Torque compensation method for joint actuators and related products

Abstract

This application relates to the field of robotics and provides a torque compensation method and related products for joint actuators. The method includes: constructing a saturated degradation physical model; based on the acquired temporal ontology perception signals of N joint actuators, constructing observation tensors corresponding to each of the N joint actuators within a set time window; inputting the N observation tensors into a coupled temporal degradation identification network; extracting the local temporal features of each joint actuator through causal temporal convolution; performing cross-actuator multi-head self-attention operations at the actuator dimension to model the torque redistribution coupling relationship among the N joint actuators; outputting preliminary estimates of the degradation parameters corresponding to each of the N joint actuators; based on the preliminary estimates of the degradation parameters corresponding to each of the N joint actuators, performing mathematical inverse solving on the saturated degradation physical model to obtain a torque compensation calculation formula; and converting the expected torque output by the motion strategy into a compensated command torque according to the torque compensation calculation formula.

Description

Technical Field

[0001] This application relates to the field of robotics, and in particular to a torque compensation method for a joint actuator and related products. Background Technology

[0002] With the continuous development of robotics technology, various robots are gradually being applied in practical scenarios such as industrial inspection, service interaction, and disaster relief. Robots rely on the torque output of their joint actuators to complete various motion tasks such as standing, walking, turning, and jumping. During prolonged continuous operation, joint actuators can degrade due to factors such as temperature rise, mechanical friction, component wear, and motor winding aging. This alters the actuator's torque output capability and command torque response characteristics. If the desired torque output from the motion strategy is directly applied to the degraded actuator after it degrades, the robot's actual output torque will deviate from the expected value, easily leading to decreased motion control accuracy and, in severe cases, robot instability or even tipping over. Therefore, identifying and compensating for the degradation of robot joint actuators is a key technology for ensuring the long-term reliable operation of robots.

[0003] Currently, the industry has developed several technical means to address the gap between robot simulation and reality, as well as dynamic changes in hardware. These mainly include domain randomization, pre-deployment system identification, policy-level adaptive adjustment, and actuator-level fault compensation. Among them, domain randomization randomly perturbs physical parameters during the simulation training phase to improve the robustness of the trained motion strategy to dynamic changes; system identification completes the calibration of physical parameters before robot deployment, reducing the static difference between simulation and real hardware; policy-level adaptive methods adapt to dynamic changes brought about by hardware by adjusting the output behavior of the motion controller online; and actuator-level compensation schemes correct the actuator output residuals by establishing models to achieve control output correction under fault conditions.

[0004] However, the aforementioned existing technical solutions still have many limitations, including: the strategy tends to be conservative under healthy hardware conditions; the control strategy will fail when the actual degree of actuator degradation exceeds the training randomization range; it cannot adapt to the dynamic actuator degradation that gradually evolves during robot deployment and operation; it cannot provide online compensation capabilities during operation; most actuator-level compensation schemes perform independent observation and modeling of a single actuator, making it difficult to fully utilize the correlation information between multiple joints to identify the degradation state; and it cannot be decoupled from existing motion control strategies, etc. Summary of the Invention

[0005] Based on this, it is necessary to address the problems of existing technologies being unable to utilize multi-joint coupling information and being unable to adapt to dynamically evolving actuator degradation online without modifying strategies. Therefore, a torque compensation method and related products for joint actuators are proposed.

[0006] In a first aspect, a torque compensation method for a joint actuator is provided, the method comprising: Construct a saturated degradation physical model to characterize the parametric relationship between the joint actuator in the nominal state and the degradation state; Based on the collected temporal ontology perception signals of N joint actuators, observation tensors corresponding to the N joint actuators are constructed under a set time window. The N observation tensors are input into a coupled temporal degradation identification network. After extracting the local temporal features of each joint actuator through causal temporal convolution, cross-actuator multi-head self-attention operation is performed in the actuator dimension to model the torque redistribution coupling relationship between the N joint actuators. The preliminary estimates of the degradation parameters corresponding to the N joint actuators are output, where N is a positive integer. The coupled temporal degradation identification network is adapted to the saturated degradation physical model. Based on the preliminary estimates of the degradation parameters corresponding to the N joint actuators, the saturated degradation physical model is mathematically inversely solved to obtain the torque compensation calculation formula. According to the torque compensation calculation formula, the expected torque output by the motion strategy is converted into the compensated command torque.

[0007] Secondly, a torque compensation device for a joint actuator is provided, the device comprising: The building module is used to construct a saturated degradation physical model that characterizes the parametric relationship between the joint actuator in the nominal state and the degradation state; An estimation module is used to construct observation tensors corresponding to the N joint actuators under a set time window based on the collected temporal ontology perception signals of the N joint actuators. The N observation tensors are input into a coupled temporal degradation identification network. After extracting the local temporal features of each joint actuator through causal temporal convolution, cross-actuator multi-head self-attention operation is performed in the actuator dimension to model the torque redistribution coupling relationship between the N joint actuators. The module outputs preliminary estimates of the degradation parameters corresponding to the N joint actuators, where N is a positive integer. The coupled temporal degradation identification network is adapted to the saturated degradation physical model. The conversion module is used to perform mathematical inverse solution on the saturated degradation physical model based on the preliminary estimated values ​​of the degradation parameters corresponding to the N joint actuators to obtain the torque compensation calculation formula, and convert the expected torque output by the motion strategy into the compensated command torque according to the torque compensation calculation formula.

[0008] Thirdly, a humanoid robot is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the torque compensation method of the aforementioned joint actuator.

[0009] Fourthly, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, implements the torque compensation method of the aforementioned joint actuator.

