Humanoid robot simulation method based on autonomous driving technology

Abstract

The application discloses a humanoid robot simulation method based on automatic driving technology, which is executed in a virtual environment constructed by a MuJoCo physical engine and a Robosuite simulation framework. The method comprises the following steps: constructing a simulation environment integrated with dynamic elements; integrating a Realman-RM75 mechanical arm model and a multi-modal automatic driving perception system, configuring tactile, visual and other sensors and realizing multi-source data fusion; constructing a hierarchical decision and planning layer containing tasks, behaviors and motion planning, and realizing collaborative decision-making in a dynamic scene in combination with various algorithms; designing a perception-decision-planning-control closed-loop simulation for collaborative task completion; collecting multi-dimensional data and performing performance evaluation; and based on the evaluation results, using a Bayesian optimization framework to perform parameter iterative optimization. The application realizes the deep integration of automatic driving logic and humanoid robot operation, improves the navigation, operation and collaborative performance of the robot in a dynamic and complex scene, and provides reliable simulation support for efficient research and development of humanoid robots.

Description

Technical Field

[0001] This invention relates to the field of robot simulation and intelligent control technology, specifically to a humanoid robot simulation method based on autonomous driving technology. Background Technology

[0002] Humanoid robots are hailed as one of the ultimate forms of robotics, aiming to work autonomously in environments designed for humans, thereby unleashing enormous potential in service, rescue, and industrial collaboration scenarios. However, humanoid robots are high-dimensional, nonlinear, and underactuated complex systems integrating movement and manipulation, facing two core challenges in their development: first, achieving stable and intelligent movement and decision-making in unstructured, dynamically changing real-world environments; and second, completing precise and compliant object manipulation tasks. These two aspects are closely coupled, placing extremely high demands on the realism and completeness of simulation technology.

[0003] Meanwhile, autonomous driving technology has made significant progress in dynamic environment perception, behavior prediction, risk avoidance, and global path planning. Its technological paradigm is well-suited to solving the "where to go" and "how to arrive safely" problems of humanoid robots in open environments. However, the decision-making and planning outputs of autonomous driving (such as vehicle trajectories) cannot directly drive the joints of humanoid robots. Currently, there is a lack of an effective bridge to seamlessly connect and verify the "intelligent decision-making" of autonomous driving with the "physical execution" of humanoid robots in simulation.

[0004] Mujoco, a high-precision physics engine, is renowned for its accurate contact dynamics, joint friction, and flexible body simulation, making it ideal for high-fidelity simulation of complex balance control, gait, and hand-object interaction in humanoid robots. Robosuite is a modular simulation framework built on Mujoco, providing a rich set of robot models (extendable to humanoid robots), standardized operational task environments (such as grasping and assembly), and clear controller interfaces.

[0005] However, the existing Mujoco+Robosuite combination is designed primarily for manipulating robots with fixed bases or simple mobile platforms, and lacks simulation support for full-body humanoid robots that integrate autonomous driving-level navigation intelligence. Specifically, there is a lack of a unified simulation framework that can simultaneously handle macroscopic navigation commands from the autonomous driving stack and the dexterity manipulation of the upper body for specific tasks.

[0006] Therefore, there is an urgent need in this field for an innovative simulation method that can combine the environmental cognition and decision-making capabilities of autonomous driving with Mujoco's high-fidelity dynamics simulation and Robosuite's standardized task management to build a safe, efficient, and realistic simulation verification platform for humanoid robot operation, so as to accelerate the development and deployment of its intelligent control algorithms. Summary of the Invention

[0007] The present invention proposes a humanoid robot simulation method based on autonomous driving technology, which can at least solve one of the technical problems in the background art.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A humanoid robot simulation method based on autonomous driving technology includes the following steps: S100, Construct a simulation environment integrating dynamic elements; S200: In a simulation environment, integrate the Realman-RM75 robotic arm model with an autonomous driving perception system; S300, a decision-making and planning layer based on the autonomous driving perception system to construct autonomous driving logic; S400, design collaborative tasks are combined with decision-making and planning layers to perform closed-loop simulation; S500 performs data acquisition and multi-dimensional performance evaluation on closed-loop simulation results; S600 performs iterative parameter optimization based on multi-dimensional performance evaluation results.

[0009] Furthermore, the method for constructing a simulation environment integrating dynamic elements in step S100 of the present invention includes: Based on Robosuite's Arena environment building module, create complex scenes containing static obstacles, dynamic pedestrians, and interactive objects; utilize MuJoCo's high-fidelity physics engine to configure physical properties for all environment elements; Among them, the motion planning of dynamic pedestrians adopts a motion model based on social forces:

[0010] in, The force that moves toward the target The interaction force between pedestrians, This refers to the repulsive force between pedestrians and obstacles.

[0011] Furthermore, the integration of the Realman-RM75 robotic arm model in step S200 of the present invention includes: In the RoboSuite simulation framework, the MJCF dynamic model of the Realman-RM75 high-DOF manipulator was loaded and integrated, and the Rohand-A001 multi-fingered dexterous hand was installed at the end of the manipulator as an actuator. The overall dynamic model is described as follows:

[0012] in, For the generalized mass matrix, For the Coriolis force and centrifugal force terms, For gravity, For joint driving torque, To access the Jacobian matrix, For contact force, Joint angle, Joint velocity, Joint acceleration; Simultaneously, based on the physical structure of the robotic arm system, a multimodal sensor system is constructed: The system adopts a layered distributed architecture, including core modules: environmental perception module, ontology status module, and data fusion and synchronization module; The environmental perception module and the body state module collect data in parallel and transmit them to the data fusion and synchronization module respectively. The data fusion and synchronization module achieves timestamp alignment of the data from each sensor with a deviation of <100μs through a precise time protocol, and uses an extended Kalman filter to perform multi-sensor fusion to generate a unified state estimate that includes robot pose, joint state, environmental obstacles, and operation target. Finally, it provides state feedback to the decision planning layer and motion controller through a unified data bus.

[0013] Furthermore, the environmental perception module of the present invention is installed on the robotic arm base or external support, and is equipped with a binocular stereo camera, simulating RGB image data stream and a depth camera to collect 3D point cloud data of the environment; The body state sensors include: a six-axis inertial measurement unit (IMU) mounted on the robotic arm base to monitor the acceleration and angular velocity of the robotic arm system in real time for dynamic disturbance compensation and motion stability assessment; and encoders configured at each joint to provide real-time feedback of joint angles. With angular velocity , for the dynamic model , and Provides real-time status input.

