target object; The gripping force of the humanoid robot's hand during the insertion action; To describe the causal relationship between insertion depth and target object pose and gripping force, it represents the theoretically achievable insertion depth under given target object pose and gripper gripping force. These are random factors that affect the insertion depth but are not observed (i.e., external variables of the structural causal model, which cannot be interfered with), such as uneven friction at the contact surface, minute deformation, sensor noise, etc. The lateral force (an internal variable in the structural causal model, which can be manipulated); For causal relationships related to lateral forces, To describe the causal relationship between the lateral force and the target object's pose and insertion depth, it represents the theoretically achievable lateral force given the target object's pose and insertion depth. These are random factors that affect lateral forces but are not observed (i.e., external variables of the structural causal model, which cannot be interfered with). This serves as a marker of task success (an internal variable of the structural causal model, which can be manipulated). An indicator function that outputs 1 when the condition inside the parentheses is true, and 0 otherwise; The insertion depth threshold; This is the lateral force threshold.

[0123] When the task fails, a three-step reasoning mechanism is triggered: (1) Extrapolation mechanism: Based on observed facts (e.g., insertion depth of only 0.3 cm), calculate the external variables of the structural causal model (e.g., random factors). The posterior distribution of ).

[0124] (2) Intervention mechanism: By changing the action variable, such as in the structural causal model Adjusted to (For example, changing from 10N to 15N), delete in the structural causal model The relevant description, and replace it with Related descriptions.

[0125] (3) Prediction mechanism: New result parameters (insertion depth, lateral force and mission success indicator) are calculated in the new structural causal model.

[0126] By repeatedly executing the three-step reasoning mechanism, new task success indicators are repeatedly obtained. exist Under the given conditions, the action variable that minimizes the number of variables is denoted as the counterfactual optimal action. The updated structural causal model can then be used for subsequent counterfactual reasoning.

[0127] After determining the counterfactual optimal action, the correction influence parameters of the trajectory correction parameters are defined, and the state vector at time t is defined. ,in, The contact force at time t (measured by a six-dimensional force sensor). The insertion depth at time t (e.g., the distance a cable is inserted into a socket). Define the action at time t. , Let be the clamping force applied to the gripper at time t. Define the counterfactual optimal action. , The action variable is the one that minimizes the variable. Define the value estimation function. It is obtained by online fitting using a small neural network and is used to evaluate the quality of the next state after the action is performed.

[0128] Taking into account the influence parameters of the above correction terms, the correction terms are calculated. The specific formula for calculating the correction terms is as follows: ; in, For trajectory correction parameters, set the trajectory correction parameters Input nonlinear forced term In this process, it is used by humanoid robots to generate initial control trajectory information in the next iteration, and to correct the generated initial control trajectory information; Let be the mathematical expectation of the posterior distribution of the external variables in the structural causal model; Given an observation matrix (i.e., environmental feature data and multi-source state data), this represents the posterior distribution of the external variables in a structural causal model. For external variables in structural causal models; For the specific values ​​of the external variables in the external variables of the structural causal model; Indicates from the posterior distribution An external variable that samples a specific value; Let be the gradient of the value function with respect to actions, representing the direction and sensitivity of the impact of changes in actions on long-term returns.

[0129] Track correction parameters Input nonlinear forced term In the middle, we get: ; in, This is the corrected nonlinear forcing term; The original nonlinear forcing term is the term before correction. , The weighting coefficients for trajectory correction parameters.

[0130] As can be seen from the above, this humanoid robot control device acquires the first sensor data and target task information of the humanoid robot. Based on the first sensor data and combined with the preset set of command actions, it constructs a symbol field to obtain target neural symbol field data. According to the target neural symbol field data and the preset set of command actions, it generates directed acyclic task graph data of the target neural symbol field data. Based on the directed acyclic task graph data and the target neural symbol field data, it performs motion planning on the target task information to obtain the target control trajectory information and target control quantity information of the humanoid robot. According to the target control trajectory information and target control quantity information, it controls the humanoid robot. Thus, by using the directed acyclic task graph data as a basis and the target neural symbol field data to perform motion planning on the target task information, it obtains the target control trajectory information and target control quantity information of the humanoid robot to control the humanoid robot. This solves the problem that existing humanoid robot control methods are difficult to perform high-precision and high-reliability control operations in complex and unstructured environments. It can deeply integrate environmental perception, task understanding and motion planning, and improve the control accuracy of the humanoid robot.