[0010] As can be seen from the technical solutions provided in this application: On the one hand, by constructing a saturated degradation physical model to characterize the parameter relationship between joint actuators in the nominal and degraded states, a clear physical modeling object and an analytical mathematical carrier are provided for degradation state identification and torque compensation; on the other hand, by performing cross-actuator multi-head self-attention operations in the actuator dimension, the torque redistribution coupling relationship between N joint actuators is explicitly modeled and the preliminary estimate of the degradation parameters of each joint actuator is output. Instead of performing independent signal analysis on each actuator, degradation identification is carried out using the objectively existing torque redistribution coupling relationship between robot joints, and the preliminary estimate of the degradation parameters of each joint actuator can be obtained based on multi-joint body perception information; thirdly, based on the preliminary estimate of the degradation parameters, the two-parameter saturated degradation physical model is further refined. The inverse kinematics method is used to obtain the torque compensation calculation formula. Based on the torque compensation calculation formula, the expected torque output by the motion strategy is converted into the compensated command torque. This processing method is based on the inverse kinematics of the physical model to obtain the compensation command, which is different from the black-box compensation mapping. At the same time, a safety boundary processing is set to avoid the control risks caused by the abnormal value of the inverse function. The expected torque output by the motion strategy is converted into the command torque adapted to the degenerate actuator. As an intermediate processing flow between the motion strategy and the actuator, the degeneracy identification and torque compensation can be continuously iterated during the robot's task operation without modifying the motion strategy itself. This makes the actual torque output of the degenerate joint actuator close to the expected torque of the motion strategy, improves the robot's motion control effect under actuator degradation conditions, and avoids the problems of decreased motion accuracy and motion instability caused by actuator degradation. In summary, the technical solution of this application uses multi-joint coupling for identification, does not require modification of the motion strategy, compensates for actuator degradation online, and improves the stability of robot motion control. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart of a torque compensation method for a joint actuator in one embodiment; Figure 2 This is a structural block diagram of the torque compensation device for a joint actuator in one embodiment; Figure 3 This is a structural block diagram of a humanoid robot in one embodiment. Detailed Implementation

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] Currently, the industry has developed several technical means to address the gap between robot simulation and reality, as well as dynamic changes in hardware. These mainly include domain randomization, pre-deployment system identification, policy-level adaptive adjustment, and actuator-level fault compensation. Among them, domain randomization randomly perturbs physical parameters during the simulation training phase to improve the robustness of the trained motion strategy to dynamic changes; system identification completes the calibration of physical parameters before robot deployment, reducing the static difference between simulation and real hardware; policy-level adaptive methods adapt to dynamic changes brought about by hardware by adjusting the output behavior of the motion controller online; and actuator-level compensation schemes correct the actuator output residuals by establishing models to achieve control output correction under fault conditions. However, the aforementioned existing technical solutions still have many limitations, including: the strategy tends to be conservative under healthy hardware conditions; the control strategy will fail when the actual degree of actuator degradation exceeds the training randomization range; it cannot adapt to the dynamic actuator degradation that gradually evolves during robot deployment and operation; it cannot provide online compensation capabilities during operation; most actuator-level compensation schemes perform independent observation and modeling of a single actuator, making it difficult to fully utilize the correlation information between multiple joints to identify the degradation state; and it cannot be decoupled from existing motion control strategies, etc.

[0014] To address the aforementioned problems in the prior art, this application proposes a torque compensation method for a joint actuator, the main process of which is as follows: Figure 1 As shown. The following explanation uses the lower limb joint actuator of a humanoid robot as an example. However, the joint actuator can also be used for bipedal robots, wheeled robots, robotic dogs, robotic arms, composite workstations with manipulators, and other similar devices. Figure 1 The example method mainly includes steps S1 to S3, which are detailed below: Step S1: Construct a saturated degradation physical model to characterize the parametric relationship between the joint actuator in the nominal state and the degradation state.

[0015] Existing actuator degradation modeling often employs ray-like models or black-box residual mapping, which cannot simultaneously characterize two different physical degradation phenomena: the attenuation of actuator torque output capability and the change in the shape of the torque saturation curve. These models lack a physical interpretive basis, hindering degradation state identification and failing to provide an analytical mathematical framework for subsequent physical-guided torque compensation. Therefore, this application adopts a two-parameter saturation-type degradation physical model to characterize the parameter relationship between the joint actuator in its nominal and degraded states. This approach uses the parameters in the degradation physical model to represent two independent degradation physical effects, establishing a degradation characterization system with clear physical meaning. This provides a clear modeling object for the subsequent coupled temporal degradation identification network and an analytical mathematical foundation for the subsequent inverse solution of torque compensation.

[0016] It should be noted that, in the embodiments of this application, the joint actuator can be the joint actuator of a bipedal robot, wheeled robot, robotic dog, robotic arm, or composite workstation with a manipulator arm. Under the healthy nominal operating condition of the robot joint actuator, the command torque and feedback torque satisfy a saturated nonlinear relationship. The saturation constraint of the actuator's inherent torque output is characterized by the tanh (hyperbolic tangent function), and the healthy state model is as follows: (Formula 1) in, For the first i Actual measured feedback torque of each lower limb joint actuator For the first i The rated maximum torque of each lower limb joint actuator is a known constant determined by the robot hardware manufacturer. To be distributed to the first i Command torque of a lower limb joint actuator It is a hyperbolic tangent saturation function used to characterize the torque output saturation characteristics at the physical level of the actuator.

[0017] When the joint actuator degrades, its torque output capability and saturation response curve change due to factors such as temperature rise, mechanical friction, reducer wear, and motor winding aging. Therefore, a two-parameter saturation-type degradation physical model is established. The degradation model is as follows: (Formula 2) in, It is the output gain. , This indicates that the actuator is in a healthy state. This indicates a decrease in the actuator's torque output capability, for example, This means that the actuator can only output 70% of its rated maximum torque; It is a saturation characteristic parameter. , For nominal saturation characteristics, This indicates that the linear operating range of the actuator is shrinking, and it will saturate at lower command torques. This indicates that the linear operating range of the actuator has been expanded.

[0018] It can be seen that when and When Equation 2 degenerates into Equation 1 for the healthy actuator, the model possesses self-consistent nominal regression properties.

[0019] If a degradation model is simply constructed without distinguishing parameter definitions and verifying nominal conditions, multiple degradation effects can easily be coupled together. Furthermore, the model may introduce unexpected biases under healthy operating conditions, reducing identification accuracy. Therefore, as an embodiment of this application, constructing a two-parameter saturated degradation physical model to characterize the parameter relationship between the joint actuator in its nominal and degraded states can be achieved through steps S31 to S33, detailed below: Step S31: Define the output gain to characterize the degree of attenuation of the maximum torque output capability of the joint actuator, and define the saturation characteristic parameter to characterize the change in the shape of the torque output saturation curve.