[0014] Furthermore, the contact force F of the present invention C The measurement methods include: Adaptive gripping force control: It adopts a force closed-loop control law based on real-time tactile feedback to achieve compliant or stable gripping of objects of different weights and stiffness.

[0015] in, To capture the joint output torque, For the desired grasping power, This refers to the actual measured gripping force. For the proportional, integral, and derivative gains of the PID controller; Object slip detection and suppression: By monitoring the ratio of tangential to normal forces, slip prediction and dynamic gripping force compensation are achieved.

[0016]

[0017] when Approaching or exceeding the preset object-hand static friction coefficient At that time, it is determined that slippage is about to occur;

[0018] in, The observed real-time friction coefficient, The tangential force vector at the fingertip Normal force at the fingertip For the adjusted gripping force, To adjust the gain coefficient, This is the slip detection threshold.

[0019] Furthermore, the method for constructing the decision and planning layer of autonomous driving logic in step S300 of the present invention includes: Construct a hierarchical decision-making system comprising a task decision layer, a behavior decision layer, and a motion planning layer, and then build a planning architecture based on the hierarchical decision-making system: Among them, the task decision layer uses a hierarchical task network planner to decompose tasks and uses a predefined task method library to recursively decompose high-level goals into executable actions. By incorporating partially observable Markov decision process models to address environmental uncertainty, long-term returns can be optimized. The system integrates a large language model to perform natural language task parsing and common sense reasoning, and finally outputs a structured task sequence diagram. The task sequence diagram includes the execution order, preconditions, resource requirements and exception handling plans for each subtask. The behavior decision layer implements deterministic behavior logic based on a finite state machine (FSM), defining states, transition conditions, and corresponding behaviors. It adopts a utility-based behavior selection approach, comprehensively considering multi-objective optimization such as task progress, safety risks, energy consumption, and social suitability. A reinforcement learning policy network is introduced to learn adaptive behavior policies through simulation experience to cope with complex dynamic scenarios. Finally, a specific set of behavior instructions is output, which includes behavior type, parameters, execution conditions, expected duration, and priority. For the moving base, the motion planning layer uses a time elastic band algorithm to optimize the collision-free path in the local map in real time, while considering both time optimality and dynamic feasibility. For the whole body motion, a task priority control framework is adopted to prioritize multiple tasks such as base movement, arm operation, and balance maintenance, and achieve coordination through zero-space projection. Model predictive control is applied to optimize the rolling time domain and solve the optimal control sequence that satisfies multiple constraints online. The final output includes a sequence of joint trajectory points with a frequency greater than or equal to 100Hz, control torque commands for each joint, and related compliant control parameters.

[0020] Furthermore, the hierarchical decision-based planning architecture construction method of the present invention includes: This planning framework adopts a modular, layered, closed-loop architecture, mainly including: The module consists of an SmolVLA model module, a hierarchical decision and planning engine, a trajectory optimization and verification module, and an execution monitor. These modules are connected through a unified data bus to form a closed loop of perception, decision-making, planning, execution, and feedback. The planning engine employs a model-based predictive control framework.

[0021] in, For system status, To control the input, Q, R, P are weight matrices. This is the system dynamics equation.

[0022] Furthermore, the principle of the motion planning layer of the present invention includes: First, the composite behavior instruction is broken down into multiple subtasks, and ordered according to the following strict priority: Priority 1: Maintain dynamic balance; Priority 2: Obstacle avoidance and safety constraints; Priority 3: Tracking of the moving base; Priority 4: End effector operation tasks; For the moving base, algorithms such as TimedElasticBand are used to generate local trajectories that satisfy dynamic constraints:

[0023] in It is a sequence of path points. For time intervals. Indicates the first k The position vectors of the path points; Indicates the time of movement between adjacent path points; and The robot at each path point Velocity and acceleration at the point; For reference speed; , These are the weighting coefficients for each optimization objective; path point With the i The distance to each obstacle; This is the distance sensitivity scaling factor.

[0024] The optimized discrete path points are interpolated into a continuous trajectory:

[0025] In the formula, For mobile base at any time t Reference position vector; spline( ) is the spline interpolation function; This represents the total motion time.

[0026] And calculate the corresponding reference speed. and acceleration .

[0027] Simultaneously, RoboSuite's built-in or custom full-body controllers are invoked to coordinate chassis movements with upper body posture adjustments and arm operation tasks, ensuring overall balance and task execution during the movement: Let the total degrees of freedom of the robot be n, and initialize the null space projection matrix:

[0028] In the formula, Let be the initial null space projection matrix. for n × n An identity matrix. During initialization, all degrees of freedom are unoccupied by tasks. Subsequently, high-priority tasks will occupy the system's degrees of freedom first, while low-priority tasks will execute in the remaining degrees of freedom space through null projection, thus achieving multi-task collaboration. First priority task - Balance control: Calculate ZMP control force :

[0029] In the formula, The space inertia matrix for the first priority task. The ZMP reference acceleration vector, The ZMP reference position vector (the desired zero torque point position). This is the current actual ZMP position vector. The ZMP reference velocity vector, This represents the current actual ZMP velocity vector. , The proportional gain matrix and differential gain matrix are for the first priority task.

[0030] Corresponding joint torque for:

[0031] In the formula, Let be the Jacobian matrix corresponding to ZMP.

[0032] Update zero-space projection:

[0033] In the formula, The null projection matrix updated for the first priority task; Second priority task - obstacle avoidance constraint: Calculate obstacle avoidance repulsion force :

[0034] In the formula, M represents the total number of obstacles that the robot needs to avoid. Here, p represents the obstacle avoidance repulsion force gain coefficient, and p is the robot's current position vector. For the first i The position vectors of the obstacles; Corresponding joint torque for:

[0035] In the formula, The Jacobian matrix corresponding to the obstacle avoidance constraint; Update zero-space projection:

[0036] In the formula, The null projection matrix is ​​updated for the second priority task.

[0037] Third priority task - base trajectory tracking: Calculate the trajectory tracking force:

[0038] In the formula, The space inertia matrix for the third priority task. For the base at all times t The reference acceleration vector, The proportional gain matrix and differential gain matrix are for the third priority task. For the base at all times t The reference position vector, This is the vector representing the current actual position of the base. For the base at all times t The reference velocity vector, This represents the current actual velocity vector of the base. The corresponding joint torque is:

[0039] In the formula, It is the transpose of the Jacobian matrix corresponding to the base.

