Humanoid robot body design method and device and medium

Abstract

The invention discloses a humanoid robot body design method and device and a medium, and relates to the technical field of kinetic energy control, and the method comprises the steps: constructing an interactive dynamics digital twin of a humanoid robot; the method comprises the following steps: simulating an interactive dynamics digital twin in a virtual environment, and obtaining real-time relative space relation data through a distributed visual sensor network; in the process that the humanoid robot executes the self-adaptive motion trail, environment interaction is monitored through a distributed visual sensor network, and an actual kinetic energy loss value is collected; and inputting the energy consumption deviation value between the expected kinetic energy loss value and the actual kinetic energy loss value into the reinforcement learning model, and iteratively optimizing the prediction precision of the contact event type and the motion planning parameters of the online motion planner. According to the method, the expected kinetic energy loss value serves as a dynamic optimization target and is embedded into the online motion planner to generate the self-adaptive motion track, the motion planning process can be optimized in real time, and therefore it is ensured that task constraint and dynamic limitation are met while the kinetic energy loss is minimized.

Description

Technical Field

[0001] This invention relates to the field of kinetic energy control technology, and in particular to a humanoid robot body design method, equipment and medium. Background Technology

[0002] In recent years, humanoid robot technology has made significant progress. Digital twins are widely used in robot dynamics simulation and behavior prediction, enabling real-time simulation in virtual environments by constructing interactive dynamic models. Distributed visual sensor networks can simultaneously capture multi-view environmental image streams and generate high-precision environmental depth point cloud data using stereo matching algorithms, providing a foundation for spatial relationship analysis. Online motion planners combine dynamic constraints and task requirements, employing numerical optimization algorithms to generate motion trajectories, thereby enhancing the robot's autonomous movement capabilities.

[0003] However, existing technologies often fail to accurately predict the type of contact event and the contact force vector when dealing with humanoid robot interactions with the environment, resulting in a lack of effective integration of kinetic energy loss assessment during motion planning. This makes the robot's motion trajectory prone to high-energy-consuming contact events in real-world environments, increasing energy consumption and reducing task efficiency, especially due to the lack of adaptive optimization capabilities in dynamic and unknown environments. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a humanoid robot body design method to solve the problem of excessive kinetic energy loss caused by environmental contact events during the movement of humanoid robots.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for designing a humanoid robot body, comprising: constructing an interactive dynamics digital twin of the humanoid robot; simulating the interactive dynamics digital twin in a virtual environment and acquiring real-time relative spatial relationship data through a distributed visual sensor network; predicting contact event types and contact force vectors based on the real-time relative spatial relationship data, and calculating the expected kinetic energy loss value generated by external interaction; embedding the expected kinetic energy loss value as a dynamic optimization target into an online motion planner to generate an adaptive motion trajectory that can avoid high-energy-consuming contact events; monitoring environmental interaction and collecting actual kinetic energy loss values ​​through a distributed visual sensor network during the execution of the adaptive motion trajectory by the humanoid robot; and inputting the energy consumption deviation between the expected kinetic energy loss value and the actual kinetic energy loss value into a reinforcement learning model to iteratively optimize the prediction accuracy of contact event types and the motion planning parameters of the online motion planner.

[0007] As a preferred embodiment of the humanoid robot body design method of the present invention, the specific steps for constructing the interactive dynamics digital twin of the humanoid robot are as follows: Acquire the three-dimensional geometric model data and physical characteristic parameters of the humanoid robot; Import the three-dimensional geometric model data and physical property parameters into the dynamic simulation software, and define the joint constraints and range of motion of the humanoid robot based on the mechanical structure of the humanoid robot; Based on the joint constraints and range of motion of the humanoid robot, the deployment location and field of view parameters of the distributed vision sensor network are determined, and an interactive dynamic digital twin is constructed in the dynamic simulation software.

[0008] As a preferred embodiment of the humanoid robot body design method of the present invention, the specific steps for acquiring real-time relative spatial relationship data are as follows: Simultaneous capture of multi-view environmental image streams via a distributed visual sensor network; Perform stereo matching on multi-view environmental image streams to generate environmental depth point cloud data; The environmental depth point cloud data is registered with the humanoid robot body model to calculate the relative distance and angle between the humanoid robot and environmental objects, and output real-time relative spatial relationship data.

[0009] As a preferred embodiment of the humanoid robot body design method of the present invention, the predicted contact event type and contact force vector are realized by forward inference through a pre-trained contact mechanics deep learning model based on real-time relative spatial relationship data.

