Whole-body dynamics real-time online optimization method and system for humanoid robot

Abstract

The invention provides a whole-body dynamics real-time online optimization method and system for a humanoid robot, and the method comprises the steps: receiving real-time joint signals of whole-body joints of the humanoid robot, fusing the real-time joint signals with motion historical optimization data of the same type to generate fused motion signals, and inputting the fused motion signals into a pre-trained hierarchical dynamics optimization model to obtain an initial joint torque distribution scheme; and inputting the initial scheme into a motion simulation layer, loading current scene environment constraint data, and outputting a motion attitude feedback signal and torque distribution deviation traceability information. Positioning a core deviation joint and an associated joint chain in combination with deviation traceability information, and adjusting the weight of the model in combination with a feedback signal; and inputting the fused motion signal into the adjusted model, and outputting an optimized joint torque distribution scheme through a built-in joint-crossing torque coupling optimization unit. And finally, the optimization scheme is transmitted to a joint driving system, matched data are stored in a historical optimization database, dynamic online optimization is completed, and the motion performance of the humanoid robot is improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and more specifically, to a method and system for real-time online optimization of the whole-body dynamics of a humanoid robot. Background Technology

[0002] In the field of humanoid robot technology, achieving efficient, stable, and adaptable motion control across various scenarios is a key objective. Currently, humanoid robot dynamics control methods suffer from the following main problems. Firstly, traditional methods often rely solely on real-time joint signals for control, lacking full utilization of historical data. This prevents robots from quickly optimizing their motion strategies based on past experience when facing complex or similar motion scenarios, resulting in low motion efficiency and difficulty adapting to diverse task requirements. Secondly, existing dynamics optimization methods lack effective feedback and adjustment mechanisms after generating joint torque allocation schemes. If the initial scheme deviates in the actual scenario, it becomes impossible to promptly locate and accurately adjust problematic joints, thus affecting the robot's posture and overall performance. Furthermore, the coupling relationship between joints is not fully considered during joint torque allocation, easily leading to motion conflicts between joints and reducing the stability and coordination of robot motion. Therefore, a novel real-time online optimization method for whole-body dynamics is urgently needed to address these issues. Summary of the Invention

[0003] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for real-time online optimization of the whole-body dynamics of a humanoid robot, the method comprising: The system receives real-time joint signals from all joints of the humanoid robot, integrates historical optimization data of similar movements of the humanoid robot to generate a fused motion signal, and inputs the fused motion signal into a pre-trained hierarchical dynamics optimization model to generate an initial joint torque allocation scheme. The initial joint torque allocation scheme is input into the motion simulation layer and the current scene environment constraint data is loaded. The motion simulation layer outputs motion posture feedback signals and torque allocation deviation traceability information. By combining the torque distribution deviation source information to locate the core deviation joint and related joint chain, and by combining the motion posture feedback signal and the deviation location result to adjust the weight of the hierarchical dynamic optimization model, the adjusted hierarchical dynamic optimization model is obtained. After adjusting the fused motion signal input, the hierarchical dynamic optimization model outputs the optimized joint torque distribution scheme through the built-in cross-joint torque coupling optimization unit. The optimized joint torque distribution scheme is transmitted to the joint drive system, and the matching data of the optimized joint torque distribution scheme and real-time joint signals is stored in the historical optimization database to complete the online dynamics optimization.

[0004] In another aspect, embodiments of the present invention also provide a real-time online optimization system for the whole-body dynamics of a humanoid robot, including a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the machine-readable storage medium to implement the above-described method.

[0005] Based on the above, this embodiment of the invention generates a fused motion signal by fusing real-time joint signals with historical optimization data, and inputs it into a pre-trained model to obtain an initial joint torque allocation scheme. This fully utilizes historical experience to improve the rationality and adaptability of the initial scheme. Next, the initial scheme is input into the motion simulation layer, combined with current scene environmental constraint data, to output motion posture feedback signals and torque allocation deviation traceability information. This accurately locates problematic joints and related joint chains, providing an accurate basis for model weight adjustment. Then, the model weights are adjusted based on the feedback information to obtain the adjusted model, which is then input into the fused motion signal to generate an optimized joint torque allocation scheme, achieving effective optimization of the initial scheme. Finally, the optimized scheme is transmitted to the joint drive system and the matching data is stored, completing online optimization. This enables the robot to adapt to different scenarios in real time, improving motion stability, coordination, and efficiency, and enhancing the humanoid robot's adaptability and task execution capabilities in complex environments. Attached Figure Description

[0006] Figure 1 This is a schematic diagram of the execution flow of the real-time online optimization method for the whole-body dynamics of a humanoid robot provided in an embodiment of the present invention.

[0007] Figure 2 This is a schematic diagram of exemplary hardware and software components of the real-time online optimization system for the whole-body dynamics of a humanoid robot provided in an embodiment of the present invention. Detailed Implementation

[0008] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a real-time online optimization method for the whole-body dynamics of a humanoid robot according to an embodiment of the present invention. The following is a detailed description of this real-time online optimization method for the whole-body dynamics of a humanoid robot.

[0009] Step S110: Receive real-time joint signals from all joints of the humanoid robot, fuse historical optimization data of similar movements of the humanoid robot to generate a fused motion signal, and input the fused motion signal into a pre-trained hierarchical dynamics optimization model to generate an initial joint torque allocation scheme.

[0010] In this embodiment, the first step involves receiving real-time joint signals from all joints of the humanoid robot and fusing them with historical data to generate a fused motion signal, thereby obtaining an initial joint torque allocation scheme. This is illustrated using a fixed scenario of the humanoid robot walking on flat ground as an example. In this scenario, the robot needs to complete continuous walking movements, involving the coordinated movement of multiple joints throughout the body. First, the system receives real-time joint signals from all joints of the robot. These signals include information such as the rotation angle, movement rate, and force feedback of each joint. Each joint signal exists in the form of multi-dimensional data; for example, the rotation angle data records the angle changes of the joint at different times, the movement rate data reflects the speed of joint rotation, and the force feedback data reflects the magnitude and direction of the force experienced by the joint during movement. Next, the system matches historical optimization data from the historical optimization database that is similar to the current flat ground walking movement. This historical optimization data consists of optimized joint torque allocation data generated when the robot previously performed flat ground walking movements. Then, the system fuses the real-time joint signals and historical optimization data according to a preset scenario adaptation weight rule to generate a fused motion signal. During the fusion process, different weights are assigned to real-time joint signals and historical optimization data based on the characteristics of the current flat-ground walking scenario. For example, joint angle and motion rate data with high real-time requirements are given higher weights, while stable torque distribution experience data from historical data are also given certain weights. Through this weighted fusion method, the advantages of real-time and historical data are combined to form a more comprehensive fused motion signal. The fused motion signal is then input into a pre-trained hierarchical dynamics optimization model. This model has been trained on a large amount of flat-ground walking and similar scenario data and can process the input fused motion signal to output an initial joint torque distribution scheme.

[0011] Step S111: Receive real-time joint signals including joint rotation angle, movement rate and force feedback, and match them with historical optimization data of the same type of movement.

[0012] In this embodiment, during the flat ground walking scenario, each joint of the robot, such as the hip, knee, ankle, shoulder, and elbow joints, generates corresponding real-time joint signals. For example, in the real-time joint signals of the hip joint, rotation angle data records the sequence of angle changes from the initial position to the maximum extension position and then to the retracted position during walking; motion rate data reflects the speed of rotation of the hip joint at each angle change stage; and force feedback data reflects the forces acting on the hip joint when supporting body weight and propelling the body forward. The system collects these real-time joint signals using sensors located at each joint. The sensors convert the collected analog signals into digital signals and transmit them to the system's data receiving module. Simultaneously, based on the current motion type, i.e., flat ground walking, the system searches the historical optimization database to match all historical optimization data related to flat ground walking. This historical data includes joint torque distribution schemes, corresponding joint motion parameters, and optimization effect evaluation data from previous flat ground walking actions. During the matching process, the system will filter based on the similarity of the movements. For example, it will consider the similarity between the current movement and factors such as the walking speed, stride length, and road surface conditions of the past movements, and select historical data with higher similarity as the matching result.