[0020] In the embodiments of this application, the output gain Specifically designed to address the decrease in maximum torque output capability caused by actuator hardware wear, this indicator directly reflects faults such as motor winding aging and permanent magnet demagnetization, which can lead to a reduction in the upper limit of torque. Saturation characteristic parameters. It is specifically designed to address the saturation response curve deformation caused by wear and increased friction in reducers. By using two independent parameters, it decouples two types of degradation phenomena with different physical causes, avoiding confusion between different degradation effects.

[0021] Step S32: Establish the saturation function mapping relationship of the feedback torque under the degenerate state with respect to the command torque, output gain, saturation characteristic parameters, and the rated maximum torque of the joint actuator.

[0022] The saturation function mapping is the mathematical relationship shown in Formula 2 above. This mapping relationship takes the input command torque as the independent variable and the output gain, saturation characteristic parameter, and rated maximum torque as model parameters to solve for the actuator feedback torque under degradation conditions. This mapping is the core mathematical carrier used by the subsequent identification network and compensation calculation.

[0023] Step S33: Verify the saturation function mapping relationship. When both the output gain and the saturation characteristic parameter are taken as nominal values, the saturation function mapping relationship is degenerated into the torque input-output relationship of a healthy actuator.

[0024] nominal value set to and Substituting the nominal parameters into Formula 2 above, the model is now equivalent to Formula 1 for a healthy actuator. This verification is used to ensure the model's logical consistency, avoid introducing additional model biases in healthy operating conditions where the actuator does not degenerate, and prevent interference with the identification and compensation under healthy conditions.

[0025] From the examples in steps S31 to S33, it can be seen that the construction of a two-parameter saturated degradation physical model used to characterize the parameter relationship between the joint actuator in the nominal state and the degradation state clearly defines two types of physical degradation parameters, establishes a saturation function mapping relationship, and verifies the characteristics of the model degrading to a healthy actuator under the nominal parameters, ensuring the physical self-consistency of the model and providing a rigorous physical basis for subsequent identification and compensation.

[0026] Step S2: Based on the collected temporal ontology perception signals of N joint actuators, construct observation tensors corresponding to the N joint actuators under a set time window. Input the N observation tensors into a coupled temporal degradation identification network. Adapt the coupled temporal degradation identification network to the saturated degradation physical model. After extracting the local temporal features of each joint actuator through causal temporal convolution, perform cross-actuator multi-head self-attention operation in the actuator dimension to model the torque redistribution coupling relationship between the N joint actuators. Output the preliminary estimates of the degradation parameters corresponding to the N joint actuators, where N is a positive integer. The coupled temporal degradation identification network is adapted to the saturated degradation physical model.

[0027] Robots are strongly coupled dynamic systems. When an actuator in a lower limb joint degenerates, the upper-level motion strategy will automatically redistribute torque among multiple lower limb joints to maintain the balance of the robot. However, traditional solutions analyze the signals of each actuator independently, which cannot utilize this cross-joint coupling information. The estimation of degradation parameters is easily affected by noise from single-joint sensors, resulting in insufficient accuracy of the identification results. In view of this, the present application can adopt the following technical solution: based on the collected temporal ontology perception signals of N joint actuators, construct observation tensors corresponding to the N joint actuators under a set time window, input the N observation tensors into a coupled temporal degradation identification network, adapt the coupled temporal degradation identification network to the saturated degradation physical model, extract the local temporal features of each joint actuator through causal temporal convolution, perform cross-actuator multi-head self-attention operation in the actuator dimension, model the torque redistribution coupling relationship between the N joint actuators, and output the preliminary estimated values ​​of the degradation parameters corresponding to the N joint actuators, where N is a positive integer, the coupled temporal degradation identification network adapts to the saturated degradation physical model, where the ontology perception signal refers to the state information collected by the humanoid robot's own sensors, including but not limited to joint command torque, feedback torque, joint angle, joint angular velocity, etc., and the coupled temporal degradation identification network is a neural network model based on deep learning, used to estimate the degradation parameters of each joint actuator online from the multi-joint temporal ontology perception signal. The above-mentioned solution of this application utilizes cross-actuator multi-head self-attention to explicitly mine the torque redistribution coupling features between multiple joints, integrates all lower limb joint temporal ontology perception information to carry out joint identification, improves the reliability of the preliminary estimation results of degradation parameters, and ensures the causal attributes of network inference through causal temporal convolution.

[0028] First, define the proprioception vector at a single time step for a single actuator. For the i-th lower limb joint actuator, the proprioception vector at time step t is... for: (Formula 3) in, for t Time of the first i Joint actuator command torque, for t Time of the first i Joint actuator feedback torque, for t Time of the first i The joint angles of the joints, for t Time of the first i Joint angular velocity.

[0029] A time window length of H = 30 steps is set. This window represents the use of ontology perception data from the current moment and the past 29 historical control steps, excluding future moment data to ensure causal inference in the network. Based on the above single actuator time sequence vectors, the observation tensors of all lower limb actuators are constructed: (Formula 4) Where H represents the total number of steps in the timing window, which can be 30 in this embodiment. n This represents the total number of actuators in the robot's lower limb joints. These are tensor dimensions, corresponding to the time step dimension, executor number dimension, and single executor perception feature dimension, respectively.

[0030] The overall architecture of the coupled temporal degradation identification network consists of three sub-modules: a causal temporal convolutional layer, a cross-actuator multi-head self-attention layer, and a shared prediction head. The network adapts to the saturated degradation physical model obtained in step S1, which is a two-parameter saturated degradation physical model. The network's learning objective is to estimate the degradation parameters, such as the output gain and saturation characteristic parameters, of this model. The following is a detailed explanation of each of the three sub-modules: the causal temporal convolutional layer, the cross-actuator multi-head self-attention layer, and the shared prediction head. 1) Causal Temporal Convolutional Layer: Its kernel size can be 3, stride can be 1, and output channel number can be 64. It independently performs one-dimensional causal convolution along the time dimension for each executor. The convolution kernel of the causal convolution only uses historical and current time data and does not introduce future time information, ensuring the causality of online inference. The convolution output obtains the local temporal feature vector corresponding to each executor. .