[0010] In a preferred embodiment of the humanoid robot body design method of the present invention, the specific steps for calculating the expected kinetic energy loss value generated by external interaction are as follows: Based on the predicted type of contact event, determine the corresponding contact dynamics equations; Input the predicted contact force vector into the contact dynamics equation to calculate the work done by the instantaneous contact force; Integrate the work done by the instantaneous contact force within the predicted contact time interval to output the expected kinetic energy loss value.

[0011] As a preferred embodiment of the humanoid robot body design method of the present invention, the specific steps for generating an adaptive motion trajectory that can avoid high-energy-consuming contact events are as follows: The online motion planner receives the expected kinetic energy loss value as a dynamic optimization target; Invoke predefined humanoid robot motion task constraints and interaction dynamics digital twin dynamic constraints, and construct a multi-objective optimization function based on an online motion planner; A multi-objective optimization function is solved using a numerical optimization algorithm to generate candidate motion trajectories. The candidate motion trajectory with the smallest expected kinetic energy loss value is selected from the candidate motion trajectories and used as the adaptive motion trajectory.

[0012] As a preferred embodiment of the humanoid robot body design method described in this invention, the reinforcement learning model pre-training process includes using energy consumption deviation value as a reward signal, updating the prediction algorithm parameters of the contact event type and the motion planning parameters of the online motion planner through the policy gradient method, and minimizing the long-term cumulative energy consumption deviation.

[0013] As a preferred embodiment of the humanoid robot body design method described in this invention, the prediction accuracy of contact event types and the motion planning parameters of the online motion planner are iteratively optimized through a reinforcement learning model. The specific steps are as follows: The energy consumption deviation value is input into the reinforcement learning model as a reward signal for the current state; The reinforcement learning model calculates parameter adjustments for the prediction algorithm parameters and motion planning parameters based on the reward signal; the parameter adjustments are then applied to update the prediction algorithm parameters and the online motion planner.

[0014] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the humanoid robot body design method as described in the first aspect of the present invention.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the humanoid robot body design method as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: By embedding the expected kinetic energy loss value as a dynamic optimization objective into the online motion planner to generate adaptive motion trajectories, the motion planning process can be optimized in real time, thereby minimizing kinetic energy loss while ensuring the satisfaction of task constraints and dynamic limitations. The integration of the expected kinetic energy loss value allows the motion trajectory to pre-assess the energy consumption impact of contact events in a virtual environment, thereby avoiding high-energy-consuming interactions, improving energy utilization efficiency, further enhancing the robot's motion endurance and overall performance, adapting to the long-term operational needs in complex environments, and continuously improving the quality of trajectory generation through an iterative optimization mechanism. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the humanoid robot body design method.

[0019] Figure 2 Flowchart for constructing an interactive dynamics digital twin.

[0020] Figure 3 A flowchart for generating real-time relative spatial relationship data.

[0021] Figure 4 Flowchart for predicting contact events and calculating expected kinetic energy loss. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides a humanoid robot body design method, including the following steps: S1: Constructing an interactive dynamics digital twin of a humanoid robot.

[0026] S1.1: Obtain the three-dimensional geometric model data and physical characteristic parameters of the humanoid robot.

[0027] Specifically, three-dimensional geometric model data, including the shape, size, and assembly relationships of each component, is extracted from the computer-aided design (CAD) files of the humanoid robot. Simultaneously, physical property parameters, including mass, moment of inertia, center of mass location, and material stiffness, are obtained from material databases or experimental test reports.

[0028] It should be noted that the 3D geometric model data is stored in a standard format (such as STEP or IGES); physical property parameters are recorded in the form of structured data tables to ensure data integrity and accuracy.

[0029] S1.2: Import the three-dimensional geometric model data and physical property parameters into the dynamic simulation software, and define the joint constraints and range of motion of the humanoid robot based on the mechanical structure of the humanoid robot.

[0030] Specifically, launch the dynamics simulation software (such as Adams or Simulink). The 3D geometric model data and physical property parameters are loaded using the data import function of the dynamics simulation software. Within the dynamics simulation software, based on the humanoid robot's mechanical structure, the type of each joint (e.g., rotary or translational joint), degrees of freedom, range of motion (e.g., minimum and maximum angle or displacement limits), and constraints (e.g., joint limits and collision detection) are set. When setting the type of each joint, the dynamics simulation software automatically identifies the joint type and connection relationships by importing the robot's CAD model or URDF file. It then infers the degrees of freedom and range of motion based on predefined templates or learning algorithms, while automatically generating constraints to reduce manual intervention. This ensures the simulation model conforms to the physical structure and possesses efficient computational adaptability. It should be noted that predefined templates refer to the software's built-in parameter library of common joints (such as hinges and sliders), which are automatically applied by matching the model's geometric features. Learning algorithms, on the other hand, analyze joint motion trajectory data (such as learning from CAD assembly constraints or motion capture) and automatically infer their type and range of motion using machine learning (such as neural networks) or geometric inference algorithms.