[0131] Please refer to Figure 3 , Figure 3This application provides a schematic diagram of the structure of an electronic device, including a processor 301 and a memory 302. The processor 301 and the memory 302 are interconnected and communicate with each other through a communication bus 303 and / or other forms of connection mechanism (not shown). The memory 302 stores a computer program executable by the processor 301. When the electronic device is running, the processor 301 executes the computer program to execute the humanoid robot control method in any optional implementation of the above embodiments, so as to achieve the following functions: acquiring first sensor data and target task information of the humanoid robot; constructing a symbol field based on the first sensor data and combined with a preset set of instruction action information to obtain target neural symbol field data; generating directed acyclic task graph data of the target neural symbol field data according to the target neural symbol field data and the preset set of instruction action information; performing motion planning on the target task information based on the directed acyclic task graph data and the target neural symbol field data to obtain target control trajectory information and target control quantity information of the humanoid robot; and controlling the humanoid robot according to the target control trajectory information and target control quantity information.

[0132] This application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it executes the humanoid robot control method in any optional implementation of the above embodiments to achieve the following functions: acquiring first sensor data and target task information of the humanoid robot; constructing a symbol field based on the first sensor data and combined with a preset set of instruction actions to obtain target neural symbol field data; generating directed acyclic task graph data of the target neural symbol field data based on the target neural symbol field data and the preset set of instruction actions; performing motion planning on the target task information based on the directed acyclic task graph data and the target neural symbol field data to obtain target control trajectory information and target control quantity information of the humanoid robot; and controlling the humanoid robot according to the target control trajectory information and target control quantity information. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0133] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0134] Furthermore, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0135] Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0136] In this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, without necessarily requiring or implying any such actual relationship or order between these entities or operations.

[0137] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A humanoid robot control method, characterized in that, include: Acquire first sensor data and target task information of the humanoid robot; Based on the data from the first sensor, and combined with a preset set of command actions, a symbol field is constructed to obtain target neural symbol field data; Based on the target neural symbol field data and the preset instruction action set information, a directed acyclic task graph data of the target neural symbol field data is generated; Based on the directed acyclic task graph data and the target neural symbol field data, motion planning is performed on the target task information to obtain the target control trajectory information of the humanoid robot and the target control quantity information of the target control trajectory information. The humanoid robot is controlled according to the target control trajectory information and the target control quantity information.

2. The humanoid robot control method according to claim 1, characterized in that, Based on the data from the first sensor, and combined with a preset set of command actions, a symbol field is constructed to obtain target neural symbol field data, including: Based on the first sensor data, combined with a preset set of command actions, a symbol field is constructed to obtain the first initial neural symbol field data; Based on the first initial neural symbol field data, a preset Gaussian distribution function is used to calculate the first confidence level of the first initial neural symbol field data. When the first confidence level is greater than or equal to a preset unconfidence threshold, the first initial neural symbol field data corresponding to the first confidence level is determined as the target neural symbol field data.

3. The humanoid robot control method according to claim 2, characterized in that, Based on the first initial neural symbol field data, after calculating the first confidence level of the first initial neural symbol field data using a preset Gaussian distribution function, the method further includes: When the first confidence level is less than the preset unconfidence threshold, the current pose of the humanoid robot is adjusted, and a second target neural symbol field data is constructed based on the second sensor data of the humanoid robot after the pose adjustment, so as to determine the target neural symbol field data based on the second target neural symbol field data.