[0013] Step S112: Fuse real-time joint signals and historical optimization data according to the scene adaptation weight rules to generate fused motion signals.

[0014] In this embodiment, in the flat ground walking scenario, the scenario adaptation weighting rule is formulated based on the characteristics and requirements of joint movement in that scenario. For example, for real-time knee joint signals, since the knee joint plays an important role in support and cushioning during walking, its movement rate and force feedback data are crucial in reflecting the current movement state, and therefore are given a high weight. For empirical knee joint torque allocation data in historical optimization data, since these data are verified effective solutions, they are also given a certain weight, but the weight is relatively lower than that of real-time data. The specific weight allocation will be adjusted according to the function of the joint and the needs of the scenario. For example, for the ankle joint, which needs to adapt to changes in the road surface during walking, the real-time force feedback data will have a higher weight. During the fusion process, the system will weight and combine each data item in the real-time joint signal with the corresponding historical optimization data item. For example, the real-time knee joint movement rate data is multiplied by the corresponding weight, and then the historical knee joint movement rate-related torque allocation data is multiplied by the corresponding weight. These two weighted data are then merged to form the fused movement rate-related data item. By performing such weighted fusion processing on each data item of all joints, a fused motion signal is finally generated. This fused motion signal contains comprehensive information on real-time motion status and historical optimization experience.

[0015] Step S113: Input the fused motion signal into the joint-level signal parsing unit of the hierarchical dynamic optimization model according to limb partitions to obtain the partitioned joint feature encoding results.

[0016] In this embodiment, in a flat-ground walking scenario, the humanoid robot's limbs can be functionally divided into lower limbs, upper limbs, and torso, etc. The lower limbs can be further divided into sub-regions such as thighs, calves, and feet, each containing multiple joints. The joint-level signal analysis units of the hierarchical dynamics optimization model are set up according to these limb regions, with each region corresponding to one joint-level signal analysis unit. The system will split the fused motion signal according to the limb region; for example, fused motion signals related to the lower limbs will be transmitted to the joint-level signal analysis unit corresponding to the lower limbs, and fused motion signals related to the upper limbs will be transmitted to the joint-level signal analysis unit corresponding to the upper limbs. Each joint-level signal analysis unit will process the input fused motion signal of the corresponding region. First, it will extract the basic motion features from the signal, such as the angle change range of the joints, the trend of motion rate changes, and the peak and trough values ​​of force feedback. Then, it will process these basic motion features, transforming the key dynamic information, for example, combining angle changes and motion rates to obtain angular acceleration-related information of the joints, and combining force feedback and motion rates to obtain power-related information of the joints. Next, the transformed key dynamic information is converted into a standardized encoding result. This encoding result is a format that can be recognized and processed by subsequent modules of the model. Each joint-level signal parsing unit of each partition will output the corresponding standardized encoding result. The encoding results of all partitions are summed up to form a set of partition joint feature encoding results.

[0017] Step S1131: Activate the joint-level signal parsing unit set according to limb functional zones to receive the fused motion signals of the corresponding zones.

[0018] In this embodiment, during a flat-ground walking scenario, limb functional zones include lower limb functional zones, upper limb functional zones, and trunk functional zones. Each functional zone has a corresponding joint-level signal parsing unit. The system first activates all joint-level signal parsing units configured according to limb functional zones, such as the lower limb joint-level signal parsing unit, the upper limb joint-level signal parsing unit, and the trunk joint-level signal parsing unit. After activation, these parsing units are in a state of waiting to receive signals. The system transmits the fused motion signals of the corresponding zones, which were previously separated, to each parsing unit. For example, the fused motion signals related to the lower limbs are transmitted to the lower limb joint-level signal parsing unit, which receives these signals and prepares for subsequent processing. During transmission, the system ensures the integrity and accuracy of the signals, avoiding signal loss or errors during transmission.

[0019] For example, step S11311: Start the hierarchical dynamics optimization model that has been jointly trained with multiple scene data and multiple sets of robot body parameters.

[0020] In this embodiment, the training dataset for the hierarchical dynamics optimization model includes joint motion data from multiple scenarios and ontological parameter data from multiple sets of robots with different hardware configurations. The training process employs a transfer learning strategy: the model is first pre-trained using flat-ground walking data from a robot with standard parameters, and then fine-tuned using cross-scenario data from robots with other parameters, ensuring stable output across different scenarios and hardware conditions. The system starts the model by calling the model interface and simultaneously loads the model's partition configuration file, clarifying the input and output requirements of each parsing unit, ensuring that the model can directly receive the split and fused signals after initialization.

[0021] Step S11312: Compare the limb partitions of the joint-level signal analysis unit in the hierarchical dynamics optimization model with the robot functional partitions according to the partitioning criteria, and generate a partition matching comparison table.

[0022] In this embodiment, the joint-level signal analysis unit within the model divides the limbs into zones according to the priority of motor function. The lower limbs are the set of joints involved in weight-bearing and propulsion, the upper limbs are the set of joints involved in balance and assistance, and the trunk is the set of joints involved in posture stabilization. The actual functional zones of the robot are divided according to mechanical structure and control logic. The lower limbs are the weight-bearing joint group, the upper limbs are the balance joint group, and the trunk is the stabilizing joint group. The system uses a rule-matching algorithm to map the model zones to the actual robot zones one-to-one, generating a zone matching lookup table to avoid signal transmission misalignment.

[0023] Step S11313: The fused motion signal is split according to the partition criteria in the partition matching lookup table and transmitted to the joint-level signal parsing unit of the corresponding limb partition.

[0024] In this embodiment, the system splits the fused motion signal according to a lookup table, extracts signals related to the lower limbs, upper limbs, and trunk, and transmits them to the corresponding parsing units. A reliable transmission mechanism is employed during the transmission process, with a partition identifier header added before each sub-signal to ensure that the parsing unit can accurately identify and receive the corresponding signal.

[0025] Step S11314: After receiving the signal, the joint-level signal parsing unit activates the feature processing link of the corresponding depth convolutional architecture.

[0026] In this embodiment, the feature processing links of each parsing unit are customized deep convolutional networks. The lower limb parsing unit activates a link designed for high-frequency force feedback, which can extract features such as the time difference of force peaks and the slope of angle and velocity linkage; the upper limb parsing unit activates a link designed for low-frequency swinging, which can extract features such as swing amplitude stability and velocity smoothness; the torso parsing unit activates a link designed for static posture, which can extract features such as angle offset and posture holding time. After the links are activated, the computational resources of the parsing unit are fully utilized to ensure the real-time performance of feature extraction.

[0027] Step S11315: Output the fused motion signal reception completion signal, and prepare to perform subsequent fused motion signal basic motion feature extraction processing.

[0028] In this embodiment, after each parsing unit receives the signal and activates the link, it outputs a standardized completion signal, which includes information such as partition identifier and signal status. After the system aggregates the completion signals of all partitions, it sends a start command to the feature extraction sublayer to ensure seamless integration of subsequent basic motion feature extraction.

[0029] Step S1132: Extract the basic motion features of the fused motion signal through the feature extraction sublayer of the joint-level signal parsing unit.