[0031] S1.3: Determine the deployment location and field of view parameters of the distributed vision sensor network based on the joint constraints and motion range of the humanoid robot, and construct an interactive dynamic digital twin in the dynamic simulation software.

[0032] Specifically, based on the joint constraints and range of motion of the humanoid robot, the possible motion postures and workspace of the humanoid robot are analyzed. Coverage analysis algorithms (such as frustum calculation) are used to calculate the optimal deployment locations (e.g., head, shoulder, or waist) and field-of-view parameters (e.g., horizontal and vertical field of view) of the distributed vision sensor network to ensure that the sensors can fully cover the humanoid robot's interaction environment. The optimal deployment location refers to the placement point that maximizes the coverage of the target area while minimizing resource consumption.

[0033] It should be noted that, in the process of analyzing the possible motion postures and workspace of the humanoid robot, inverse kinematics is used or Monte Carlo random sampling is performed in the workspace to simulate all reachable poses of the humanoid robot under joint constraints, thereby determining its workspace boundary in point cloud form.

[0034] In dynamics simulation software, the digital representation of a distributed visual sensor network is integrated into the humanoid robot dynamics model, and sensor characteristics (such as resolution and frame rate) are added to complete the construction of an interactive dynamics digital twin. The interactive dynamics digital twin includes the humanoid robot body model, dynamic attributes, and sensor network.

[0035] It should be noted that the humanoid robot body model is a mathematical description of the robot's mechanical structure (including physical properties such as links, joints, mass, and inertia) based on the dynamics theory of multi-rigid-body systems in robotics. The digital representation of a distributed visual sensor network refers to a virtual software entity that accurately describes the geometric deployment, functional characteristics, and data flow of a physical sensor network within a digital twin or simulation environment.

[0036] After integrating the humanoid robot dynamics model, the motion data of the actual robot joints need to be recorded through a motion acquisition system. Based on the principle of minimizing the error between the predicted torque and the actual driving torque of the humanoid robot dynamics model, the dynamic parameters (such as link mass, inertia tensor and friction coefficient) are iteratively optimized to achieve consistency calibration between the humanoid robot dynamics model and the physical entity.

[0037] It should also be noted that the deployment location of the distributed visual sensor network is determined in three-dimensional coordinates; the field of view parameter is specified in angle values; and the interactive dynamics digital twin is represented in the simulation software as a virtual model that can simulate dynamic interaction and visual perception.

[0038] S2: Simulate interactive dynamics digital twins in a virtual environment and acquire real-time relative spatial relationship data through a distributed visual sensor network.

[0039] S2.1: Synchronously capture multi-view environmental image streams through a distributed visual sensor network.

[0040] Specifically, in the distributed visual sensor network, each visual sensor continuously acquires two-dimensional color images at a fixed frame rate; at the same time, in order to achieve time synchronization of multi-view image streams, a precision clock synchronization protocol is adopted to ensure that the image acquisition actions of all visual sensors in the distributed visual sensor network remain highly consistent, laying the foundation for accurate collaborative analysis of multi-view data; each visual sensor transmits the acquired image stream to the central processing unit (CPU) after adding a unified timestamp.

[0041] S2.2: Perform stereo matching on multi-view environmental image streams to generate environmental depth point cloud data.

[0042] Specifically, after receiving a multi-view image stream with timestamps, the central processing unit selects images from any two visual sensors as a stereo pair. By calculating the disparity value of each pixel in the stereo pair, it converts the two-dimensional image coordinates into three-dimensional spatial coordinates using the principle of triangulation. Specifically, a semi-global stereo matching algorithm is used to calculate the disparity value of each pixel. Based on the known baseline distance between the visual sensors and the intrinsic parameter matrix of the visual sensors, the three-dimensional coordinate points corresponding to each pixel are calculated through perspective geometry, generating environmental depth point cloud data composed of several three-dimensional coordinate points.

[0043] It should be noted that the mathematical foundation of perspective geometry is perspective projection transformation, which describes the process of projecting from the three-dimensional world to a two-dimensional image, and can inversely reconstruct the three-dimensional spatial structure from two-dimensional image information through computational rules and constraints. The baseline distance is the geometric baseline length of the three-dimensional measurement, while the intrinsic parameter matrix of the vision sensor represents the mapping rules for camera imaging.

[0044] S2.3: Register the environmental depth point cloud data with the humanoid robot body model, calculate the relative distance and angle between the humanoid robot and environmental objects, and output real-time relative spatial relationship data.