4. The humanoid robot control method according to claim 3, characterized in that, Adjusting the current pose of the humanoid robot, and constructing second target neural symbol field data based on the second sensor data of the humanoid robot after pose adjustment, to determine target neural symbol field data based on the second target neural symbol field data, including: Based on the first initial neural symbol field data, the variational free energy of the humanoid robot when performing each candidate sensory action is calculated; the candidate sensory action is the sensory action in the preset candidate sensory action data. Based on the variational free energy, target adjustment operation information is determined from the preset candidate sensing action data; Acquire the second sensor data of the humanoid robot; the second sensor data is the sensor data of the humanoid robot after performing the candidate perception action corresponding to the target adjustment operation information. Construct second target neural symbol field data corresponding to the second sensor data, and calculate based on the second sensor data to obtain the second confidence level of the second target neural symbol field data; When the second confidence level is greater than or equal to the preset unbelievable threshold, the second target neural symbol field data corresponding to the second confidence level is determined as the target neural symbol field data; When the second confidence level is less than a preset unbelievable threshold, new target adjustment operation information is re-determined based on the second target neural symbol field data corresponding to the second confidence level. After the humanoid robot executes the candidate perception action corresponding to the new target adjustment operation information, third target neural symbol field data is constructed so that the third confidence level corresponding to the third target neural symbol field data is greater than or equal to the preset unbelievable threshold, and the third target neural symbol field data is determined as the target neural symbol field data.

5. The humanoid robot control method according to claim 1, characterized in that, Based on the target neural symbol field data and the preset set of instruction actions, a directed acyclic task graph data of the target neural symbol field data is generated, including: Extract each instruction action information from the preset instruction action set information; Based on the target neural symbol field data, determine the task execution completion state transition conditions between each of the instruction action information; Using each instruction action information as a node and the task completion state transition condition as an edge, a directed acyclic task graph data of the target neural symbol field data is generated.

6. The humanoid robot control method according to claim 1, characterized in that, Based on the directed acyclic task graph data and the target neural symbol field data, motion planning is performed on the target task information to obtain the target control trajectory information of the humanoid robot and the target control quantity information of the target control trajectory information, including: Based on the directed acyclic task graph data, and using the target neural symbol field data and the target task information, the Riemannian geometric motion planning method is employed to generate the reference trajectory on the Riemannian manifold through integration, thereby obtaining the initial control trajectory information. Based on the initial control trajectory information, a constrained manifold control algorithm is used to calculate the desired acceleration information and desired control torque information of the initial control trajectory information. The model predictive path integral control algorithm is used to optimize the initial control trajectory information, the expected acceleration information and the expected control torque information of the initial control trajectory information to obtain the target control trajectory information and the target control quantity information of the humanoid robot.

7. The humanoid robot control method according to claim 1, characterized in that, After controlling the humanoid robot according to the target control trajectory information and the target control quantity information, the process further includes: When the humanoid robot fails to execute the target task information, it uses counterfactual reasoning to calculate trajectory correction parameters based on the target neural symbol field data. These trajectory correction parameters are used by the humanoid robot to correct the generated initial control trajectory information in the next generation.

8. A humanoid robot control device, characterized in that, include: The acquisition module is used to acquire the first sensor data and target task information of the humanoid robot; The construction module is used to construct a symbol field based on the data from the first sensor and combined with a preset set of instruction actions to obtain target neural symbol field data; The generation module is used to generate directed acyclic task graph data of the target neural symbol field data based on the target neural symbol field data and the preset instruction action set information; The planning module is used to perform motion planning on the target task information based on the directed acyclic task graph data and the target neural symbol field data to obtain the target control trajectory information of the humanoid robot and the target control quantity information of the target control trajectory information. The control module is used to control the humanoid robot according to the target control trajectory information and the target control quantity information.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a computer program executable by the processor, and when the processor executes the computer program, it performs the steps in the humanoid robot control method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it performs the steps in the humanoid robot control method as described in any one of claims 1-7.

Images