[0030] In this embodiment, taking the lower limb joint-level signal parsing unit as an example, its feature extraction sublayer processes the input lower limb fused motion signal. First, the sublayer performs preliminary filtering and organization of the signal data, removing some noise data and retaining valid motion-related data. Then, it extracts basic motion features, including the maximum, minimum, and average values ​​of joint rotation angles, movement speeds, force feedback values, and the trend characteristics of these data over time, such as the slope of angle changes and the slope of speed changes. The basic motion features of each joint are extracted; for example, the basic motion features of the hip joint include its maximum extension angle, maximum flexion angle, average movement speed, and maximum force feedback value during walking, while the basic motion features of the knee joint include its maximum flexion angle, average movement speed, and the trend of force feedback changes. The feature extraction sublayer outputs these basic motion features in a specific format and passes them to the next sublayer for processing.

[0031] Step S1133: The key dynamic information in the basic motion features of the motion signal is transformed and fused through the feature processing sub-layer.

[0032] In this embodiment, the feature processing sublayer receives basic motion features from the feature extraction sublayer and then further processes and transforms these features to extract key dynamic information. For example, for the basic motion features of the hip joint, the feature processing sublayer combines the change in rotation angle with the motion rate to calculate the angular acceleration information of the joint, which reflects the rate of change in the joint's motion state. Simultaneously, it combines force feedback with the motion rate to calculate the power information of the joint, which reflects the energy consumption of the joint during motion. For the basic motion features of the knee joint, the feature processing sublayer combines the change in flexion angle with force feedback to analyze the force changes in the joint during the support and swing phases, and the relationship between these changes and the motion rate, thereby extracting key dynamic information reflecting the cushioning and propulsion functions of the knee joint. Through this processing, the feature processing sublayer transforms the basic motion features into more targeted and representative key dynamic information, which better reflects the dynamic state of the joint.

[0033] Step S1134: The feature coding sublayer converts the basic motion features after the key dynamic information is transformed into standardized coding results.

[0034] In this embodiment, the feature encoding sublayer receives the transformed key dynamic information from the feature processing sublayer and then converts this information into standardized encoding results. The standardized encoding results are encoded in a unified format. During the encoding process, the feature encoding sublayer maps different types of key dynamic information to specific encoding intervals or symbols according to preset encoding rules. For example, for angular acceleration-related information, it is mapped to different encoding values ​​based on its magnitude and trend; for power-related information, it is also mapped to corresponding encoding values ​​based on its numerical range and variation characteristics. The key dynamic information of each joint is converted into corresponding encoding values, and the combination of all joint encoding values ​​forms the standardized encoding result for that region. For example, the standardized encoding result for the lower limb region includes encoding values ​​for joints such as the hip, knee, and ankle. These encoding values ​​are arranged in a certain order to form an encoding sequence, which can completely represent the dynamic characteristics of the joints in that region.

[0035] Step S1135: Summarize the standardized coding results of each limb region to form a set of region joint feature coding results.

[0036] In this embodiment, the system collects the standardized encoding results output by the joint-level signal parsing units corresponding to each limb partition, such as the standardized encoding results for the lower limb partition, upper limb partition, and trunk partition. Then, these standardized encoding results from different partitions are arranged and combined in a preset order to form a complete set of partition joint feature encoding results. This set contains the dynamic feature encoding information of all limb partition joints, comprehensively reflecting the motion state and dynamic characteristics of the robot's entire body joints. During the aggregation process, the system identifies the encoding results of each partition to distinguish the encoding information of different partitions. For example, a specific identifier is added before the encoding results of the lower limb partition, and another identifier is added before the encoding results of the upper limb partition. This allows subsequent modules of the model to accurately identify the encoding information of each partition when processing this set.

[0037] Step S114: Generate global joint moment correlation coefficients by processing the partitioned joint feature encoding results through the system-level association mapping unit.

[0038] In this embodiment, during a flat-ground walking scenario, the system-level association mapping unit receives a set of partitioned joint feature encoding results. This unit's role is to establish the association relationships between joints in different partitions and generate global joint torque association coefficients. First, the system-level association mapping unit analyzes the set of partitioned joint feature encoding results to identify the kinematic coordination relationships between joints in different partitions. For example, during walking, the movement of the hip joint is closely related to the movement of the knee joint, and the movement of the ankle joint influences the movement of the knee joint. The swinging of the upper limbs also coordinates with the movement of the lower limbs. Then, based on these coordination relationships, the degree of torque association between different joints is calculated. For example, the influence of hip joint torque changes on knee joint torque changes, and the influence of knee joint torque changes on ankle joint torque changes, etc. These degrees of association are quantified into specific association coefficients, each representing the magnitude of the correlation between torque changes between two joints. Finally, all the association coefficients between joints are combined to form a global joint torque association coefficient. This coefficient reflects the mutual influence and coordination relationships between the torques of all joints throughout the body.

[0039] Step S115: Combining the joint hardware constraint parameters, the initial joint torque distribution scheme is output through the torque distribution generation unit.

[0040] In this embodiment, the joint hardware constraint parameters include the maximum torque output value, minimum torque output value, and torque output accuracy range for each joint. These parameters are determined by the hardware performance of the robot joints. The torque allocation generation unit receives the global joint torque correlation coefficient and processes it in conjunction with the joint hardware constraint parameters. First, the torque allocation generation unit determines the torque allocation ratio between each joint based on the global joint torque correlation coefficient. For example, based on the correlation coefficient between the hip and knee joints, it determines the torque ratio that the hip and knee joints should receive during walking. Then, in conjunction with the joint hardware constraint parameters, these torque allocation ratios are adjusted. For example, if the torque allocation ratio of a joint exceeds its maximum torque output value, the torque allocation ratio of that joint will be appropriately reduced, and the excess torque will be allocated to other joints that can bear the load. If the torque allocation ratio of a joint is lower than its minimum torque output value, the torque allocation ratio of that joint will be appropriately increased to ensure normal joint movement. Through such adjustments, an initial joint torque distribution scheme is generated. This scheme assigns a specific torque value to each joint, ensuring that all joints can work together to complete the walking motion on flat ground.

[0041] Step S120: Input the initial joint torque allocation scheme into the motion simulation layer and load the current scene environment constraint data. The motion simulation layer outputs motion posture feedback signals and torque allocation deviation traceability information.

[0042] In this embodiment, in the flat ground walking scenario, the initial joint torque allocation scheme is first input into the motion simulation layer, a module capable of simulating the robot's motion state. Simultaneously, environmental constraint data for the current scene is loaded. This data includes the ground friction coefficient, ground flatness, and the presence of obstacles. In the flat ground walking scenario, the ground friction coefficient is a crucial environmental constraint, affecting the robot's grip and stability during walking. Based on the initial joint torque allocation scheme and environmental constraint data, the motion simulation layer simulates the motion process of each joint of the robot, generating motion posture feedback signals. These signals contain information about the robot's posture changes during the simulated motion, such as the body's tilt angle and joint position changes. Furthermore, the motion simulation layer analyzes the simulated motion process, identifying deviations in the torque allocation scheme. For example, excessively high or low torque allocation to a joint may lead to robot instability, or uncoordinated torque allocation between different joints may result in uneven motion. Then, these deviations are traced back to determine the cause of the deviation, such as whether the torque value of a certain joint in the initial torque distribution scheme is set unreasonably, or whether the environmental constraint data has an unexpected impact on the motion, thereby generating torque distribution deviation traceability information.

[0043] Step S121: Input the initial joint torque allocation scheme into the motion simulation layer, and simultaneously load the environmental constraint data and robot body dynamic parameters.