[0045] Specifically, the environmental depth point cloud data and the 3D model of the humanoid robot body are placed in the same world coordinate system. An iterative nearest-point algorithm is used to achieve precise registration between the environmental depth point cloud data and the surface of the humanoid robot body model. Furthermore, environmental depth point cloud data is defined as... The point set of the humanoid robot body model is ;for and Establish initial correspondence: For Each environment point to be registered ,exist Find the nearest neighbor This forms a set of corresponding point pairs. The expression is: .

[0046] The optimal rotation matrix is ​​calculated by minimizing the Euclidean distance error between corresponding point pairs. and optimal translation vector And iteratively solve the problem; the expression is: ; In the formula, This represents the first point in the environmental depth point cloud data. One environment point to be registered. The first point in the humanoid robot body model point set represents the... The nearest neighbor point, This represents the Euclidean distance (L2 norm). This indicates that for all corresponding point pairs Summation, To optimize the objective, we mean finding the optimal... and To minimize the error; Each iteration updates the set of corresponding point pairs. and recalculate and The process continues until the error is less than the ICP registration convergence threshold or the maximum number of iterations is reached, at which point accurate registration is complete. The ICP registration convergence threshold is specifically determined based on the geometric accuracy requirements of the humanoid robot body model surface and the resolution of the environmental depth point cloud data, using the change in the Frobenius norm of two adjacent transformation matrices during the iteration process as the convergence criterion.

[0047] It should be noted that the maximum number of iterations represents a safety limit set based on computational budget and real-time requirements to avoid infinite loops in the iteration process due to poor point cloud data quality or excessive initial pose deviation. The safety limit is determined based on the ratio of the maximum computation time to the average time of a single iteration.

[0048] After registration, environmental depth point cloud data The data is transformed into a new point set to ensure the accuracy of the registered environmental depth point cloud data. Point set of humanoid robot body model Precise alignment; The expression for the registered environmental depth point cloud data set is: ; In the formula, Indicates to Apply the optimal and This enables the registered environmental depth point cloud data set to be used for registration. Point set of humanoid robot body model Precise alignment; After accurate registration, a predefined set of key points is extracted from the surface of the humanoid robot body model. (e.g., hand tips and foot tips); for each key point In the registered environmental depth point cloud data set Search for nearest neighbor (i.e., the point with the smallest Euclidean distance) is used as the relative distance; the relative distances of all key points are stored as a distance dataset; The relative distance is calculated and expressed as: ; In the formula, Indicates the first The relative distances between key points and environmental objects. For index variables of key points, The first surface of the humanoid robot body model represents the... One key point, Indicating environmental depth point cloud data with nearest neighbor; It should be noted that, , and The surface of the humanoid robot body model is represented by the first... Key points Three-dimensional coordinates in the world coordinate system; , and This indicates that in the registered environmental depth point cloud data, the key points... nearest neighbor The three-dimensional coordinates, 、 and These represent the component values ​​along the three mutually perpendicular coordinate axes in a three-dimensional spatial coordinate system.

[0049] For each key point and its corresponding nearest neighbor Calculate the key points separately Normal vector and nearest neighbor Normal vector ; normal vector and Surface normals are estimated using the surface geometry of the humanoid robot model and the registered environmental depth point cloud data (normal estimation is based on the existing triangular mesh method). The angles between the normal vectors are calculated, and all the angles between the normal vectors are stored as an angle dataset. The angle between the normal vectors is calculated using the following expression: ; In the formula, Indicates the first The angle between the key point and the normal vector of the surface of the surrounding object. The magnitude of the normal vector is represented (all values ​​are 1 after normalization). It is the inverse cosine function, used to calculate the angle between two normal vectors; Distance and angle datasets are integrated using data fusion algorithms (such as Kalman filtering or weighted averaging) to generate real-time relative spatial relationship data. This data is output in a structured format (such as a table or vector) containing the relative distance and angle values ​​for each keypoint.

[0050] S3: Based on real-time relative spatial relationship data, predict the contact event type and contact force vector, and calculate the expected kinetic energy loss value generated by external interaction; the prediction of contact event type and contact force vector is achieved by forward inference through a pre-trained deep learning model of contact mechanics based on real-time relative spatial relationship data.

[0051] It should be noted that the deep learning model for contact mechanics must be trained before deployment. The training data specifically includes real-time relative spatial relationship data (including distance and angle values) and corresponding labels for real contact event types (such as rigid collisions and sliding friction), as well as contact force vector measurements (three-dimensional force and torque). The deep learning model for contact mechanics employs a multi-layer fully connected neural network structure. The input layer contains real-time relative spatial relationship data, and the output layer consists of two branches: a contact event type classification branch and a contact force vector regression branch. Specifically, the contact event type classification branch outputs a discrete probability distribution through the Softmax activation function, representing the probability of occurrence of each type of contact event; the contact force vector regression branch outputs a three-dimensional contact force vector through a linear activation function; the training process adopts supervised learning, and the loss function is the weighted sum of classification cross-entropy and regression mean square error.