[0031] 2) Multi-head self-attention layer across actuators: This layer can employ 8 attention heads with a latent vector dimension of 128. It takes the set of local temporal feature vectors from all actuators as input and performs multi-head self-attention computation along the actuator dimension. This explicitly models the pairwise interactive coupling relationships between multiple joints caused by torque redistribution. The attention output is then average-pooled to obtain the global latent vector. This layer possesses permutation equivariance, meaning that changing the order of actuator inputs will not alter the output result, thus improving the network's adaptability to different robot hardware configurations.

[0032] 3) Shared prediction head: It can consist of two fully connected layers, with a hidden layer dimension of 64 and an output layer dimension of 2 (corresponding to respectively). and (estimated value), shared prediction head It receives the global latent vector, combines it with the learnable actuator index embedding, and outputs a preliminary estimate of the degradation parameters corresponding to each joint actuator. The parameter estimation output relationship is shown in the following formula 5: (Formula 5) in, This is a preliminary estimate of the output gain of the actuator at the i-th joint. This is a preliminary estimate of the saturation characteristic parameters of the i-th joint actuator. The parameter is A shared prediction head is used, and all executors share the weights of this prediction head, thereby improving sample utilization efficiency. i It is the actuator index number, which distinguishes different joints through learnable index embedding.

[0033] Given that training existing degradation identification networks requires collecting real degradation parameter labels, which is difficult to obtain from the robot's actual hardware, , Real-value labels are difficult to use efficiently for training the identification network in a simulation environment. Therefore, the coupled temporal degradation identification network in the above embodiment uses torque reconstruction loss for training, and its training process is described in steps S21 to S23: Step S21: Configure multiple sets of different degradation parameters for N joint actuators in the simulation environment, control the motion strategy to drive the robot to perform various types of motion tasks, and collect data pairs of command torque and feedback torque.

[0034] This application uses a multi-joint dynamics with Contact (MuJoCo) simulation environment to traverse multiple sets of degradation parameter combinations, covering single-joint degradation and simultaneous multi-joint degradation. Motion strategies execute various motion tasks such as walking, stepping in place, and single-leg support, and batch-collect command torque and feedback torque data pairs corresponding to the body perception as training datasets. This step does not require intrusive modifications to the underlying motor and reducer source code of the simulation; the degradation effect is achieved through a post-processing transformation layer.

[0035] It should be noted that, in the embodiments of this application, the motion strategy is the control algorithm executed by the upper-level motion controller of the robot, which outputs the expected command torque of each joint actuator, including but not limited to neural network strategies based on reinforcement learning, optimization strategies based on model predictive control, and feedback control laws based on classical control theory, etc.

[0036] Step S22: Input the observation data obtained from the simulation into the coupled temporal degradation identification network, and use the degradation parameters output by the network to reconstruct the reconstruction feedback torque.

[0037] Preliminary estimates of degradation parameters from network output , Substituting into the two-parameter degradation physical model represented by Equation 2 above, the reconstruction feedback torque is calculated. The calculation formula is as follows: (Formula 6) Step S23: Calculate the mean square error between the reconstructed feedback torque and the simulated measured feedback torque as the torque reconstruction loss. Update the network weights based on the torque reconstruction loss through backpropagation. The training process does not require real degradation parameter labels.

[0038] The torque reconstruction loss function is shown in Formula 7 below: (Formula 7) in, For torque reconstruction loss, n This represents the total number of lower limb actuators. This training method uses real-world measured feedback torque in a simulated environment. It does not require the dataset to contain real-world data. , The label is used to backpropagate the loss and update all weights of the identification network to complete the network training, relying only on instruction-feedback torque data pairs.

[0039] The coupled temporal degradation identification network in the examples of steps S21 to S23 above uses torque reconstruction loss to complete the training process, realizing network training without real value labels. It can complete network parameter learning only by relying on the command torque and feedback torque observation data, reducing the threshold for obtaining datasets and facilitating the generation of training samples in large quantities in the simulation environment.

[0040] Step S3: Based on the preliminary estimates of the degradation parameters corresponding to the N joint actuators, perform mathematical inverse solution on the saturated degradation physical model to obtain the torque compensation calculation formula. According to the torque compensation calculation formula, convert the expected torque output by the motion strategy into the compensated command torque.

[0041] Since the expected torque output by the upper-level motion strategy is designed for a healthy actuator, directly issuing this expected torque when the actuator has degraded will result in the actual output torque failing to meet the strategy's expectations. Therefore, it is necessary to perform an inverse mathematical solution on the saturated degradation physical model based on the preliminary estimate of the degradation parameters to obtain the torque compensation calculation formula, and then convert the expected torque into the compensated command torque according to this formula. This scheme obtains the compensation command based on the analytical inverse solution of the physical model, which differs from the residual compensation obtained by black-box network learning. It has a clear physical mechanism and is conducive to zero-shot generalization across different motion modes.

[0042] Specifically, firstly, the saturated degradation physical model in step S1 is mathematically inversely solved to obtain the desired torque. The feedback torque that should be output after degradation Solve for the corresponding command torque The following original compensation formula without safety constraints is obtained: (Formula 10) in, It is the command torque after compensation that needs to be solved. It is the first i The rated maximum torque of the joint actuator These are the smoothed saturation characteristic parameters output in step S3. It is the output gain after smoothing the output in step S3. It is the inverse hyperbolic tangent function. The inverse function of , whose effective domain is the interval (-1, 1) where the input must fall within the interval. It is the first output of the upper-level motion strategy i The desired torque of the joint actuator.

[0043] Given that the preliminary estimates of degradation parameters output by the coupled temporal degradation identification network are mixed with high-frequency random jitter caused by torque sensor measurement noise and instantaneous torque disturbances in the motion strategy, directly using these jittery preliminary estimates in subsequent torque compensation calculations would cause frequent jumps in the compensated command torque, inducing oscillations in robot joint control and hindering robot motion stability, this application adopts a technical solution of smoothing the preliminary estimates of degradation parameters of each joint actuator using exponential moving average (EMA) to obtain smoothed degradation parameters. Specifically, exponential moving average is used to perform temporal low-pass filtering on the preliminary estimates of degradation parameters output by the network, and smoothing operations are performed independently on the two estimates of output gain and saturation characteristic parameters. The corresponding calculation formulas are as follows: (Formula 8) (Formula 9) in, for t Time of the first i The output gain of each joint actuator after smoothing for t -1 moment i The smoothed output gain of each joint actuator, and the output result of the previous control step. for t The initial estimate of the output gain of the network at time step S2, i.e., the output result of step S2. for t Time of the first i Saturation characteristic parameters of each joint actuator after smoothing for t -1 moment i Saturation characteristic parameters of each joint actuator after smoothing for t Preliminary estimates of the saturation characteristic parameters of the network output at time step [time]. The smoothing coefficient is the exponential moving average; in this embodiment, the control period is 100Hz (10ms per step), and the selected value is... =0.95.