[0044] In this embodiment, in the flat ground walking scenario, the initial joint torque allocation scheme is first transmitted to the motion simulation layer. The motion simulation layer receives and parses this scheme to obtain the torque allocation value for each joint. Simultaneously, environmental constraint data for the current scene, such as the friction coefficient data of the ground, is loaded. This data is stored in the system's environmental database in a specific format, and the motion simulation layer reads the corresponding friction coefficient data from this database. In addition, robot body dynamics parameters are also loaded. These parameters include the rotational inertia of each joint, mass distribution, and connection stiffness between joints. These parameters are inherent properties of the robot and affect its motion response and dynamic characteristics. The motion simulation layer then integrates the initial joint torque allocation scheme, environmental constraint data, and robot body dynamics parameters.

[0045] Step S122: Simulate joint motion based on the initial joint torque allocation scheme, environmental constraint data and robot body dynamics parameters, generate basic virtual motion posture data and perform scene adaptability verification.

[0046] In this embodiment, in the flat ground walking scenario, the motion simulation layer simulates the motion process of each joint of the robot based on the initial joint torque allocation scheme, environmental constraint data, and robot body dynamics parameters. First, based on the torque values ​​in the initial joint torque allocation scheme, combined with the rotational inertia and mass distribution in the robot body dynamics parameters, the angular acceleration of each joint is calculated. Then, based on the angular acceleration and the initial motion state, the motion rate and position change of the joint are calculated, thereby simulating the joint's motion trajectory. During the simulation, the influence of environmental constraint data is considered; for example, the ground friction coefficient affects the friction between the robot's feet and the ground, thus affecting the joint's motion torque and trajectory. Through this simulation, basic virtual motion posture data is generated. This data contains the robot's posture information at various moments during the simulated motion, such as body position, joint angles, and body tilt angles. Then, a scenario adaptability check is performed, that is, checking whether the basic virtual motion posture data meets the requirements of the current flat ground walking scenario, such as checking whether the robot's posture is stable, whether it can maintain balance, and whether it can adapt to the ground friction coefficient. If the posture data is found to be inconsistent with the scenario requirements, the corresponding anomalies are marked, and the simulation process is adjusted.

[0047] Step S1221: Associate the initial joint torque distribution scheme with the robot's body dynamics parameters to simulate single joint torque transmission and motion response.

[0048] In this embodiment, in the flat ground walking scenario, the torque value of each joint in the initial joint torque distribution scheme is first correlated with the corresponding joint's moment of inertia, mass distribution, and other parameters in the robot's body dynamics parameters. Then, based on the relationship between torque and moment of inertia, the angular acceleration of each joint is calculated. Angular acceleration reflects the rate of change of the joint's motion state. Next, based on the angular acceleration and the initial motion rate, the change in the joint's motion rate at different times is calculated. Combined with the initial joint position, the change in the joint's position at different times is calculated, thereby simulating the torque transmission and motion response process of a single joint. For example, for the knee joint, based on the knee joint torque value and the knee joint's moment of inertia in the initial torque distribution scheme, the angular acceleration of the knee joint is calculated. Then, based on the angular acceleration, the motion rate of the knee joint at each time moment is calculated, thus obtaining the change in the knee joint's position and simulating the bending and extension movements of the knee joint during walking.

[0049] Step S1222: Predict multi-joint motion interference based on robot limb structure parameters, and adjust and integrate the motion trajectories of each single joint.

[0050] In this embodiment, the robot's limb structural parameters include the distances between joints, the range of motion of the joints, and the length of the limbs. In a flat-ground walking scenario, the motion simulation layer uses these limb structural parameters to predict whether multiple joints will interfere with each other during movement. For example, it might predict whether the movement trajectory of the knee joint will overlap with that of the hip joint, or whether the movement trajectory of the ankle joint will conflict with that of the knee joint. If motion interference is predicted, the motion simulation layer adjusts the movement trajectories of the relevant joints, such as adjusting the joint's movement speed or angle, to avoid interference. Then, the adjusted individual joint movement trajectories are integrated to form the robot's overall motion trajectory, which reflects the robot's overall movement state during walking.

[0051] Step S1223: Generate basic virtual motion posture data containing the combined postures of all joints in the body.

[0052] In this embodiment, the motion simulation layer generates basic virtual motion posture data containing the combined postures of all joints based on the integrated whole-body motion trajectory. This data records the position and angle information of all joints of the robot at different times, such as the angle combination of the hip, knee, and ankle joints during the support phase of walking, and the angle changes of each joint during the swing phase of walking. The basic virtual motion posture data is stored in the form of a time series, with each time point corresponding to a set of posture information of all joints. This information can completely reflect the posture changes of the robot during walking.

[0053] Step S1224: Verify the scene adaptability of the basic virtual motion posture data based on environmental constraint data, and mark adaptation anomalies.

[0054] In this embodiment, in a flat-ground walking scenario, environmental constraints such as the ground friction coefficient affect the robot's posture. The motion simulation layer calculates the frictional force between the robot's feet and the ground during walking based on the ground friction coefficient, and then checks the stability of the robot's posture in the basic virtual motion posture data based on this frictional force. For example, if the ground friction coefficient is low, the robot may slip during walking. The motion simulation layer checks whether there are signs of such slippage in the basic virtual motion posture data; if so, it is marked as an adaptation anomaly. In addition, it also checks whether the robot's posture meets the basic requirements for flat-ground walking, such as whether the body remains upright and whether the stride length is uniform. If these requirements are not met, it is also marked as an adaptation anomaly.

[0055] Step S1225: Output the basic virtual motion posture data and adaptation anomaly marker information that have passed the scene adaptability verification.

[0056] In this embodiment, the motion simulation layer filters out basic virtual motion posture data that passes the scene adaptability check, i.e., posture data that has no adaptation anomalies or whose adaptation anomalies are within the allowable range. At the same time, it outputs adaptation anomaly labeling information, which records in detail the location, type, and degree of the anomalies. For example, it marks an abnormal knee joint angle at a certain moment, or marks a body tilt angle that exceeds the allowable range at a certain moment.

[0057] Step S123: Compare the basic virtual motion posture data that has passed the scene adaptability verification with the preset posture reference information to generate posture deviation information.

[0058] In this embodiment, in the flat ground walking scenario, the preset posture reference information is standard posture information developed based on extensive flat ground walking experiments and theoretical analysis. It includes information such as the robot's standard joint angles, body position, and tilt angle at various moments during the walking process. The motion simulation layer compares the basic virtual motion posture data, which has passed scene adaptability verification, with the preset posture reference information, comparing parameters such as joint angles, body position, and tilt angle at each moment. For example, it compares the hip joint angle at a certain moment in the basic virtual motion posture data with the corresponding hip joint angle in the preset posture reference information, calculating the difference between the two; it also compares the body tilt angle at a certain moment with the corresponding tilt angle in the preset posture reference information, calculating the difference. These differences are summarized and organized to generate posture deviation information, which reflects the degree of difference between the basic virtual motion posture data and the standard posture.

[0059] Step S124: Tracing back to the source of the deviation to locate the core deviation joint and related joint chain, and generating torque distribution deviation tracing information.

[0060] In this embodiment, during a flat-ground walking scenario, the motion simulation layer traces the source of the deviation based on the posture deviation information. First, it analyzes the deviation of each parameter in the posture deviation information, identifying parameters with significant deviations. For example, if the hip joint angle deviation is large, the hip joint is the focus of analysis. Then, it studies the relationship between these parameters with significant deviations and joint torque distribution. For instance, it analyzes whether the hip joint angle deviation is caused by an unreasonable torque distribution in the hip joint, or whether it is caused by an unreasonable torque distribution in joints associated with the hip joint. Through this analysis, the core deviation joint is located—the joint that directly causes the posture deviation—and the joint chains associated with the core deviation joint are identified, such as the joint chains formed by the hip joint, knee joint, and ankle joint. Finally, torque distribution deviation tracing information is generated. This information records the location of the core deviation joint, the type and degree of deviation, and information about the associated joint chains.