[0040] Update zero-space projection:

[0041] In the formula, The null projection matrix updated for the third priority task; Fourth priority task - end-point operation task: Calculate the end effector force:

[0042] In the formula, The space inertia matrix for the fourth priority task. For the end effector at time t Reference pose vector, For the end effector at time t The reference acceleration vector, The proportional gain matrix and differential gain matrix are for the fourth priority task. This represents the current actual pose vector of the end effector. For the end effector at time t The reference velocity vector, This represents the current actual velocity vector of the end effector. The corresponding joint torque is:

[0043] Update zero-space projection:

[0044] In the formula, The null projection matrix updated for the fourth priority task; Fifth priority task - Attitude optimization: Calculate attitude optimization torque :

[0045] in Here is the attitude gain matrix. This is the robot's nominal joint angle vector. This is the current actual joint angle vector of the robot.

[0046] Furthermore, the parameter iterative optimization method in step S600 of the present invention includes: Based on the performance evaluation report generated in step S500, the parameters are automatically tuned using a Bayesian optimization framework:

[0047] In the formula, x is the parameter vector to be optimized, X is the parameter search space, and f(x) is the objective function (, The optimal combination of parameters obtained through the search; The acquisition function is selected from the desired improvement function:

[0048] E[ [This represents the mathematical expectation.] f ( x ) for probabilistic surrogate model parameters x Performance predictions f ( x +) represents the optimal performance value among the currently evaluated parameters; The MuJoCo+Robosuite environment is used to quickly verify the optimization effect, forming a closed development loop of simulation-evaluation-optimization.

[0049] As can be seen from the above technical solution, this invention, by constructing a hierarchical decision-making and planning layer, deeply integrates the behavioral decision-making logic of autonomous driving (such as environmental perception, behavior prediction, and path planning) with the full-body motion control of humanoid robots (such as balance and grasping) within a unified framework. Through RoboSuite's standardized controller interface, precise execution of underlying actions is achieved, forming a complete perception-decision-planning-control closed loop, providing core technical support for verifying humanoid robot systems driven by autonomous driving logic.

[0050] Addressing the limitations of native simulation frameworks like MuJoCo+Robosuite, which focus primarily on fixed-base operations, this invention innovatively integrates high-precision models of dynamic pedestrians, interactive objects, the Realman-RM75 robotic arm, and the Rohand-A001 multi-fingered dexterous hand based on a social force model. By constructing complex scenes encompassing static obstacles, dynamic pedestrians, and interactive objects, high-fidelity simulations of humanoid robot movement and manipulation processes are achieved, significantly enhancing the realism coverage of dynamic and complex simulation scenarios.

[0051] This invention designs a detailed evaluation system covering multiple indicators such as navigation error, grasping success rate, torque smoothness, and perception latency, and combines it with a Bayesian optimization framework. This enables multi-dimensional quantitative evaluation of robot performance and constructs an efficient "simulation-evaluation-optimization" R&D closed loop. Based on the quantitative comparison results, algorithm bottlenecks can be clearly identified, and data-driven automatic parameter optimization replaces traditional manual experience-based debugging, significantly improving algorithm iteration efficiency.

[0052] This invention integrates a force feedback control law for a fingertip tactile sensor array, an object slip detection mechanism, and a task-priority-based zero-space projection control method to strongly couple multimodal sensor data (visual, tactile, etc.) with the robot's dynamic state. This high-fidelity physical interaction simulation allows the grasping strategy, force control parameters, and motion planning results trained in the simulation environment to be more easily transferred to real physical robots, effectively shortening the cycle from simulation algorithm development to actual deployment. Attached Figure Description

[0053] Figure 1 This is a flowchart of the humanoid robot simulation method based on autonomous driving technology of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0055] like Figure 1 As shown, the humanoid robot simulation method based on autonomous driving technology described in this embodiment is executed in a virtual environment constructed using the MuJoCo physics engine and the Robosuite simulation framework, and includes the following steps: S100, Construct a simulation environment integrating dynamic elements; S200: In a simulation environment, integrate the Realman-RM75 robotic arm model with an autonomous driving perception system; S300, a decision-making and planning layer based on the autonomous driving perception system to construct autonomous driving logic; S400, design collaborative tasks are combined with decision-making and planning layers to perform closed-loop simulation; S500 performs data acquisition and multi-dimensional performance evaluation on closed-loop simulation results; S600 performs iterative parameter optimization based on multi-dimensional performance evaluation results.

[0056] The following provides a detailed explanation of each step: S100, Construct a simulation environment integrating dynamic elements; Based on Robosuite's Arena environment building module, create complex scenes containing static obstacles, dynamic pedestrians (controlled by preset paths or random strategies), and interactive objects. Utilize MuJoCo's high-fidelity physics engine to configure precise physical properties for all environmental elements.

[0057] Static obstacles refer to objects that remain stationary in the simulation environment but affect the robot's path planning and navigation. These objects typically have physical collision properties but do not participate in active movement or manipulation tasks.

[0058] Interactive object scene models are objects that robots can grasp, push, manipulate, or interact with in their environment. These objects have full dynamic properties and degrees of freedom and can respond to external forces and contact.

[0059] Robosuite includes a set of basic geometric shapes in its MJCF files, and users can also write custom MJCF files to define complex static obstacles and interactive object scene models.

[0060] Among them, the motion planning of dynamic pedestrians adopts a motion model based on social forces:

[0061] in, The force that moves toward the target The interaction force between pedestrians, This represents the repulsive force between a pedestrian and an obstacle. α and β are indices used to distinguish different pedestrians, where α represents the target pedestrian whose force is currently being calculated, and β represents other pedestrians besides α. : Represents the index of a static obstacle in the environment.

[0062] S200, in a simulation environment, integrates a Realman-RM75 robotic arm model with an autonomous driving perception system. In the RoboSuite simulation framework, a high-precision MJCF dynamic model of the Realman-RM75 high-DOF robotic arm was loaded and integrated, and a Rohand-A001 multi-fingered dexterous hand was installed at its end effector as an actuator, forming a mobile manipulation unit with fine manipulation capabilities. This robotic arm has joint configurations, link masses, and rotational inertia consistent with the physical entity, and its overall dynamic model can be described as follows:

[0063] in, For the generalized mass matrix, For the Coriolis force and centrifugal force terms, For gravity, For joint driving torque, To access the Jacobian matrix, For contact force, Joint angle, Joint velocity, Joint acceleration.