[0044] During the system power-on initialization phase, if there are no historical smoothing parameters, the initial value will be... and This corresponds to the nominal health status of the actuator. After this step, the smoothed degradation parameters are output. , This serves as the input for subsequent torque compensation calculations.

[0045] If a smoothing coefficient is directly assigned manually, it is easy for the coefficient to mismatch with the actual hardware characteristics. This can result in either insufficient filtering with excessive noise, or over-smoothing that fails to respond to true degradation. Therefore, Figure 1 The example method may also include the technical solutions exemplified by steps S111 to S113 below: Step S111: Obtain the physical timescale of the humanoid robot's joint actuator degradation and the controller's control cycle.

[0046] The physical timescale of actuator degradation characterizes the time required for significant performance degradation of the actuator. Degradation caused by temperature rise is generally on the order of seconds, while degradation caused by mechanical wear is on the order of minutes or even hours. The controller control cycle is an inherent operating parameter of the robot motion control hardware. In the embodiments of this application, the controller control cycle is 10ms (100Hz). These two types of parameters are the original inputs for determining the smoothing coefficient.

[0047] Step S112: Calculate the exponential moving average equivalent time constant based on the physical time scale and the controller control cycle.

[0048] Exponential moving average equivalent time constant The unit is the number of control steps, satisfying the relationship. The equivalent time constant needs to be less than the number of control steps corresponding to the degradation physical time scale, while it needs to be greater than the number of short-term fluctuation steps corresponding to the sensor noise, so as to balance noise suppression and degradation response speed.

[0049] Step S113: Determine the smoothing coefficient of the exponential moving average based on the equivalent time constant of the exponential moving average.

[0050] Substituting the target equivalent time constant The smoothing coefficient is obtained by reverse calculation. Simulation experiments verified that the equivalent time constant can effectively suppress sensor noise (improving the signal-to-noise ratio by approximately 15 dB) within a range of 10 to 30 steps, while ensuring that the response delay to second-level degradation does not exceed 0.5 seconds. This embodiment takes the middle value, selecting an equivalent time constant of 20 steps, and obtains the corresponding... .

[0051] The technical solutions exemplified in steps S111 to S113 above determine the smoothing coefficient based on the actual time scale of the robot hardware, so that the filtering characteristics of the exponential moving average can be adapted to the slowly changing physical characteristics of the actuator degradation, thus avoiding performance imbalance caused by arbitrary parameter settings.

[0052] because The input must be strictly within the interval (-1, 1) when the desired torque Approaching or even exceeding the maximum output torque of the actuator after degradation , It will approach As the gradient of the function approaches infinity, even a small perturbation in the parameter estimation can cause... Given the drastic change, the technical solution for trimming the first-layer security boundary domain can be implemented using the following steps S51 to S53: Step S51: Obtain the desired torque, the output gain in the smoothed degradation parameters, and the rated maximum torque of the joint actuator, and calculate the normalized ratio of the desired torque.

[0053] Normalized ratio calculation formula: , This is the normalized ratio of the desired torque to the maximum output torque after degradation.

[0054] Step S52: Use a preset small positive value to shrink and trim the upper and lower boundaries of the normalized ratio.

[0055] When the input approaches ±1, the gradient of the arctanh function tends to infinity. With ε = 0.02, the maximum absolute value of the input is 0.98, corresponding to an arctanh gradient of approximately 25, which is much smaller than the gradient of 500 when the input is 0.999, effectively suppressing the risk of numerical explosion. Simulation verification shows that ε can guarantee numerical stability within the range of 0.01 to 0.05. Therefore, the embodiments of this application can preset a small positive value. The effective interval of the ratio is narrowed to ( , ),Should It is determined by the gradient characteristics of the inverse hyperbolic tangent function itself and does not change with the actuator hardware model or the degree of degradation.

[0056] Step S53: Substitute the clipped normalized ratio into the inverse hyperbolic tangent function obtained from the inverse solution of the two-parameter saturated degenerate physical model to avoid the function gradient explosion causing drastic changes in the compensation command.

[0057] The compensation formula after domain clipping is as follows: (Formula 11) in, This is a truncation function that forces the input value to be limited to [...]. Within the range, In this embodiment of the application, a preset small positive value is used to define the domain shrinkage. .

[0058] As can be seen from the domain pruning techniques described in steps S51 to S53 above, by performing inward pruning on the inverse function input, the gradient explosion phenomenon is suppressed, ensuring the numerical stability of the compensation calculation process.

[0059] The intermediate command torque after inverse computation may exceed the physical output range allowed by the actuator hardware. Directly issuing this command could trigger hardware protection or even damage the actuator. Therefore, a hard limit on the command amplitude for the second-layer safety boundary is implemented through steps S71 to S73, as detailed below: Step S71: Obtain the intermediate command torque output by the inverse hyperbolic tangent function.

[0060] The intermediate command torque is calculated using the formula 11 above. Hardware amplitude constraints have not yet been implemented.

[0061] Step S72: Compare the intermediate command torque with the rated maximum torque of the corresponding joint actuator.

[0062] intermediate command torque and , Perform a size comparison to identify whether the torque exceeds the hardware's allowable range.

[0063] Step S73: Cut the intermediate command torque to between the negative rated maximum torque and the positive rated maximum torque, and output the final compensated command torque.

[0064] Command amplitude hard limit calculation as follows (Formula 12).

[0065] The above steps S73 to S73 example demonstrate hardware amplitude hard limiting of the compensation instruction to ensure that the output instruction torque is always within the physically achievable range of the actuator, thus forming a hardware-level safety protection.

[0066] In summary, by combining domain clipping and instruction amplitude hard limiting as two layers of safety boundary processing, the problems of singular inverse function values ​​and hardware torque exceeding limits are solved respectively, unreasonable instruction torque output is avoided, and the safety of engineering operation in the compensation process is ensured.