[0061] Step S125: Integrate the basic virtual motion posture data, posture deviation information and torque distribution deviation traceability information that have passed the scene adaptability verification to form a motion posture feedback signal.

[0062] In this embodiment, the motion simulation layer integrates the basic virtual motion posture data, posture deviation information, and torque distribution deviation source information that have passed scene adaptability verification to form a motion posture feedback signal. During the integration process, this information is arranged and combined according to a preset format; for example, the basic virtual motion posture data is placed at the beginning of the feedback signal, the posture deviation information in the middle, and the torque distribution deviation source information at the end. Each information component is also labeled to distinguish different types of information. This motion posture feedback signal comprehensively reflects the robot's posture, deviations, and the causes of these deviations during simulated motion.

[0063] Step S130: Combine the torque distribution deviation source information to locate the core deviation joint and related joint chain, and adjust the weights of the hierarchical dynamic optimization model by combining the motion posture feedback signal and the deviation positioning result to obtain the adjusted hierarchical dynamic optimization model.

[0064] In this embodiment, during a flat-ground walking scenario, the core deviation joint and its associated joint chain are first located based on the torque distribution deviation tracing information. For example, the hip joint is identified as the core deviation joint, and the associated joint chain is hip-knee-ankle. Then, the motion posture feedback signal and the deviation location results are jointly analyzed. The posture deviation information in the motion posture feedback signal reflects the specific degree of deviation, while the deviation location results reflect the location and correlation of the deviation. Based on these analysis results, the weights of the hierarchical dynamics optimization model are adjusted. Adjusting the model weights affects the model's emphasis on different joint features. For example, if the core deviation joint is the hip joint, the weights of features related to the hip joint are increased, while the weights of features related to other non-core deviation joints are decreased to improve the model's ability to correct hip joint deviations. Through this weight adjustment, an adjusted hierarchical dynamics optimization model is obtained. This model can better adapt to the current motion scenario and improve the accuracy of torque distribution.

[0065] Step S131: Process the torque distribution deviation source information, locate the core deviation joint, deviation type and related joint chain, and form the deviation location result.

[0066] In this embodiment, during the flat ground walking scenario, the received torque distribution deviation tracing information is first parsed to extract information about the core deviation joint, deviation type, and associated joint chain. For example, the core deviation joint is extracted as the hip joint, the deviation type as insufficient torque distribution, and the associated joint chain as hip-knee-ankle. Then, this information is organized and classified to form the deviation localization result. The deviation localization result is stored in a structured data format, including the identifier of the core deviation joint, a description of the deviation type, and the composition of the associated joint chain.

[0067] Step S132: Extract the attitude deviation information from the motion attitude feedback signal and input it together with the deviation positioning result into the model weight adjustment module.

[0068] In this embodiment, posture deviation information is first extracted from the motion posture feedback signal. This information includes the differences between the postures of each joint and the standard posture, such as hip joint angle deviations and knee joint angle deviations. Then, the extracted posture deviation information is combined with the deviation localization results; for example, the hip joint angle deviation value is correlated with the deviation localization result where the core deviation joint is the hip joint. The combined information is then input into the model weight adjustment module, which analyzes the received information to determine the direction and magnitude of the model weight adjustment.

[0069] Step S133: Correct the feature encoding weights of the corresponding joint-level signal parsing unit through the joint-level adjustment submodule.

[0070] In this embodiment, during the flat ground walking scenario, the joint-level adjustment submodule corrects the feature encoding weights of the corresponding joint-level signal analysis units based on the analysis results of the model weight adjustment module. First, it identifies the joint-level signal analysis units that need adjustment; for example, if the core deviation joint is the hip joint, it determines the joint-level signal analysis unit corresponding to the hip joint to be adjusted. Then, based on the posture deviation information and deviation localization results, it determines the adjustment direction of the feature encoding weights. For example, if the hip joint's angle deviation is due to an excessively low encoding weight for angle features, the encoding weight for angle features will be increased; if it is due to an excessively high encoding weight for force features, the encoding weight for force features will be decreased. Next, it determines the magnitude of the weight adjustment based on the degree of deviation; the greater the deviation, the greater the potential magnitude of the weight adjustment. Through this correction, the joint-level signal analysis unit's ability to extract and encode features related to the core deviation joint is improved.

[0071] Step S1331: Extract the core deviation joint location and deviation type information from the deviation positioning results.

[0072] In this embodiment, during the flat ground walking scenario, the deviation localization results are first analyzed to extract information about the core deviation joint location and deviation type. For example, the core deviation joint location is extracted as the hip joint, and the deviation type is insufficient torque distribution. This information is crucial for determining the direction and magnitude of feature encoding weight adjustment. The core deviation joint location indicates the joint-level signal parsing unit that needs adjustment, and the deviation type indicates the feature type that needs adjustment.

[0073] Step S1332: Match the deviation type and the feature encoding weight adjustment direction to generate the feature encoding weight adjustment strategy for the core deviation joint and related joints.

[0074] In this embodiment, in the flat ground walking scenario, the first step is to match the corresponding feature encoding weight adjustment direction based on the extracted deviation type. For example, if the deviation type is insufficient torque distribution, it may be due to the low encoding weight of joint angle features, resulting in insufficient sensitivity of the model to changes in joint angles, thus leading to insufficient torque distribution. In this case, the adjustment direction is to increase the encoding weight of angle features. If the deviation type is excessive torque distribution, it may be due to the high encoding weight of force features, causing the model to overemphasize force feedback, thus leading to excessive torque distribution. In this case, the adjustment direction is to decrease the encoding weight of force features. Then, based on the core deviation joint and the associated joint chain, the joint range of feature encoding weights that need to be adjusted is determined. For example, if the core deviation joint is the hip joint and the associated joint chain is hip-knee-ankle, then the feature encoding weights of the joint-level signal parsing units corresponding to the hip, knee, and ankle joints need to be adjusted. Combining the adjustment direction and adjustment range generates a feature encoding weight adjustment strategy.

[0075] Step S1333: Combine the deviation values ​​in the attitude deviation information to quantify the adjustment range of the feature encoding weights.

[0076] In this embodiment, the deviation values ​​in the posture deviation information reflect the degree of difference between the posture and the standard posture. The larger the deviation value, the more unreasonable the feature encoding weights of that joint are in the model, and the greater the adjustment required. In the flat ground walking scenario, the deviation values ​​in the posture deviation information are first obtained, for example, the hip joint angle deviation value is within a certain range. Then, according to a preset quantization rule, the deviation values ​​are mapped to the feature encoding weight adjustment range. For example, when the deviation value is within a small range, the adjustment range is small; when the deviation value is within a large range, the adjustment range is large. Through this quantization, the specific adjustment range for each feature encoding weight that needs to be adjusted is determined.

[0077] Step S1334: Convert the feature encoding weight adjustment strategy and feature encoding weight adjustment magnitude into parameter instructions to correct the feature encoding weights of the joint-level signal analysis units corresponding to the core deviation joint and associated joints.

[0078] In this embodiment, during the flat ground walking scenario, the feature encoding weight adjustment strategy and adjustment range are first converted into parameter instructions that the machine can recognize and execute. These parameter instructions include information such as the identifier of the joint-level signal analysis unit to be adjusted, the type of feature to be adjusted, the adjustment direction, and the adjustment range. Then, these parameter instructions are sent to the corresponding joint-level signal analysis unit, which will correct its own feature encoding weights according to the parameter instructions. For example, for the joint-level signal analysis unit corresponding to the hip joint, the encoding weight of angle features is increased and the encoding weight of force features is decreased according to the parameter instructions. The adjusted weights can more accurately reflect the motion characteristics of the hip joint and improve the model's ability to correct hip joint deviations.