[0052] S3.1: Determine the corresponding contact dynamics equations based on the predicted contact event types.

[0053] Specifically, real-time relative spatial relationship data is input into a pre-trained deep learning model of contact mechanics to perform forward inference; the contact event type classification branch outputs the contact event type with the highest probability, including contact event type and sliding friction, and the contact force vector regression branch outputs the contact force vector. Based on the predicted contact event type, a predefined contact dynamics equation is selected. It should be noted that the predefined contact dynamics equations refer to mathematical relationships with partially assigned parameters that are explicitly set based on physical laws for each type of contact event (such as collision and sliding) before the algorithm runs. These equations are used to convert the predicted contact force vectors into specific dynamic effects (such as kinetic energy loss). Rigid collisions are defined as instantaneous high-impact contact (such as a humanoid robot's leg colliding with an obstacle); sliding friction is defined as continuous relative surface motion contact (such as a hand sliding along a table).

[0054] If it is a rigid collision, then the impulse model equation is used, and the expression is: ; In the formula, It is a three-dimensional vector representing the change in velocity. ), Let be a scalar, representing the local mass of the humanoid robot. The symbol for definite integral is . Indicates the moment of contact initiation. The duration of instantaneous contact. This is a three-dimensional vector, representing the predicted contact force vector. Represents the time differential variable; If it is sliding friction, then the Coulomb friction equation is used, and the expression is: ; In the formula, This is a three-dimensional vector, representing the friction force vector. A negative sign indicates that the direction of friction is opposite to the direction of motion. Indicates the coefficient of friction. Represents the absolute value of the normal component of the contact force. Represents a three-dimensional unit vector, a unit vector along the tangential direction of motion; coefficient of friction The materials are obtained from a material pairing database, which is based on the surface materials that the humanoid robot may come into contact with (such as the material of the robot's feet and the type of flooring), and is derived from a pre-compiled table of friction coefficients. Values. For example, common pairings include rubber-concrete (μ≈0.6−0.8) and metal-plastic (μ≈0.2−0.3), etc.

[0055] It should be noted that the impulse model equation is used for rigid collisions because the impulse model equation directly relates momentum changes through the integral of force over time, which can efficiently calculate the kinetic energy loss caused by instantaneous collisions; the Coulomb friction equation is used for sliding frictions because the Coulomb friction equation quantifies energy dissipation through the friction coefficient and normal force, which can accurately predict the kinetic energy decay during continuous sliding.

[0056] The impulse model equation simplifies the complex dynamic analysis of transient collisions, while the Coulomb friction equation provides a stable and physically interpretable calculation of frictional energy consumption. The combination of the two can cover the kinetic energy loss prediction needs of humanoid robots in the main contact scenarios.

[0057] S3.2: Input the predicted contact force vector into the contact dynamics equation to calculate the work done by the instantaneous contact force.

[0058] Specifically, the contact force vector Substituting the selected contact dynamics equations, the instantaneous power is calculated, and the expression is: ; In the formula, Indicates time Instantaneous power, Indicates time The predicted contact force vector, This represents the dot product operation. The velocity vector at the point of contact at any given moment is the sum of the products of the corresponding vector components. Furthermore, The instantaneous velocity of the humanoid robot at the point of contact (provided by real-time simulation from an interactive dynamics digital twin); the work done by the instantaneous contact force. This is the integral of power over time; Work done by instantaneous contact force The expression is: ; In the formula, This indicates the step size (preset constant) for dynamic simulation. Let be an integral infinitesimal element, representing an infinitesimal increment with respect to time. Indicates the upper limit of integration, and indicates the time. After a simulation step size The point in time that follows.

[0059] It should be noted that the dynamic simulation step size The preset value is determined based on the joint motion control cycle of the humanoid robot, and the specific value is the minimum time interval of the control command (e.g., 0.001 seconds).

[0060] S3.3: Integrate the work done by the instantaneous contact force within the predicted contact time interval and output the expected kinetic energy loss value.

[0061] Set the predicted contact time interval (The contact duration is output by the deep learning model of contact mechanics) Within the predicted contact time range Internal accumulation, output expected kinetic energy loss value The expression is: ; In the formula, The moment of contact initiation. This is the moment when contact ends.

[0062] In contrast to traditional prediction methods based on rules or simple physical models, the end-to-end inference prediction method based on three-dimensional spatial relationship data of accurately registered environmental point clouds and pre-trained contact mechanics deep learning model does not rely on accurate environmental modeling, has stronger generalization ability, and can accurately predict complex contact events such as rigid collisions to sliding friction transitions. The output contact force vector and event type prediction results have higher accuracy.