[0067] To achieve continuous online identification and torque compensation, a sending and iterative process can be introduced after step S3. That is, the compensated command torque is sent to the corresponding joint actuator, new data of robot body perception is collected to update the observation tensor, and steps S2 to S5 are executed cyclically to achieve continuous online identification and torque compensation of joint actuator degradation.

[0068] During robot movement, actuator degradation is a slow and continuous process that evolves over time. Simultaneously, the body perception window needs to continuously incorporate the latest perception data and discard the oldest historical data to ensure that the observation tensor always reflects the robot's current true state, enabling uninterrupted online identification and compensation. To further restrict the temporal observation window configuration and causal constraints based on the observation tensor update logic, Figure 1 The example method may also include the technical solutions exemplified in steps S91 to S93, as detailed below: Step S91: Configure the time step of the time series observation window, and select historical data of robot body perception to construct the observation tensor of the set time window.

[0069] Window time step in this embodiment H The configuration is set to 30; in each iteration, historical ontology-sensing data of the corresponding length is selected to assemble the observation tensor. .

[0070] Step S92: Filter the data sources for the observation tensor and construct the observation tensor using only the ontology perception data at the current time and historical time.

[0071] When updating the observation tensor by sliding, only the ontology-aware samples of the current moment and historical moments that have already been collected are used, and virtual data of future moments that have not yet occurred are not introduced.

[0072] Step S93: Filter the data perceived in future moments to ensure the causality of the coupled temporal degradation identification network.

[0073] Whether it's window sliding updates or network inference, any perceptual data from future moments is removed; this avoids leaking future information during inference and ensures that the algorithm can be used for real-time online operation of a real robot.

[0074] The technical solutions exemplified in steps S91 to S93 above strictly constrain the data source of the observation tensor, ensuring that the entire algorithm possesses causal properties and meets the engineering constraints of real-world robot online deployment.

[0075] In fact, the complete implementation process of step S5 above is as follows: 1) The compensated command torque output in step S4 is sent to the corresponding lower limb joint actuators, and the hardware driver layer executes the torque output; 2) The robot's various sensors collect a new round of body perception data in real time, that is, the new command torque, feedback torque, joint angle, and joint angular velocity of each joint; 3) The newly collected body perception data is used to update the observation tensor. The timing window maintains a length of H=30 steps, discarding the oldest set of historical observations and pushing in the newly acquired observation vector at the current moment to complete the observation tensor sliding update. This update logic follows the window configuration constraints of the above embodiment; 4) The program flow jumps back to step S2 and enters the next control cycle, repeatedly executing all processes from step S2 to step S5. This middleware layer serves as a software intermediary between the upper-level motion strategy and the lower-level actuator. Throughout the entire loop, it does not require modification of the internal logic of the upper-level motion strategy, nor does it interrupt the motion task being performed by the robot. By constructing a complete closed-loop data flow, it enables continuous iterative updates of degradation parameters and torque compensation while the robot is continuously operating; it achieves plug-and-play functionality and compatibility with any type of upper-level motion control strategy.

[0076] Furthermore, recognition networks trained solely on simulation datasets suffer from simulation-to-real-world transfer bias, making it difficult to obtain degraded real-value labels in a real-world environment. Offline retraining necessitates system shutdown and interruption of operational tasks, failing to meet the robot's continuous operation requirements. Therefore, Figure 1 The example method may also include the technical solution for the online fine-tuning process during the deployment phase, i.e., steps S41 to S43, as detailed below: Step S41: Continuously collect real command torque and feedback torque during the actual operation of the humanoid robot.

[0077] During actual operation of the L1 humanoid robot, the body perception and acquisition link is reused to continuously collect the command torque output by the real robot. With measured feedback torque The acquired data is stored in a circular data buffer; the circular buffer retains real-machine observation samples from the most recent control steps as the data source for online fine-tuning. This part of the data acquisition and the body perception acquisition hardware path of the main compensation loop (steps S2 to S5) are completely reused, and no additional sensor hardware is required.

[0078] Step S42: Reuse the torque reconstruction loss and fine-tune the network weights of the coupled temporal degradation identification network online using real-time acquired data.

[0079] Step S42 can directly reuse the torque reconstruction loss function from Formula 7 above: (Formula 13) in, This refers to the torque reconstruction loss used during the online fine-tuning phase of the actual machine. The reconstruction feedback torque is obtained by reconstructing the current network weights and the real machine instruction torque. The feedback torque is obtained from actual measurements of the actual machine. n This represents the total number of actuators for the lower limb joints.

[0080] Online fine-tuning uses a small learning rate, based on Loss backpropagation is performed, and the weights of the coupled temporal degradation identification network are updated in small increments. Note that online fine-tuning is an asynchronous background task that does not compete with the main inference flow for high-priority real-time control cycles; the inference chain still uses the network weights before or partially after the update to ensure real-time control. This process inherits the core characteristic of the simulation training phase—"no need for real degradation parameter labels"—and still does not require them in a real-device environment. , The actual value can be updated by simply relying on the command-feedback torque data.

[0081] Step S43: During the online fine-tuning process, the network inference link is kept uninterrupted and the robot's current motion task is not interrupted.

[0082] By setting the network weight update as a background asynchronous task, the robot's main control loop continues to execute the complete process described in steps S1 to S5 above, continuously outputting degradation parameter estimates and compensation command torques. The weight update will not stop the robot's movement, and there is no need to stop the robot or reload the model. The robot can continue to complete business tasks such as inspection and walking.

[0083] The technical solutions in steps S41 to S43 above reuse the validated torque reconstruction loss paradigm. Under the condition that the robot does not stop, the network weights are fine-tuned using real-time data from the real machine, which reduces the domain offset between the simulation and the real machine and further improves the accuracy of degradation parameter estimation on the real machine.

[0084] Considering that traditional simulations of actuator degradation require modifying the underlying motor and reducer dynamics source code, which is highly invasive and difficult to maintain, relying on artificially inducing wear and degradation using real physical prototypes results in high sample acquisition costs, poor repeatability, and an inability to generate diverse degradation condition training samples in large quantities. Therefore, Figure 1 The example method may also include the technical solutions exemplified in steps S81 to S83 as follows: Step S81: Keep the underlying motor and reducer dynamics models of the simulation environment unchanged, and insert a degenerate transformation layer before the command torque is transmitted to the joint.