[0079] Step S1335: Input historical similarity deviation data to verify the effect of feature encoding weight adjustment and lock the final feature encoding weight parameters of the joint-level signal parsing unit.

[0080] In this embodiment, historical deviation data similar to the current deviation are first retrieved from the historical optimization database. This data includes previously occurring similar deviations, corresponding feature encoding weight adjustment strategies, and the effects of the adjustments. Then, the adjusted feature encoding weights of the joint-level signal parsing unit are applied to these historical similar deviation data. The model's processing of this data is simulated, and the improvement in processing results is observed. If the processing results show significant improvement, the feature encoding weight adjustment is effective, and the current feature encoding weight parameters are locked. If the processing results do not improve, or the improvement is not significant, the feature encoding weights need to be readjusted, and the above steps are repeated until the adjustment effect reaches the expected level.

[0081] Step S134: Correct the association weights of the system-level association mapping units and update the global joint torque association coefficient generation logic through the system-level adjustment submodule.

[0082] In this embodiment, during the flat ground walking scenario, the system-level adjustment submodule corrects the association weights of the system-level association mapping unit based on the analysis results of the model weight adjustment module. First, it determines the association weights that need adjustment. For example, if the core deviation joint is the hip joint and its association weight with the knee joint is unreasonable, it determines to adjust the association weight between the hip and knee joints. Then, based on the posture deviation information and deviation positioning results, it determines the adjustment direction and magnitude of the association weights. For example, if the association weight between the hip and knee joints is too low, resulting in insufficient motion coordination, this association weight will be increased; if the association weight is too high, leading to motion interference, this association weight will be decreased. Simultaneously, the global joint moment association coefficient generation logic is updated, enabling the generated global joint moment association coefficients to more accurately reflect the association relationships between joints and improve the model's ability to handle coordinated movements of all joints.

[0083] Step S135: Store the adjusted parameters of the hierarchical dynamics optimization model to form the adjusted hierarchical dynamics optimization model.

[0084] In this embodiment, during the flat ground walking scenario, the system stores the adjusted parameters of the hierarchical dynamics optimization model. These parameters include the feature encoding weights of the joint-level signal parsing unit, the association weights of the system-level association mapping unit, and the global joint torque association coefficient generation logic. The stored parameters are marked as valid parameters for the current scenario. When the robot performs a similar flat ground walking action again, these parameters can be directly invoked, improving the model's processing efficiency and accuracy. By storing these adjusted parameters, an adjusted hierarchical dynamics optimization model is formed. This model can better adapt to the current motion scenario and improve the accuracy of torque distribution.

[0085] Step S140: Input the fused motion signal into the adjusted hierarchical dynamic optimization model, and output the optimized joint torque distribution scheme through the built-in cross-joint torque coupling optimization unit.

[0086] In this embodiment, during the flat ground walking scenario, the fused motion signal is first input into an adjusted hierarchical dynamics optimization model. The adjusted model, after weight correction, can more accurately process the current motion signal. The model analyzes and processes the fused motion signal, extracting feature information and generating global joint torque correlation coefficients. Then, the built-in cross-joint torque coupling optimization unit optimizes the global joint torque correlation coefficients, considering the torque coupling relationships between different joints. For example, during walking, changes in the hip joint torque affect changes in the knee joint torque, which in turn affect changes in the ankle joint torque. Based on these coupling relationships, the cross-joint torque coupling optimization unit adjusts the joint torque distribution, eliminating conflicts and unreasonable aspects in torque distribution, and generating an optimized joint torque distribution scheme. This scheme can better coordinate the movement of all joints, improving the stability and efficiency of robot walking.

[0087] Step S141: Input the fused motion signal into the adjusted hierarchical dynamic optimization model, and generate the optimized global joint torque correlation coefficient through joint-level analysis and system-level correlation processing.

[0088] In this embodiment, during a flat-ground walking scenario, the fused motion signal is first input into the adjusted hierarchical dynamic optimization model. The model's joint-level signal parsing unit parses the fused motion signal, extracts its feature information, and encodes it. Then, the system-level correlation mapping unit processes the encoded feature information, establishes the correlation between different joints, and generates an optimized global joint torque correlation coefficient. This correlation coefficient takes into account the adjusted model weights and can more accurately reflect the degree of torque correlation between joints.

[0089] Step S142: Input the optimized global joint moment correlation coefficient into the cross-joint moment coupling optimization unit and load the joint coupling constraint rules.

[0090] In this embodiment, during a flat-ground walking scenario, the cross-joint torque coupling optimization unit receives the optimized global joint torque correlation coefficients and simultaneously loads joint coupling constraint rules. These joint coupling constraint rules are formulated based on the robot's motion characteristics and joint structure, specifying the limitations on torque distribution between different joints. For example, during walking, the sum of the torques of the hip and knee joints cannot exceed a certain maximum value, and the torque of the ankle joint cannot fall below a certain minimum value. After loading these constraint rules, the cross-joint torque coupling optimization unit optimizes the global joint torque correlation coefficients according to these rules.

[0091] Step S143: Verify and correct the torque distribution coefficient in the optimized global joint torque correlation coefficient that has coupling conflicts based on the joint coupling constraint rules.

[0092] In this embodiment, during a flat-ground walking scenario, the cross-joint torque coupling optimization unit verifies the torque distribution coefficients in the optimized global joint torque correlation coefficients according to joint coupling constraint rules. First, it checks whether each torque distribution coefficient meets the requirements of the constraint rules. For example, it checks whether the sum of the torques of the hip and knee joints exceeds the maximum value, and whether the torque of the ankle joint is below the minimum value. If a coupling conflict is found, i.e., the torque distribution coefficients do not meet the constraint rules, these coefficients are corrected. For example, if the sum of the torques of the hip and knee joints exceeds the maximum value, the torque distribution coefficient of the hip or knee joint is appropriately reduced to make the sum meet the constraint rules; if the torque of the ankle joint is below the minimum value, the torque distribution coefficient of the ankle joint is appropriately increased to make it meet the constraint rules.

[0093] Step S1431: Group the global joint moment correlation coefficients according to the joint coupling relationship, and match the joint coupling constraint rules corresponding to each group of joints.

[0094] In this embodiment, during a flat-ground walking scenario, the global joint torque correlation coefficients are first grouped according to the coupling relationships between joints. For example, the hip and knee joints are grouped together, the knee and ankle joints are grouped together, and the upper limb and trunk joints are grouped together, etc. Then, corresponding joint coupling constraint rules are matched for each group. For example, the hip and knee joint group is matched with a torque sum constraint rule, and the knee and ankle joint group is matched with a torque transmission constraint rule. Through such grouping and matching, the torque distribution coefficients can be more effectively verified and corrected.

[0095] Step S1432: Verify whether the global joint moment correlation coefficients of each group meet the corresponding joint coupling constraint rules, and locate the coupling conflict parts and conflict types.

[0096] In this embodiment, during the flat ground walking scenario, the cross-joint torque coupling optimization unit verifies the global joint torque correlation coefficient of each group to check whether it meets the corresponding joint coupling constraint rules. For example, it checks whether the sum of the torques of the hip and knee joint groups exceeds the maximum value, and whether the torque transmission of the knee and ankle joint groups conforms to the constraint rules. If it finds that the torque correlation coefficient of a certain group does not meet the constraint rules, it will locate the coupling conflict location, such as locating a conflict between the hip and knee joint groups; at the same time, it will determine the conflict type, such as whether the conflict type is that the sum of the torques exceeds the maximum value, or whether the torque transmission does not conform to the rules.

[0097] Step S1433: Develop a correction strategy for different coupling conflict types and adjust the torque distribution coefficient in the global joint torque correlation coefficient corresponding to the joint with coupling conflict.