[0063] S4: Embed the expected kinetic energy loss value as a dynamic optimization target into the online motion planner to generate an adaptive motion trajectory that can avoid high-energy-consuming contact events.

[0064] S4.1: The online motion planner receives the expected kinetic energy loss value as the dynamic optimization target.

[0065] Specifically, the expected kinetic energy loss value, as a core parameter of the dynamic optimization objective, is loaded into the optimization objective queue of the online motion planner (such as the Model Predictive Control (MPC) framework or an optimizer based on Sequential Quadratic Programming (SQP), with its priority set as a real-time optimization objective. The expected kinetic energy loss value is then marked as a minimization objective in subsequent optimization processes.

[0066] S4.2: Call the predefined humanoid robot motion task constraints and the dynamic constraints of the interaction dynamics digital twin, and construct a multi-objective optimization function based on the online motion planner.

[0067] Specifically, the online motion planner reads humanoid robot motion task constraint data (including end effector target position, posture accuracy requirements, and motion completion time limits) from the storage block, and extracts dynamic constraint data (including joint torque upper limit, joint velocity limit, and zero torque point stability region) from the interactive dynamics digital twin. A multi-objective optimization function is constructed based on the expected kinetic energy loss value, predefined humanoid robot motion task constraint data, and dynamic constraint data, specifically as follows: The online motion planner sets the expected kinetic energy loss value as the core dynamic optimization objective to be minimized; simultaneously, it loads predefined humanoid robot motion task constraint data, quantifying the end effector target position deviation, posture angle error, and motion completion time deviation as task violation penalty items; it also extracts dynamic constraint data from the interactive dynamics digital twin, quantifying the difference between the actual joint torque exceeding the torque limit, the difference between the actual joint speed exceeding the joint speed limit, and the deviation distance of the actual position of the zero torque point from the stable region as dynamic violation penalty items; and linearly superimposes the expected kinetic energy loss value objective item, the task violation penalty item, and the dynamic violation penalty item according to preset weight coefficients to form a multi-objective optimization function. It should be noted that the multi-objective optimization function aims to minimize the weighted sum, achieving coordinated optimization of energy consumption targets, tasks, and physical constraints. Preset weight coefficients are determined by standardization methods to unify dimensions and are allocated according to objective priority. Specifically, objective function normalization is used to balance the differences in the numerical ranges of each optimization term, and then weight values ​​are allocated based on a preset proportional relationship between the humanoid robot's task requirements and physical limitations.

[0068] S4.3: Solve the multi-objective optimization function using a numerical optimization algorithm to generate candidate motion trajectories.

[0069] Specifically, a sequential quadratic programming algorithm is used to solve the constructed multi-objective optimization function. Using the current joint state of the humanoid robot as the initial value, an iterative search is performed in the state space. Each iteration generates candidate solutions for joint trajectory parameters. The dynamic feasibility of the trajectory is verified using an interactive dynamics digital twin. Candidate solutions that satisfy all constraints are stored as candidate motion trajectories. Each candidate motion trajectory contains time-series data of joint angles, velocities, and accelerations.

[0070] S4.4: Select the candidate motion trajectory with the minimum expected kinetic energy loss value from the candidate motion trajectories as the adaptive motion trajectory.

[0071] Specifically, all candidate motion trajectories are traversed, and the expected kinetic energy loss value corresponding to each candidate motion trajectory is extracted. The magnitudes of the expected kinetic energy loss values ​​of each candidate motion trajectory are compared, and the candidate motion trajectory with the smallest value is selected as the final output. When multiple candidates have the same minimum expected kinetic energy loss value, the candidate motion trajectory with the smallest degree of violation of the humanoid robot motion task constraints is selected.

[0072] In a superior manner, by comparing the deviation between expected and actual kinetic energy loss, a reward signal is generated in real time, which synchronously drives the online iterative update of the prediction algorithm parameters of the contact mechanics model and the multi-objective optimization weights of the motion planner. This can continuously correct prediction errors and adjust planning preferences (such as placing more emphasis on collision avoidance or tolerating slight slippage), thereby continuously reducing cumulative energy consumption in long-term execution and achieving self-evolution and performance improvement that traditional static design or control methods cannot achieve.

[0073] S5: During the process of the humanoid robot executing the adaptive motion trajectory, the distributed visual sensor network monitors the environmental interaction and collects the actual kinetic energy loss value.