[0064] To support autonomous driving-level perception and collaborative control, a multimodal sensor system tightly coupled with the robot's kinematics and dynamics is constructed based on the physical structure of the robotic arm system. The system adopts a layered distributed architecture, including core modules: environmental perception module, ontology status module, and data fusion and synchronization module; The environmental perception module and the body state module collect data in parallel and transmit them to the data fusion and synchronization module respectively. The fusion module achieves timestamp alignment of the data from each sensor (deviation <100μs) through the Precise Time Protocol (PTP) and uses an extended Kalman filter to perform multi-sensor fusion to generate a unified state estimate that includes robot pose, joint state, environmental obstacles, and operation target. Finally, it provides synchronous and consistent state feedback to the decision planning layer and motion controller through a unified data bus.

[0065] Among them, the environmental perception sensor (mounted on the robotic arm base or external support) is equipped with a binocular stereo camera to simulate RGB image data streams for object recognition, localization, and scene understanding within the workspace. A depth camera (such as a simulated RGB-D sensor) is also installed to provide 3D point cloud data of the environment for real-time obstacle detection and target reconstruction.

[0066] Body state sensors: A six-axis IMU (Inertial Measurement Unit) is installed on the robot arm base to monitor the acceleration and angular velocity of the robot arm system in real time for dynamic disturbance compensation and motion stability assessment. High-precision encoders are configured at each joint to provide real-time feedback of joint angles. With angular velocity , for the dynamic model , and Provides real-time status input.

[0067] The roles of each module in the layered distributed architecture are as follows: (1) Environment perception module: responsible for acquiring external environment information, including stereo vision unit, three-dimensional perception unit and panoramic perception unit, and outputting raw perception data through GMSL2.0 interface.

[0068] (2) Body status module: responsible for monitoring the robot's own status, including joint status unit, inertial measurement unit and force sensing unit, and transmitting high-frequency body data through EtherCAT bus.

[0069] (3) Data fusion and synchronization module: responsible for spatiotemporal alignment and state estimation of multi-source data, including time synchronization unit, sensor fusion unit and state estimation unit, and publishes unified state through ROS2-DDS middleware.

[0070] The Rohand-A001 multi-finger dexterity hand integrates tactile sensor arrays at each fingertip to measure contact force in real time. The distribution and magnitude of these forces enable the following specific force feedback and control functions: (1) Adaptive gripping force control: The force closed-loop control law based on real-time tactile feedback is adopted to achieve compliant or stable gripping of objects of different weights and stiffnesses.

[0071]

[0072] in, To capture the joint output torque, For the desired grasping power, This refers to the actual measured gripping force. For the proportional, integral, and derivative gains of the PID controller.

[0073] (2) Object slippage detection and suppression: By monitoring the ratio of tangential to normal force, slippage prediction and dynamic gripping force compensation are achieved.

[0074]

[0075] when Approaching or exceeding the preset object-hand static friction coefficient At that time, it is determined that slippage is about to occur.

[0076]

[0077] in, The observed real-time friction coefficient, The tangential force vector at the fingertip Normal force at the fingertip For the adjusted gripping force, To adjust the gain coefficient, The slip detection threshold (usually set to) ×0.8).

[0078] (3) Force guidance for precision operation: By using multi-point force distribution information, identify jamming, twisting and other states, and guide the robot to complete precision operation.

[0079] Sensor data fusion and interface: Through Robosuite's Observers module and the aforementioned data fusion and synchronization module, the synchronous acquisition, timestamp alignment, and unified publication of multi-source data from the environmental perception sensor and the body state sensor are realized.

[0080] Sensor data stream and robotic arm dynamics and contact force Real-time binding provides consistent and reliable state observation input for upper-level decision-making, motion planning, and force control.

[0081] S300, a decision-making and planning layer based on the autonomous driving perception system to construct autonomous driving logic; Methods for constructing the decision and planning layers of autonomous driving logic include: Construct a hierarchical decision-making system comprising a task decision layer (upper layer), a behavior decision layer (middle layer), and a motion planning layer (lower layer): Among them, the task decision layer (upper layer) uses a hierarchical task network (HTN) planner to decompose tasks and uses a predefined task method library to recursively decompose high-level goals into executable actions. By incorporating a partially observable Markov decision process (POMDP) ​​model to address environmental uncertainty, long-term returns can be optimized. Integrating large language models (LLM) for natural language task parsing and commonsense reasoning enhances the generalization ability of task understanding; The final output is a structured task sequence diagram, which includes the execution order, prerequisites, resource requirements, and contingency plans for each subtask.

[0082] Decision-making process: Receive and parse high-level task instructions (such as "navigate from point A to point B and grab the cup on the table"), decompose complex tasks into an ordered sequence of subtasks, determine the timing and logical constraints between tasks, and allocate time and resource budgets.

[0083] The behavioral decision layer (middle layer) implements deterministic behavioral logic based on a finite state machine (FSM), defining states, transition conditions, and corresponding behaviors. It adopts a utility-based behavior selection approach, comprehensively considering multiple objectives such as task progress, safety risks, energy consumption, and social suitability. It introduces a reinforcement learning policy network to learn adaptive behavior policies through simulation experience to cope with complex dynamic scenarios. Finally, it outputs a specific set of behavioral instructions, which includes behavior type, parameters, execution conditions, expected duration, and priority.

[0084] Decision content: Based on the currently executed sub-task and real-time perceived information (such as dynamic pedestrian positions and obstacle status), select specific behavioral strategies (such as "go around stationary obstacles", "yield to pedestrians", "wait at the intersection").

[0085] For the moving base, the motion planning layer (lower layer) uses the Temporal Elastic Band (TEB) algorithm to optimize the collision-free path in the local map in real time, while considering both time optimality and dynamic feasibility. For the whole-body motion, a task priority control framework is adopted to prioritize multiple tasks such as base movement, arm operation, and balance maintenance, and achieve coordination through null-space projection. Model predictive control (MPC) is applied to perform rolling time-domain optimization and solve for the optimal control sequence that satisfies multiple constraints online. Final output: a high-frequency (≥100Hz) sequence of joint trajectory points, control torque commands for each joint, and related compliant control parameters.