[0085] This application uses the MuJoCo simulation engine without modifying the underlying motor dynamics code of the MuJoCo actuator, or the transmission and friction parameters of the reducer. An independent degenerate transformation intermediate layer is inserted at the software level after the motion strategy outputs the command torque and before the torque is applied to the rigid body joint. This design is a non-intrusive modification that does not require altering the simulation engine source code, facilitating cross-simulation platform portability.

[0086] Step S82: Use the two-parameter saturated degradation physical model to transform the input command torque and obtain the transformed output torque.

[0087] Specifically, the two-parameter saturated degradation physical model of Equation 2 can be called, and the simulation can be pre-configured with the corresponding actuators for each group. , The original command torque of the strategy Input the degradation model, calculate the output torque after transformation. .

[0088] Step S83: Input the transformed output torque into the simulation environment as the actual driving torque of the joint.

[0089] The result calculated after the degradation transformation layer The actuator model in the MuJoCo simulation engine serves as the actual driving torque for the joints, driving the simulated humanoid robot to complete its movements.

[0090] As can be seen from the technical solutions in the examples of steps S81 to S83 above, the non-intrusive post-processing realizes the simulation of actuator degradation in the simulation environment without modifying the underlying dynamics of the simulation. It can easily configure various degradation combinations of single joints and multi-joints, and generate a large number of training datasets with degradation effects for simulation training and algorithm ablation verification of coupled temporal degradation identification networks.

[0091] Traditional actuator fault early warning systems mostly rely on additional sensors such as vibration and temperature sensors, increasing hardware costs. Many existing degradation compensation algorithms only output compensation torque, not interpretable degradation states, making it impossible to reuse control link information for equipment health management. Therefore, in order to reuse the physically meaningful degradation parameter timing output by this compensation method and achieve online health monitoring and early warning for actuators, Figure 1 The example method may also include the technical solutions of steps S61 to S63, which are described in detail below.

[0092] Step S61: Continuously record the smoothed degradation parameters updated over time to generate a time series sequence of degradation parameters.

[0093] In each control cycle of the robot's main loop, the S3 outputs the smoothed degradation parameters. , The degradation parameters of each joint over time are saved in a non-volatile or circular time-series storage unit, constructing a time-series sequence of degradation parameters. This time-series sequence directly reuses the parameters output by the compensation calculation, eliminating the need for additional hardware sensors to collect fault information.

[0094] Step S62: Monitor the changing trend of the time series of degradation parameters.

[0095] Perform trend analysis on time series to observe output gain. The decreasing trend over time corresponds to the reduction in torque output capability caused by motor winding aging and permanent magnet demagnetization; on the other hand, the saturation characteristic parameters are observed. The deviation from the nominal value of 1 corresponds to saturated response deformation phenomena such as wear and increased friction in the reducer.

[0096] Step S63: When the output gain is lower than the first threshold and / or the saturation characteristic parameter deviates from the nominal value to the second threshold, a joint actuator maintenance warning signal is triggered.

[0097] A first threshold and a second threshold are pre-configured. For example, the first threshold is set to 0.7. When a certain joint... This indicates that the joint's torque output capability has decreased to below 70% of the rated value; the second threshold is set to 0.3, i.e. It indicates that the saturation characteristics have undergone significant distortion; if any condition is met, a maintenance warning signal will be output; the warning signal can be reported to the host computer operation and maintenance system and the human-machine interface.

[0098] The technical solutions in the examples of steps S61 to S63 above achieve online health monitoring and early warning of actuators by reusing the degradation parameter timing with physical meaning output by this compensation method. No additional sensors are required, and the data flow link of "control compensation - equipment maintenance" is opened up to support predictive maintenance services.

[0099] Furthermore, considering that although the technical solutions exemplified in steps S61 to S63 above can trigger alarms after an event, responding to degradation that has already occurred still cannot predict future degradation trends, leaving a short window for equipment maintenance. Therefore, Figure 1 The example method may also include the technical solutions exemplified in steps S101 to S103: Step S101: Input the time series of degradation parameters into the trend prediction module to predict the evolution trajectory of degradation parameters of each joint actuator in the future time period.

[0100] The degradation parameter time series obtained in the above embodiments is input into the trend prediction module. The trend prediction module can use an RNN (Recurrent Neural Network) time series network or a simple multinomial extrapolation model to output each joint within a future time window. , The module predicts the evolution trajectory. As a background task, it does not need to run within a 10ms high real-time control cycle; it can run on the host computer or the robot's non-real-time kernel.

[0101] Step S102: Evaluate the remaining working time of each joint actuator based on the predicted evolution trajectory of the degradation parameters of each joint actuator.

[0102] By comparing the predicted degradation trajectory with the aforementioned warning thresholds, the future moment when the parameters will touch the warning thresholds can be estimated, thereby assessing the remaining working time of the joint actuator and providing a reference for maintenance personnel.

[0103] Step S103: Based on the predicted evolution trajectory of the degradation parameters of each joint actuator, adjust the trigger threshold of the maintenance warning in advance.

[0104] If the prediction results indicate that the degeneration of a certain joint will deteriorate rapidly, the warning sensitivity will be dynamically increased in advance, and the first threshold and the second threshold in the above embodiment will be corrected to achieve predictive warning, rather than simply waiting for the parameters to exceed the limits before issuing an alarm.

[0105] The technical solutions in the examples of steps S101 to S103 above are based on historical degradation time series to extrapolate and predict, assess the remaining working time of the joints, and dynamically adjust the maintenance warning threshold; the alarm is upgraded from a post-event alarm to a predictive maintenance prompt, further improving the safety of robot equipment operation and maintenance.