[0074] Specifically, the online motion planner converts the adaptive motion trajectory into joint control commands, which are then transmitted to the humanoid robot's underlying motion controller via a real-time communication interface. The joint actuators of each joint of the humanoid robot receive the control commands and execute trajectory tracking. As the humanoid robot begins executing the adaptive motion trajectory, a distributed visual sensor network synchronously acquires multi-view interactive image streams at a fixed sampling frequency. Real-time stereo matching of the acquired multi-view interactive image streams is performed to reconstruct the three-dimensional geometric deformation data of the contact area between the humanoid robot and the environment at the current moment, while reading the actual movement speed of the contact point fed back by the humanoid robot's joint encoder. The pressure distribution on the contact surface is calculated based on the three-dimensional geometric deformation data. Combined with the actual movement speed of the contact point, the change in kinetic energy of the humanoid robot's mass during the duration of the contact event is calculated using the rigid body kinetic energy theorem. It should be noted that the absolute value of the change in kinetic energy is the actual kinetic energy loss value, which is recorded by the main controller and associated with the corresponding timestamp; the continuous acquisition of the actual kinetic energy loss value covers the entire adaptive motion trajectory execution cycle, forming a time series energy consumption dataset for reinforcement learning optimization in subsequent steps.

[0075] S6: Input the energy consumption deviation between the expected kinetic energy loss value and the actual kinetic energy loss value into the reinforcement learning model, and iteratively optimize the prediction accuracy of the contact event type and the motion planning parameters of the online motion planner; the pre-training process of the reinforcement learning model includes using the energy consumption deviation value as a reward signal, updating the prediction algorithm parameters of the contact event type and the motion planning parameters of the online motion planner through the policy gradient method, and minimizing the long-term cumulative energy consumption deviation.

[0076] S6.1: Input the energy consumption deviation value into the reinforcement learning model as a reward signal for the current state.

[0077] Specifically, the absolute difference between the expected kinetic energy loss and the actual kinetic energy loss is used as the energy consumption deviation value. This deviation value is input into the reward function, calculated, and converted into a standardized reward signal. The reward signal characterizes the performance of the current action strategy and is used to drive the parameter optimization direction of the reinforcement learning model.

[0078] S6.2: The reinforcement learning model calculates the parameter adjustments for the prediction algorithm parameters and motion planning parameters based on the reward signal; it then applies these parameter adjustments to update the prediction algorithm parameters and the online motion planner.

[0079] Reinforcement learning models employ a policy gradient algorithm based on reward signals. To optimize guidance; For the prediction algorithm parameters of the deep learning model of contact mechanics (i.e., neural network weights) ), calculate policy gradient The expression is: ; In the formula, For the gradient operator, represents the partial derivative with respect to the objective function. This represents the prediction algorithm parameters of the deep learning model for contact mechanics. The objective function for reinforcement learning represents the expected long-term cumulative reward. The expectation operator represents the expected value of the expression within the parentheses. It is the natural logarithm function. This is represented as a prediction strategy for contact event types (i.e., the action probability distribution output by the contact mechanics deep learning model). For actions, it refers to the predicted type of contact event (such as rigid collision). The state refers to real-time relative spatial relationship data (normalized distance and angle values). express (Prediction) is an abbreviation that specifically refers to the predictive function of a deep learning model for contact mechanics. For the motion planning parameters (i.e., the weight coefficients of the multi-objective optimization function) of the online motion planner, the policy gradient is calculated as follows: ; In the formula, This represents the motion planning parameters of the online motion planner. The strategy for generating the trajectory of the final payment (i.e., the decision probability distribution of the candidate trajectory generated by the online motion planner). yes (Planning) is an abbreviation for the motion trajectory generation function of an online motion planner; Update parameters using gradient ascent: ; ; In the formula, This is the assignment operator, indicating that the left-hand parameter is updated with the result of the right-hand calculation. For the policy gradient of the parameters of the contact mechanics deep learning model, The policy gradient for the parameters of the online motion planner. The learning rate is determined by balancing convergence speed and numerical stability under task requirements using a grid search method. The example value range is a real number greater than 0 and less than 1, such as a typical value of 0.01. Updated prediction algorithm parameters Load the contact mechanics deep learning model, overriding the original parameters; update the motion planning parameters. Load the online motion planner, covering the weight coefficients of the multi-objective optimization function.

[0080] Better yet, by iteratively reducing long-term cumulative energy consumption deviation, the accuracy of contact event type prediction and the energy consumption optimization capability of motion trajectory can be improved.

[0081] This embodiment also provides a computer device applicable to the humanoid robot body design method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the humanoid robot body design method proposed in the above embodiment.

[0082] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0083] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the humanoid robot body design method proposed in the above embodiments. 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.