[0086] Decision content: Transform behavioral instructions into precise motion trajectories that satisfy robot dynamics and kinematic constraints, and generate low-level control parameters.

[0087] Hierarchical decision-making-based planning architecture methods include: This planning framework adopts a modular, layered, closed-loop architecture, mainly including: SmolVLA model module: Enables multimodal task understanding, transforming natural language instructions and environmental awareness information into executable task representations.

[0088] Hierarchical Decision and Planning Engine: Contains the three-layer planner mentioned above and is the core computing module. Trajectory optimization and verification module: responsible for trajectory collision detection, dynamic feasibility verification, and smoothing. Execution monitor: tracks execution status, detects anomalies, and triggers replanning. The modules are connected through a unified data bus, forming a closed loop of perception → decision-making → planning → execution → feedback.

[0089] Working principle and process: Initialization: Upon receiving task instructions, the upper-level planner decomposes the tasks and generates an initial task sequence.

[0090] Real-time loop (typical period 10ms): 1. Sensor Input: Obtain the latest sensor data and input it into the SmolVLA model.

[0091] 2. Status assessment: Monitor the current task execution progress and detect anomalies (such as sudden obstacles or execution deviations).

[0092] 3. Hierarchical decision-making: Upper layer: Evaluate whether the task sequence needs adjustment.

[0093] Middle layer: Based on the SmolVLA model, select or adjust the current optimal behavior.

[0094] Lower layer: Based on the selected behavior and the precise robot state, solve for the motion trajectory and control commands.

[0095] 4. Trajectory Verification: Perform rapid collision detection and dynamic verification on the planned trajectory to ensure safety and feasibility.

[0096] 5. Command Issuance: Send the verified control command to the underlying controller.

[0097] 6. Monitoring and Feedback: Collect execution results to optimize the SmolVLA model and evaluate planning quality.

[0098] Anomaly handling: When trajectory execution failure, sudden obstacles, or system resource shortages are detected, the corresponding level of replanning or behavior switching is triggered to ensure system robustness.

[0099] Output content Upper-level output: macro-level task sequences and strategies to guide long-term behavior.

[0100] Mid-level output: Specific behavioral instructions, responding to environmental changes.

[0101] Lower-level output: precise time-state trajectory and control parameters, directly driving the actuator.

[0102] Global Path Planner: Based on the simulation environment built in step S100, it receives the target task (such as "navigate from point A to point B and grab the cup on the table") and generates a collision-free global path using the A or DLite algorithm.

[0103] Local behavior decision maker: Introducing a finite state machine (FSM) or a learning-based policy network from autonomous driving, it makes high-level behavior decisions based on the global path, real-time perception information (such as dynamic pedestrian positions), and robot state, such as "follow the path", "wait at the intersection", "go around stationary obstacles", or "avoid pedestrians".

[0104] Local motion planner: employing a model predictive control framework.

[0105] in, For system status, To control the input, Q, R, P are weight matrices. This is the system dynamics equation.

[0106] Transforming behavioral decisions into specific movement trajectories: First, the composite behavior instruction is broken down into multiple subtasks, and ordered according to the following strict priority: Priority 1: Dynamic balance maintenance (preventing tipping) Priority 2: Obstacle avoidance and safety constraints (collision prevention) Priority 3: Tracking of mobile base (navigation task) Priority 4: End effector operation tasks (grabbing / placing, etc.) Priority 5: Overall posture optimization (joint limit avoidance) For the mobile base, algorithms such as TimedElasticBand are used to generate local trajectories that satisfy dynamic constraints.

[0107] in It is a sequence of path points. For time intervals. Indicates the first k The position vectors of the path points; Indicates the time of movement between adjacent path points; and The robot at each path point Velocity and acceleration at the point; For reference speed; , These are the weighting coefficients for each optimization objective; path point With the i The distance to each obstacle; This is the distance sensitivity scaling factor.

[0108] The optimized discrete path points are interpolated into a continuous trajectory:

[0109] In the formula, For mobile base at any time t Reference position vector; spline( ) is the spline interpolation function; This represents the total motion time.

[0110] And calculate the corresponding reference speed. and acceleration .

[0111] Simultaneously, RoboSuite's built-in or custom full-body controllers are invoked to coordinate chassis movements with upper body posture adjustments and arm operation tasks, ensuring overall balance and task execution during the movement: Let the total degrees of freedom of the robot be n, and initialize the null space projection matrix:

[0112] In the formula, Let be the initial null space projection matrix. for n × n A 3D identity matrix. During initialization, all degrees of freedom are unoccupied by tasks. Subsequently, higher-priority tasks will occupy the system's degrees of freedom first, while lower-priority tasks will execute in the remaining degrees of freedom space through null projection, thus achieving multi-task collaboration.

[0113] 1. First Priority Task - Balance Control: Calculate ZMP control force :

[0114] In the formula, The space inertia matrix for the first priority task. The ZMP reference acceleration vector, The ZMP reference position vector (the desired zero torque point position). This is the current actual ZMP position vector. The ZMP reference velocity vector, This represents the current actual ZMP velocity vector. , The proportional gain matrix and differential gain matrix are for the first priority task.

[0115] Corresponding joint torque for:

[0116] In the formula, Let be the Jacobian matrix corresponding to ZMP.

[0117] Update zero-space projection:

[0118] In the formula, This is the null projection matrix updated for the first priority task.

[0119] 2. Second Priority Task - Obstacle Avoidance Constraints: Calculate obstacle avoidance repulsion force :

[0120] In the formula, M represents the total number of obstacles that the robot needs to avoid. Here, p represents the obstacle avoidance repulsion force gain coefficient, and p is the robot's current position vector. For the first i The position vectors of the obstacles.

[0121] Corresponding joint torque for:

[0122] In the formula, This is the Jacobian matrix corresponding to the obstacle avoidance constraint.

[0123] Update zero-space projection:

[0124] In the formula, The null projection matrix is ​​updated for the second priority task.

[0125] 3. Third Priority Task - Base Trajectory Tracking: Calculate the trajectory tracking force:

[0126] In the formula, The space inertia matrix for the third priority task. For the base at all times t The reference acceleration vector, The proportional gain matrix and differential gain matrix are for the third priority task. For the base at all times t The reference position vector, This is the vector representing the current actual position of the base. For the base at all times t The reference velocity vector, This represents the current actual velocity vector of the base.