[0098] In this embodiment, during the flat ground walking scenario, a corresponding correction strategy is formulated based on the identified coupling conflict type. For example, if the conflict type is that the sum of torques exceeds the maximum value, the correction strategy could be to reduce the torque distribution coefficient of one or more joints; if the conflict type is that torque transmission does not conform to the rules, the correction strategy could be to adjust the torque distribution ratio between joints. Then, according to the correction strategy, the torque distribution coefficient in the global joint torque correlation coefficient corresponding to the joints with coupling conflicts is adjusted. For example, for the case where the sum of torques of the hip and knee joints exceeds the maximum value, the torque distribution coefficient of the hip joint is reduced so that the sum of the two conforms to the constraint rules.

[0099] Step S1434: Perform a second check on the global joint torque correlation coefficient after adjustment, and investigate residual coupling conflicts and newly caused torque distribution deviations.

[0100] In this embodiment, during the flat ground walking scenario, the cross-joint torque coupling optimization unit performs a secondary verification of the adjusted global joint torque correlation coefficients to check for any remaining coupling conflicts and whether any new torque distribution deviations were introduced during the adjustment process. For example, it checks whether the sum of the torques of the adjusted hip and knee joint groups conforms to the constraint rules, and also checks whether the adjusted torque distribution coefficients cause deviations in the torque distribution of other joints. If any residual coupling conflicts or newly introduced deviations are found, the above correction steps are repeated until all torque distribution coefficients conform to the constraint rules and no new deviations are generated.

[0101] Step S1435: Repeatedly correct the global joint moment correlation coefficients until all coefficients satisfy the joint coupling constraint rules, and output the corrected global joint moment correlation coefficients.

[0102] In this embodiment, during the flat ground walking scenario, the cross-joint moment coupling optimization unit continuously corrects and verifies the global joint moment correlation coefficients until all coefficients satisfy the joint coupling constraint rules. During the correction process, the correction strategy and correction magnitude are adjusted based on the results of each verification to ensure that the corrected coefficients conform to the constraint rules without introducing new deviations. Once all coefficients satisfy the constraint rules, the corrected global joint moment correlation coefficients are output, which accurately reflect the moment correlation between joints.

[0103] Step S144: Repeatedly verify the torque distribution coefficient correction results with coupling conflicts until the conflicts are eliminated, and generate a preliminary optimized torque distribution scheme.

[0104] In this embodiment, during a flat-ground walking scenario, the cross-joint torque coupling optimization unit generates a preliminary optimized torque allocation scheme based on the corrected global joint torque correlation coefficients. This preliminary scheme is then repeatedly validated to check for any remaining coupling conflicts. If conflicts exist, the torque allocation coefficients are corrected again, and the above steps are repeated until all conflicts are eliminated. The generated preliminary optimized torque allocation scheme satisfies the joint coupling constraint rules, improving the kinematic coordination between joints.

[0105] Step S145: Extract the matching features between the preliminary optimized torque distribution scheme and the fused motion signal, and output the final optimized joint torque distribution scheme.

[0106] In this embodiment, in a flat-ground walking scenario, the matching features between the preliminary optimized torque allocation scheme and the fused motion signal are first extracted. These features include the degree of matching between the torque allocation scheme and the joint motion features in the fused motion signal. For example, whether the torque change of the hip joint in the torque allocation scheme matches the motion rate change of the hip joint in the fused motion signal, and whether the torque allocation of the knee joint matches the force feedback of the knee joint in the fused motion signal. Then, based on these matching features, the preliminary optimized torque allocation scheme is fine-tuned to improve the matching degree between the scheme and the fused motion signal. Finally, the final optimized joint torque allocation scheme is output. This scheme can accurately reflect the robot's current motion state and environmental constraints, ensuring that the robot can stably and efficiently complete the flat-ground walking action.

[0107] Step S150: Transmit the optimized joint torque distribution scheme to the joint drive system, and at the same time store the matching data of the optimized joint torque distribution scheme and the real-time joint signal to the historical optimization database to complete the online dynamics optimization.

[0108] In this embodiment, in a flat-ground walking scenario, the optimized joint torque distribution scheme is first transmitted to the robot's joint drive system. The joint drive system then controls the motors of each joint to output corresponding torques according to this scheme, enabling the robot to walk according to the optimized posture and trajectory. Simultaneously, the system stores the matching data between the optimized joint torque distribution scheme and real-time joint signals. This data includes the optimized torque distribution values, joint angles, movement speeds, force feedback, and matching degree evaluation data from the real-time joint signals. This data is stored in a historical optimization database to continuously improve the robot's motion performance. By executing these steps, the real-time online optimization process of the humanoid robot's whole-body dynamics in a flat-ground walking scenario is completed.

[0109] Figure 2 Schematic diagrams are shown of exemplary hardware and software components of a real-time online optimization system for the whole-body dynamics of a humanoid robot, which can implement the ideas of this application, according to some embodiments of this application. For example, a processor 120 can be used in the real-time online optimization system for the whole-body dynamics of a humanoid robot 100 and to perform the functions in this application.

[0110] The real-time online optimization system 100 for the whole-body dynamics of a humanoid robot can be a general-purpose server or a special-purpose server; both can be used to implement the real-time online optimization method for the whole-body dynamics of a humanoid robot as described in this application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the load.

[0111] For example, a real-time online optimization system 100 for the whole-body dynamics of a humanoid robot may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and various forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the real-time online optimization system 100 for the whole-body dynamics of a humanoid robot may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The methods of this application can be implemented according to these program instructions. The real-time online optimization system 100 for the whole-body dynamics of a humanoid robot also includes an I / O interface 150 between the computer and other input / output devices.

[0112] For ease of explanation, only one processor is described in the real-time online optimization system for whole-body dynamics of a humanoid robot 100. However, it should be noted that the real-time online optimization system for whole-body dynamics of a humanoid robot 100 of this application may also include multiple processors, and therefore the steps performed by one processor described in this application may also be performed jointly or individually by multiple processors. For example, if the processor of the real-time online optimization system for whole-body dynamics of a humanoid robot 100 performs steps A and B, it should be understood that steps A and B may also be performed jointly by two different processors or individually by one processor. For example, the first processor performs step A, the second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0113] Furthermore, this embodiment of the invention also provides a readable storage medium, wherein computer-executable instructions are preset in the readable storage medium, and when the processor executes the computer-executable instructions, the above-mentioned method for real-time online optimization of the whole-body dynamics of a humanoid robot is implemented.

[0114] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for real-time online optimization of the whole-body dynamics of a humanoid robot, characterized in that, The method includes: The system receives real-time joint signals from all joints of the humanoid robot, integrates historical optimization data of similar movements of the humanoid robot to generate a fused motion signal, and inputs the fused motion signal into a pre-trained hierarchical dynamics optimization model to generate an initial joint torque allocation scheme. The initial joint torque allocation scheme is input into the motion simulation layer and the current scene environment constraint data is loaded. The motion simulation layer outputs motion posture feedback signals and torque allocation deviation traceability information. By combining the torque distribution deviation source information to locate the core deviation joint and related joint chain, and by combining the motion posture feedback signal and the deviation location result to adjust the weight of the hierarchical dynamic optimization model, the adjusted hierarchical dynamic optimization model is obtained. After adjusting the fused motion signal input, the hierarchical dynamic optimization model outputs the optimized joint torque distribution scheme through the built-in cross-joint torque coupling optimization unit. The optimized joint torque distribution scheme is transmitted to the joint drive system, and the matching data of the optimized joint torque distribution scheme and real-time joint signals are stored in the historical optimization database to complete the online dynamics optimization.