[0084] In summary, this invention, by embedding the expected kinetic energy loss value as a dynamic optimization objective into the online motion planner to generate adaptive motion trajectories, enables real-time optimization of the motion planning process. This minimizes kinetic energy loss while ensuring the fulfillment of task constraints and dynamic limitations. The integration of the expected kinetic energy loss value allows for the pre-assessment of the energy consumption impact of contact events in the virtual environment, thereby avoiding high-energy-consuming interactions, improving energy utilization efficiency, further enhancing the robot's motion endurance and overall performance, adapting to long-term operational needs in complex environments, and continuously improving trajectory generation quality through an iterative optimization mechanism.

[0085] It should be noted that 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for designing a humanoid robot body, characterized in that: include, Constructing an interactive dynamics digital twin of a humanoid robot; Simulate interactive dynamics digital twins in a virtual environment and acquire real-time relative spatial relationship data through a distributed visual sensor network; Based on real-time relative spatial relationship data, predict the contact event type and contact force vector, and calculate the expected kinetic energy loss value generated by external interaction; The expected kinetic energy loss value is embedded as a dynamic optimization target into the online motion planner to generate an adaptive motion trajectory that can avoid high-energy-consuming contact events. During the process of the humanoid robot executing an adaptive motion trajectory, the environmental interaction is monitored through a distributed visual sensor network to collect the actual kinetic energy loss value; The energy consumption deviation between the expected kinetic energy loss value and the actual kinetic energy loss value is input into the reinforcement learning model to iteratively optimize the prediction accuracy of the contact event type and the motion planning parameters of the online motion planner.

2. The humanoid robot body design method as described in claim 1, characterized in that: The specific steps for constructing the interactive dynamics digital twin of the humanoid robot are as follows: Acquire the three-dimensional geometric model data and physical characteristic parameters of the humanoid robot; Import the three-dimensional geometric model data and physical property parameters into the dynamic simulation software, and define the joint constraints and range of motion of the humanoid robot based on the mechanical structure of the humanoid robot; Based on the joint constraints and range of motion of the humanoid robot, the deployment location and field of view parameters of the distributed vision sensor network are determined, and an interactive dynamic digital twin is constructed in the dynamic simulation software.

3. The humanoid robot body design method as described in claim 2, characterized in that: The specific steps for obtaining real-time relative spatial relationship data are as follows. Simultaneous capture of multi-view environmental image streams via a distributed visual sensor network; Perform stereo matching on multi-view environmental image streams to generate environmental depth point cloud data; The environmental depth point cloud data is registered with the humanoid robot body model to calculate the relative distance and angle between the humanoid robot and environmental objects, and output real-time relative spatial relationship data.

4. The humanoid robot body design method as described in claim 3, characterized in that: The predicted contact event type and contact force vector are achieved through forward inference using a pre-trained deep learning model of contact mechanics, based on real-time relative spatial relationship data.

5. The humanoid robot body design method as described in claim 4, characterized in that: The specific steps for calculating the expected kinetic energy loss value generated by external interaction are as follows. Based on the predicted type of contact event, determine the corresponding contact dynamics equations; Input the predicted contact force vector into the contact dynamics equation to calculate the work done by the instantaneous contact force; Integrate the work done by the instantaneous contact force within the predicted contact time interval to output the expected kinetic energy loss value.

6. The humanoid robot body design method as described in claim 5, characterized in that: The specific steps for generating an adaptive motion trajectory that avoids high-energy-consuming contact events are as follows. The online motion planner receives the expected kinetic energy loss value as a dynamic optimization target; Invoke predefined humanoid robot motion task constraints and interaction dynamics digital twin dynamic constraints, and construct a multi-objective optimization function based on an online motion planner; A multi-objective optimization function is solved using a numerical optimization algorithm to generate candidate motion trajectories. The candidate motion trajectory with the smallest expected kinetic energy loss value is selected from the candidate motion trajectories and used as the adaptive motion trajectory.

7. The humanoid robot body design method as described in claim 6, characterized in that: The pre-training process of reinforcement learning models includes using energy consumption bias as a reward signal, updating the prediction algorithm parameters for contact event types and the motion planning parameters of the online motion planner through the policy gradient method, and minimizing the long-term cumulative energy consumption bias.

8. The humanoid robot body design method as described in claim 7, characterized in that: The prediction accuracy of contact event types and the motion planning parameters of the online motion planner are optimized iteratively through a reinforcement learning model. The specific steps are as follows: The energy consumption deviation value is input into the reinforcement learning model as a reward signal for the current state; The reinforcement learning model calculates parameter adjustments for the prediction algorithm parameters and motion planning parameters based on the reward signal; the parameter adjustments are then applied to update the prediction algorithm parameters and the online motion planner.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the humanoid robot body design method according to any one of claims 1 to 8.

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 implements the steps of the humanoid robot body design method according to any one of claims 1 to 8.

Images