[0127] The corresponding joint torque is:

[0128] In the formula, It is the transpose of the Jacobian matrix corresponding to the base.

[0129] Update zero-space projection:

[0130] In the formula, The null projection matrix is ​​updated for the third priority task.

[0131] 4. Fourth Priority Task - End-of-Line Operation Task: Calculate the end effector force:

[0132] In the formula, The space inertia matrix for the fourth priority task. For the end effector at time t The reference pose vector (position + attitude). For the end effector at time t The reference acceleration vector, The proportional gain matrix and differential gain matrix are for the fourth priority task. This represents the current actual pose vector of the end effector. For the end effector at time t The reference velocity vector, This represents the current actual velocity vector of the end effector.

[0133] The corresponding joint torque is:

[0134] Update zero-space projection:

[0135] In the formula, The null projection matrix is ​​updated for the fourth priority task.

[0136] 5. Fifth Priority Task - Attitude Optimization: Calculate attitude optimization torque :

[0137] in Here is the attitude gain matrix. This is the robot's nominal joint angle vector. This is the current actual joint angle vector of the robot.

[0138] The entire collaborative planning process is completed within a single control cycle (typically 10ms):

[0139] S400, design collaborative tasks are combined with decision-making and planning layers to perform closed-loop simulation; Define complex movement operation task sequences in RoboSuite's Task module. During simulation runtime, a closed-loop architecture of perception-decision-planning-control is formed. Sensor data and robot status are obtained from the MuJoCo simulation engine via the Robosuite interface.

[0140] The perception module processes data and identifies obstacles, pedestrians, and target objects.

[0141] Based on the perception results and the current task, the decision-making and planning layer generates high-level behavioral instructions and low-level movement trajectories.

[0142] Precise force control is achieved through Robosuite's OSC controller, with the following control law:

[0143] in, Let J be the mission space inertia matrix, J be the Jacobian matrix, and M be the joint space generalized mass matrix. For the desired operating force, For dynamic consistency matrix, It is a zero-space torque.

[0144] MuJoCo calculates the next physical state and feeds it back to the simulation loop.

[0145] S500 performs data acquisition and multi-dimensional performance evaluation on closed-loop simulation results; During the simulation, the system uses RoboSuite's DataCollector module and a custom evaluation plugin to record and analyze the following multi-dimensional performance data in real time to comprehensively evaluate the performance of the robotic arm system in autonomous driving tasks. The evaluation table is as follows:

[0146] S600: Parameters are iteratively optimized based on multi-dimensional performance evaluation results; Based on the performance evaluation report generated in step S500, algorithm bottlenecks are identified. The system is then iteratively optimized by adjusting the parameters of the decision logic, optimizing the cost function of the motion planner, or retraining the learning-based perception and decision model.

[0147] Automatic parameter tuning is performed using a Bayesian optimization framework:

[0148] In the formula, x is the parameter vector to be optimized, X is the parameter search space, and f(x) is the objective function (obtained by weighting the multi-dimensional evaluation indicators in step S500). This represents the optimal combination of parameters obtained through the search.

[0149] The acquisition function is selected from the desired improvement function:

[0150] E[ [This represents the mathematical expectation.] f ( x ) for probabilistic surrogate model parameters x Performance predictions f ( x +) represents the optimal performance value among the currently evaluated parameters.

[0151] The MuJoCo+Robosuite environment is used to quickly verify the optimization effect, forming an efficient development closed loop of "simulation-evaluation-optimization".

[0152] The optimization effect was evaluated by quantitatively comparing key performance indicators before and after optimization, as shown in the table below:

[0153] The optimization process targets the key adjustable parameters of each module in the system, and their corresponding relationships and physical meanings are as follows:

[0154] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to a typical embodiment.

[0155] Example 1: Simulation of a humanoid robot driving a vehicle to perform urban road tasks This embodiment simulates a humanoid robot as a driver, controlling a traditional vehicle to complete a driving task in a complex urban road environment.

[0156] 1. Simulation environment setup; Construct a city scene that includes the following elements: Road Network: Use Robosuite's modular components to build a road system that includes intersections, curves, ramps, and set up traffic signs and traffic lights.

[0157] Vehicle Model: Create a passenger car model with a complete steering system, where the steering wheel and other control mechanisms are set as movable joints.

[0158] Dynamic environment: Configure other AI-controlled traffic vehicles that follow traffic rules but exhibit randomness. Pedestrian models were placed along the roadside to simulate random pedestrian crossing behavior. Set different weather conditions (sunny, rainy) and light variations (day, night). 2. System model and sensor integration; Humanoid robot model: A high-precision MJCF dynamics model of the Realman-RM75 (6FB version) high-DOF robotic arm is loaded and integrated into RoboSuite, and a Rohand-A001 multi-fingered dexterous hand is installed at its end as an actuator, forming a mobile manipulator unit with fine manipulation capabilities, which is then placed in the driver's seat of a vehicle. The robot's skeletal structure and joint range must ensure that it can naturally reach the steering wheel and gear lever.

[0159] Vehicle control interface modeling: The steering wheel is configured as a rotary joint, establishing a transmission relationship with the vehicle's front wheel steering mechanism. The shift lever is designed as a multi-degree-of-freedom joint with clearly defined shift points. Sensor system: Environmental perception sensors: Multiple cameras and LiDAR are installed around the vehicle to provide environmental information. Driving status sensors: Equipped with dashboard recognition camera, vehicle speed sensor, etc. Robot body sensors: joint position sensors, hand tactile sensors (for sensing steering wheel grip force). 3. Implementation of the hierarchical control system; This system adopts a three-tier control architecture: High-level decision-making level; A deep learning-based behavior prediction model is used to analyze the intentions of traffic participants. The task planner generates a global path based on the destination; The decision-making state machine switches between driving behaviors such as "going straight", "turning", "overtaking", and "stopping"; Middle-level coordination layer: Motion planner: Translates high-level decisions into specific vehicle motion trajectories, including speed curves and paths; Manipulation command generator: Converts vehicle motion requirements into robot limb movement commands; Calculate the steering wheel rotation angle and speed based on steering requirements; Plan the robot's hand movement sequence during gear shifting.