2. The method for real-time online optimization of the whole-body dynamics of a humanoid robot according to claim 1, characterized in that, The process involves receiving real-time joint signals from all joints of the humanoid robot, fusing historical optimization data of similar movements of the humanoid robot to generate a fused motion signal, and inputting the fused motion signal into a pre-trained hierarchical dynamics optimization model to generate an initial joint torque allocation scheme, including: It receives real-time joint signals including joint rotation angle, movement rate and force feedback, and matches them with historical optimization data of the same type of movement; Real-time joint signals and historical optimization data are fused according to scene-adaptive weight rules to generate fused motion signals; The fused motion signals are input into the joint-level signal parsing unit of the hierarchical dynamic optimization model according to limb partitions to obtain the partitioned joint feature encoding results. The system-level association mapping unit processes the partitioned joint feature encoding results to generate global joint moment association coefficients. Based on the joint hardware constraint parameters, the initial joint torque distribution scheme is output through the torque distribution generation unit.

3. The method for real-time online optimization of the whole-body dynamics of a humanoid robot according to claim 1, characterized in that, The process of inputting the initial joint torque allocation scheme into the motion simulation layer and loading the current scene environment constraint data, followed by the motion simulation layer outputting motion posture feedback signals and torque allocation deviation tracing information, includes: The initial joint torque allocation scheme is input into the motion simulation layer, and environmental constraint data and robot body dynamic parameters are loaded simultaneously. Based on the initial joint torque allocation scheme, environmental constraint data and robot body dynamics parameters, the joint motion is simulated to generate basic virtual motion posture data and perform scene adaptability verification. By comparing the basic virtual motion posture data that has passed the scene adaptability verification with the preset posture reference information, posture deviation information is generated. Tracing the source of the deviation in reverse, the core deviation joint and related joint chain are located, and torque distribution deviation source information is generated. Integrate basic virtual motion posture data that has passed scene adaptability verification, posture deviation information, and torque distribution deviation traceability information to form a motion posture feedback signal.

4. The method for real-time online optimization of the whole-body dynamics of a humanoid robot according to claim 1, characterized in that, The method involves locating the core deviation joint and related joint chain by combining the torque distribution deviation source information, and adjusting the weights of the hierarchical dynamic optimization model based on the motion posture feedback signal and deviation location results to obtain the adjusted hierarchical dynamic optimization model, including: Process torque distribution deviation source information, locate the core deviation joint, deviation type and related joint chain, and generate deviation location results; Extract the attitude deviation information from the motion attitude feedback signal and input it together with the deviation localization result into the model weight adjustment module; The feature encoding weights of the corresponding joint-level signal parsing unit are corrected by adjusting the joint-level adjustment submodule. The system-level adjustment submodule corrects the association weights of the system-level association mapping unit and updates the global joint torque association coefficient generation logic. The parameters of the storage hierarchy dynamic optimization model are adjusted to form the adjusted hierarchy dynamic optimization model.

5. The method for real-time online optimization of the whole-body dynamics of a humanoid robot according to claim 1, characterized in that, The hierarchical dynamic optimization model, which integrates and adjusts the fused motion signal input, outputs an optimized joint torque distribution scheme through a built-in cross-joint torque coupling optimization unit, including: The hierarchical dynamic optimization model after fusion motion signal input adjustment is used to generate optimized global joint torque correlation coefficients through joint-level analysis and system-level correlation processing. The optimized global joint moment correlation coefficients are input into the cross-joint moment coupling optimization unit, and joint coupling constraint rules are loaded. The torque distribution coefficient in the optimized global joint torque correlation coefficient with coupling conflicts is verified and corrected based on the joint coupling constraint rules. Repeatedly verify the torque distribution coefficient correction results with coupling conflicts until the conflicts are eliminated, and generate a preliminary optimized torque distribution scheme; Extract the matching features between the preliminary optimized torque distribution scheme and the fused motion signal, and output the final optimized joint torque distribution scheme.

6. The method for real-time online optimization of the whole-body dynamics of a humanoid robot according to claim 2, characterized in that, The process of inputting the fused motion signals into the joint-level signal parsing unit of the hierarchical dynamic optimization model according to limb partitions to obtain the partitioned joint feature encoding results includes: The joint-level signal analysis unit, configured according to limb function zones, is activated to receive the fused motion signals of the corresponding zones; The basic motion features of the fused motion signal are extracted through the feature extraction sublayer of the joint-level signal parsing unit. Key dynamic information in the basic motion features of motion signals is transformed and fused through feature processing sub-layers; The feature coding sublayer transforms the basic motion features after converting key dynamic information into standardized coding results; The standardized coding results of each limb region are summarized to form a set of region joint feature coding results.

7. The method for real-time online optimization of the whole-body dynamics of a humanoid robot according to claim 3, characterized in that, The process of simulating joint motion based on the initial joint torque allocation scheme, environmental constraint data, and robot body dynamics parameters to generate basic virtual motion posture data and perform scene adaptability verification includes: By associating the initial joint torque distribution scheme with the robot's body dynamics parameters, the torque transmission and motion response of a single joint are simulated. By combining the robot's limb structural parameters, multi-joint motion interference is predicted, and the motion trajectories of each single joint are adjusted and integrated. Generate basic virtual motion posture data containing the combined postures of all joints in the body; Verify the scene adaptability of the basic virtual motion posture data based on environmental constraint data, and mark adaptation anomalies. Output the basic virtual motion posture data that passes the scene adaptability check, as well as the information on the labels of adaptation anomalies.

8. The method for real-time online optimization of the whole-body dynamics of a humanoid robot according to claim 4, characterized in that, The step of correcting the feature encoding weights of the corresponding joint-level signal parsing unit through the joint-level adjustment submodule includes: Extract the core deviation joint location and deviation type information from the deviation localization results; Match the deviation type and the feature encoding weight adjustment direction to generate the feature encoding weight adjustment strategy for the core deviation joint and related joints; By combining the deviation values ​​in the attitude deviation information, the adjustment range of the feature encoding weights is quantified. The feature encoding weight adjustment strategy and the feature encoding weight adjustment magnitude are converted into parameter instructions to correct the feature encoding weights of the joint-level signal parsing units corresponding to the core deviation joints and related joints. Input historical similarity deviation data to verify the effect of feature encoding weight adjustment, and lock the final feature encoding weight parameters of the joint-level signal analysis unit.

9. The method for real-time online optimization of the whole-body dynamics of a humanoid robot according to claim 5, characterized in that, The torque distribution coefficient in the optimized global joint torque correlation coefficient, which is verified and corrected based on joint coupling constraint rules to address coupling conflicts, includes: Group the global joint moment correlation coefficients according to the joint coupling relationship, and match the joint coupling constraint rules corresponding to each group of joints. Verify whether the global joint moment correlation coefficients of each group meet the corresponding joint coupling constraint rules, and locate the coupling conflict locations and conflict types; Develop correction strategies for different types of coupling conflicts, and adjust the torque distribution coefficient in the global joint torque correlation coefficient corresponding to the joint with coupling conflict; After secondary verification and adjustment, the global joint torque correlation coefficient is checked to identify residual coupling conflicts and newly caused torque distribution deviations. Repeatedly adjust the global joint moment correlation coefficients until all coefficients satisfy the joint coupling constraint rules, and output the adjusted global joint moment correlation coefficients.

10. A real-time online optimization system for the whole-body dynamics of a humanoid robot, characterized in that, The real-time online optimization system for whole-body dynamics of a humanoid robot includes a processor and a memory, the memory and the processor being connected. The memory is used to store programs, instructions or code, and the processor is used to execute the programs, instructions or code in the memory to implement the real-time online optimization method for whole-body dynamics of a humanoid robot as described in any one of claims 1-9.

Images