[0160] Low-level execution layer: Robot motion controller: Robosuite's OSC controller precisely controls the robot's upper limbs to perform steering wheel operations and gear shifting. Fine-grained operation control: The hand grip controller maintains appropriate steering wheel grip force.

[0161] 4. Typical driving task test; The system tests the following typical scenarios: Basic handling test: basic operations such as straight driving, turning, and U-turn.

[0162] Complex traffic scenarios: Crossing an intersection: Recognize traffic signals and start or stop at the appropriate time; Lane keeping and changing: Complete lane changes only when it is safe to do so; Emergency obstacle avoidance: Responding quickly to obstacles that suddenly appear.

[0163] 5. Performance evaluation and optimization; Collect multidimensional data during the simulation process: Driving performance indicators: speed control accuracy, path tracking error; Handling quality indicators: steering wheel operation smoothness, gear shifting success rate; Safety indicators: number of collisions, traffic rule violations, emergency braking reaction time; By analyzing this data, we can: Optimize the robot's control strategy to make its driving style more similar to that of human experts; Adjust control parameters to improve vehicle stability under various operating conditions; Train better perception and decision-making models to improve the ability to cope with complex scenarios.

[0164] In summary, this embodiment constructs a highly realistic humanoid robot driving simulation system, providing a powerful experimental platform for studying human driving behavior, developing autonomous driving technology, and exploring the performance of humanoid robots in complex tasks. This system can safely and efficiently test various driving algorithms and control strategies, accelerating the development of related technologies.

[0165] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0166] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0167] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the humanoid robot simulation methods based on autonomous driving technology described above.

[0168] It is understood that the systems, devices, and storage media provided in the embodiments of the present invention correspond to the methods provided in the embodiments of the present invention, and the explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.

[0169] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0170] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0171] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0172] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A humanoid robot simulation method based on autonomous driving technology, characterized in that, Includes the following steps: S100, Construct a simulation environment integrating dynamic elements; S200: In a simulation environment, integrate the Realman-RM75 robotic arm model with an autonomous driving perception system; S300, a decision-making and planning layer based on the autonomous driving perception system to construct autonomous driving logic; S400, design collaborative tasks are combined with decision-making and planning layers to perform closed-loop simulation; S500 performs data acquisition and multi-dimensional performance evaluation of closed-loop simulation results; S600 performs iterative parameter optimization based on multi-dimensional performance evaluation results.

2. The humanoid robot simulation method based on autonomous driving technology according to claim 1, characterized in that, The method for constructing a simulation environment integrating dynamic elements in step S100 includes: Based on Robosuite's Arena environment building module, create complex scenes containing static obstacles, dynamic pedestrians, and interactive objects; utilize MuJoCo's high-fidelity physics engine to configure physical properties for all environment elements; Among them, the motion planning of dynamic pedestrians adopts a motion model based on social forces: in, The force that moves toward the target The interaction force between pedestrians, This refers to the repulsive force between pedestrians and obstacles.

3. The humanoid robot simulation method based on autonomous driving technology according to claim 1, characterized in that, Step S200, which integrates the Realman-RM75 robotic arm model, includes: In the RoboSuite simulation framework, the MJCF dynamic model of the Realman-RM75 high-DOF manipulator was loaded and integrated, and the Rohand-A001 multi-fingered dexterous hand was installed at the end of the manipulator as an actuator. The overall dynamic model is described as follows: in, For the generalized mass matrix, For the Coriolis force and centrifugal force terms, For gravity, For joint driving torque, To access the Jacobian matrix, For contact force, Joint angle, Joint velocity, Joint acceleration; Simultaneously, based on the physical structure of the robotic arm system, a multimodal sensor system is constructed: The system adopts a layered distributed architecture, including core modules: environmental perception module, ontology status module, and data fusion and synchronization module; The environmental perception module and the body state module collect data in parallel and transmit them to the data fusion and synchronization module respectively. The data fusion and synchronization module achieves timestamp alignment of the data from each sensor with a deviation of <100μs through a precise time protocol, and uses an extended Kalman filter to perform multi-sensor fusion to generate a unified state estimate that includes robot pose, joint state, environmental obstacles, and operation target. Finally, it provides state feedback to the decision planning layer and motion controller through a unified data bus.

4. The humanoid robot simulation method based on autonomous driving technology according to claim 3, characterized in that, The environmental perception module is installed on the robotic arm base or external support, and is equipped with a binocular stereo camera, a simulated RGB image data stream and a depth camera to collect 3D point cloud data of the environment. The body state sensor includes: a six-axis inertial measurement unit (IMU) installed on the base of the robotic arm to monitor the acceleration and angular velocity of the robotic arm system in real time for dynamic disturbance compensation and motion stability assessment; Encoders are installed at each joint to provide real-time feedback of joint angles. With angular velocity , for the dynamic model , and Provides real-time status input.

5. The humanoid robot simulation method based on autonomous driving technology according to claim 3, characterized in that, Contact force F C The measurement methods include: Adaptive gripping force control: It adopts a force closed-loop control law based on real-time tactile feedback to achieve compliant or stable gripping of objects of different weights and stiffness. in, To capture the joint output torque, For the desired grasping power, This is the actual measured gripping force. For the proportional, integral, and derivative gains of the PID controller; Object slip detection and suppression: By monitoring the ratio of tangential to normal forces, slip prediction and dynamic gripping force compensation are achieved; when Approaching or exceeding the preset object-hand static friction coefficient At that time, it is determined that slippage is about to occur; in, The observed real-time friction coefficient, The tangential force vector at the fingertip Normal force at the fingertip For the adjusted gripping force, To adjust the gain coefficient, This is the slip detection threshold.

6. The humanoid robot simulation method based on autonomous driving technology according to claim 1, characterized in that, The decision and planning layer method for constructing autonomous driving logic in step S300 includes: Construct a hierarchical decision-making system comprising a task decision layer, a behavior decision layer, and a motion planning layer, and then build a planning architecture based on the hierarchical decision-making system: Among them, the task decision layer uses a hierarchical task network planner to decompose tasks and uses a predefined task method library to recursively decompose high-level goals into executable actions. By incorporating partially observable Markov decision process models to address environmental uncertainty, long-term returns can be optimized. The system integrates a large language model to perform natural language task parsing and common sense reasoning, and finally outputs a structured task sequence diagram. The task sequence diagram includes the execution order, preconditions, resource requirements and exception handling plans for each subtask. The behavior decision layer implements deterministic behavior logic based on a finite state machine (FSM), defining states, transition conditions, and corresponding behaviors. It adopts a utility-based behavior selection approach, comprehensively considering multi-objective optimization such as task progress, safety risks, energy consumption, and social suitability. A reinforcement learning policy network is introduced to learn adaptive behavior policies through simulation experience to cope with complex dynamic scenarios. Finally, a specific set of behavior instructions is output, which includes behavior type, parameters, execution conditions, expected duration, and priority. For the moving base, the motion planning layer uses a time elastic band algorithm to optimize the collision-free path in the local map in real time, while also incorporating time optimality and dynamic feasibility. For the whole body motion, a task priority control framework is adopted to prioritize multiple tasks such as base movement, arm operation, and balance maintenance, and achieve coordination through zero-space projection. Model predictive control is applied to optimize the rolling time domain and solve the optimal control sequence that satisfies multiple constraints online. The final output includes a sequence of joint trajectory points with a frequency greater than or equal to 100Hz, control torque commands for each joint, and related compliant control parameters.

7. The humanoid robot simulation method based on autonomous driving technology according to claim 6, characterized in that, The hierarchical decision-based planning architecture construction method includes: This planning framework adopts a modular, layered, closed-loop architecture, mainly including: The module consists of an SmolVLA model module, a hierarchical decision and planning engine, a trajectory optimization and verification module, and an execution monitor. These modules are connected through a unified data bus to form a closed loop of perception, decision-making, planning, execution, and feedback. The planning engine employs a model-based predictive control framework. in, For system status, To control the input, Q, R, P are weight matrices. This is the system dynamics equation.

8. The humanoid robot simulation method based on autonomous driving technology according to claim 6, characterized in that, The principles of motion planning layers include: First, the composite behavior instruction is broken down into multiple subtasks, and ordered according to the following strict priority: Priority 1: Maintain dynamic balance; Priority 2: Obstacle avoidance and safety constraints; Priority 3: Tracking of the moving base; Priority 4: End effector operation tasks; For the moving base, algorithms such as TimedElasticBand are used to generate local trajectories that satisfy dynamic constraints: in It is a sequence of path points. For time intervals; Indicates the first k Position vectors of path points; Indicates the time of movement between adjacent path points; and The robot at each path point Velocity and acceleration at the point; For reference speed; , These are the weighting coefficients for each optimization objective; path point With the i The distance to each obstacle; This is the distance sensitivity scaling factor; The optimized discrete path points are interpolated into a continuous trajectory: In the formula, For mobile base at any time t Reference position vector; spline( ) is the spline interpolation function; Total motion time; And calculate the corresponding reference speed. and acceleration ; Simultaneously, RoboSuite's built-in or custom full-body controllers are invoked to coordinate chassis movements with upper body posture adjustments and arm operation tasks, ensuring overall balance and task execution during the movement: Let the total degrees of freedom of the robot be n, and initialize the null space projection matrix: In the formula, Let be the initial null space projection matrix. for n × n The system is a 3D identity matrix. During initialization, all degrees of freedom are not occupied by tasks. Subsequently, high-priority tasks will occupy the system's degrees of freedom first, while low-priority tasks will be executed in the remaining degrees of freedom space through null projection, thereby achieving multi-task collaboration. First priority task - Balance control: Calculate ZMP control force : In the formula, The space inertia matrix for the first priority task. The ZMP reference acceleration vector, This is the ZMP reference position vector. This is the current actual ZMP position vector. The ZMP reference velocity vector, This represents the current actual ZMP velocity vector. , For the first priority task, the proportional gain matrix and the differential gain matrix are respectively. Corresponding joint torque for: In the formula, This is the Jacobian matrix corresponding to ZMP; Update zero-space projection: In the formula, The null projection matrix updated for the first priority task; Second priority task - obstacle avoidance constraint: Calculate obstacle avoidance repulsion force : In the formula, M represents the total number of obstacles that the robot needs to avoid. Here, p represents the obstacle avoidance repulsion force gain coefficient, and p is the robot's current position vector. Let i be the position vector of the i-th obstacle; Corresponding joint torque for: In the formula, The Jacobian matrix corresponding to the obstacle avoidance constraint; Update zero-space projection: In the formula, The null projection matrix updated for the second priority task; Third priority task - base trajectory tracking: Calculate the trajectory tracking force: In the formula, The space inertia matrix for the third priority task. For the base at all times t The reference acceleration vector, The proportional gain matrix and differential gain matrix are for the third priority task. For the base at all times t The reference position vector, This is the vector representing the current actual position of the base. For the base at all times t The reference velocity vector, This represents the current actual velocity vector of the base. The corresponding joint torque is: In the formula, This is the transpose of the Jacobian matrix corresponding to the base. Update zero-space projection: In the formula, The null projection matrix updated for the third priority task; Fourth priority task - end-point operation task: Calculate the end effector force: In the formula, The space inertia matrix for the fourth priority task. For the end effector at time t Reference pose vector, For the end effector at time t The reference acceleration vector, The proportional gain matrix and differential gain matrix are for the fourth priority task. This represents the current actual pose vector of the end effector. For the end effector at time t The reference velocity vector, This represents the current actual velocity vector of the end effector. The corresponding joint torque is: Update zero-space projection: In the formula, The null projection matrix updated for the fourth priority task; Fifth priority task - Attitude optimization: Calculate attitude optimization torque : in Here is the attitude gain matrix. This is the robot's nominal joint angle vector. This is the current actual joint angle vector of the robot.

9. The humanoid robot simulation method based on autonomous driving technology according to claim 1, characterized in that, The parameter iterative optimization method in step S600 includes: Based on the performance evaluation report generated in step S500, automatic parameter tuning is performed using a Bayesian optimization framework: In the formula, x is the parameter vector to be optimized, X is the parameter search space, and f(x) is the objective function. The optimal combination of parameters obtained through the search; The acquisition function is selected from the desired improvement function: E[ [This represents the mathematical expectation.] f ( x For probabilistic surrogate models, the parameters are... x Performance predictions f ( x +) represents the optimal performance value among the currently evaluated parameters; The MuJoCo+Robosuite environment is used to quickly verify the optimization effect, forming a closed development loop of simulation-evaluation-optimization.

Images