[0106] From the above appendix Figure 1The torque compensation method for the example joint actuators demonstrates that, on the one hand, by constructing a saturated degradation physical model to characterize the parameter relationship between the joint actuators in their nominal and degraded states, a clear physical modeling object and an analytical mathematical carrier are provided for degradation state identification and torque compensation. On the other hand, by performing cross-actuator multi-head self-attention operations at the actuator dimension, the torque redistribution coupling relationship between N joint actuators is explicitly modeled, and preliminary estimates of the degradation parameters of each joint actuator are output. Instead of performing independent signal analysis on each actuator, degradation identification is carried out using the objectively existing torque redistribution coupling relationship between robot joints. This allows for the acquisition of preliminary estimates of the degradation parameters of each joint actuator based on multi-joint body perception information. Thirdly, based on the preliminary estimates of the degradation parameters, a two-parameter saturated degradation physical model is applied to the two-parameter saturated degradation physical model. The model undergoes mathematical inverse kinematics to obtain the torque compensation calculation formula. Based on this formula, the desired torque output by the motion strategy is converted into the compensated command torque. This processing method, based on the inverse kinematics of the physical model to obtain the compensation command, differs from black-box compensation mapping. It also incorporates safety boundary handling to avoid control risks caused by numerical anomalies in the inverse function. The desired torque output by the motion strategy is converted into a command torque adapted to the degenerate actuator, serving as an intermediate processing flow between the motion strategy and the actuator. Without modifying the motion strategy itself, it can continuously iterate and complete degradation identification and torque compensation during robot task execution, making the actual torque output of the degenerate joint actuator approach the desired torque of the motion strategy. This improves the robot's motion control performance under actuator degradation conditions and avoids motion accuracy reduction and instability problems caused by actuator degradation. In summary, the technical solution of this application utilizes multi-joint coupling for identification, without modifying the motion strategy, and compensates for actuator degradation online, thereby improving the stability of robot motion control.

[0107] Please see Figure 2 As shown, in one embodiment, a torque compensation device for a joint actuator is provided. The device may include a construction module 201, an estimation module 202, and a conversion module 203, as detailed below: Module 201 is used to construct a saturated degradation physical model that characterizes the parametric relationship between the joint actuator in the nominal state and the degradation state; The estimation module 202 is used to construct observation tensors corresponding to the N joint actuators under a set time window based on the collected temporal ontology perception signals of the N joint actuators. The N observation tensors are input into the coupled temporal degradation identification network. After extracting the local temporal features of each joint actuator through causal temporal convolution, cross-actuator multi-head self-attention operation is carried out in the actuator dimension to model the torque redistribution coupling relationship between the N joint actuators. The preliminary estimates of the degradation parameters corresponding to the N joint actuators are output, where N is a positive integer. The coupled temporal degradation identification network is adapted to the saturated degradation physical model. The conversion module 203 is used to perform mathematical inverse solution on the saturated degradation physical model based on the preliminary estimated values ​​of the degradation parameters corresponding to the N joint actuators to obtain the torque compensation calculation formula. According to the torque compensation calculation formula, the expected torque output by the motion strategy is converted into the compensated command torque.

[0108] From the above appendix Figure 2 As can be seen from the torque compensation device of the example joint actuator, on the one hand, by constructing a saturated degradation physical model to characterize the parameter relationship between the joint actuator in the nominal state and the degradation state, a clear physical modeling object and an analytical mathematical carrier are provided for degradation state identification and torque compensation; on the other hand, by performing cross-actuator multi-head self-attention operation in the actuator dimension, the torque redistribution coupling relationship between N joint actuators is explicitly modeled and the preliminary estimate of the degradation parameters of each joint actuator is output. Instead of performing independent signal analysis on each actuator, degradation identification is carried out by utilizing the objective torque redistribution coupling relationship between robot joints, and the preliminary estimate of the degradation parameters of each joint actuator can be obtained based on the multi-joint body perception information; thirdly, based on the preliminary estimate of the degradation parameters, the two-parameter saturated degradation physical model is applied to the two-parameter saturated degradation physical model. The model undergoes mathematical inverse kinematics to obtain the torque compensation calculation formula. Based on this formula, the desired torque output by the motion strategy is converted into the compensated command torque. This processing method, based on the inverse kinematics of the physical model to obtain the compensation command, differs from black-box compensation mapping. It also incorporates safety boundary handling to avoid control risks caused by numerical anomalies in the inverse function. The desired torque output by the motion strategy is converted into a command torque adapted to the degenerate actuator, serving as an intermediate processing flow between the motion strategy and the actuator. Without modifying the motion strategy itself, it can continuously iterate and complete degradation identification and torque compensation during robot task execution, making the actual torque output of the degenerate joint actuator approach the desired torque of the motion strategy. This improves the robot's motion control performance under actuator degradation conditions and avoids motion accuracy reduction and instability problems caused by actuator degradation. In summary, the technical solution of this application utilizes multi-joint coupling for identification, without modifying the motion strategy, and compensates for actuator degradation online, thereby improving the stability of robot motion control.

[0109] In one embodiment, a humanoid robot is provided, the internal structure of which can be shown in the following diagram: Figure 3As shown, the humanoid robot includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface allows for communication with external clients via a network. When executed by the processor, the computer program implements the functions or steps of a torque compensation method for a joint actuator.

[0110] In one embodiment, a humanoid robot is proposed, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, performs the following steps: Construct a saturated degradation physical model to characterize the parametric relationship between the joint actuator in the nominal state and the degradation state; Based on the temporal ontology perception signals of N joint actuators, observation tensors corresponding to the N joint actuators are constructed under a set time window. The N observation tensors are input into a coupled temporal degradation identification network. After extracting the local temporal features of each joint actuator through causal temporal convolution, cross-actuator multi-head self-attention operation is carried out in the actuator dimension to model the torque redistribution coupling relationship between the N joint actuators. The preliminary estimates of the degradation parameters corresponding to the N joint actuators are output, where N is a positive integer. The coupled temporal degradation identification network is adapted to the saturated degradation physical model. Based on the preliminary estimates of the degradation parameters corresponding to the N joint actuators, the torque compensation calculation formula is obtained by mathematical inverse solution of the saturated degradation physical model. According to the torque compensation calculation formula, the expected torque output by the motion strategy is converted into the compensated command torque.

[0111] The aforementioned computer program utilizes multi-joint coupling for identification, eliminating the need to modify the motion strategy, and compensates for actuator degradation online, thereby improving the stability of robot motion control.

[0112] In one embodiment, a computer-readable storage medium is provided that stores a computer program, which, when executed by a processor, performs the following steps: Construct a saturated degradation physical model to characterize the parametric relationship between the joint actuator in the nominal state and the degradation state; Based on the temporal ontology perception signals of N joint actuators, observation tensors corresponding to the N joint actuators are constructed under a set time window. The N observation tensors are input into a coupled temporal degradation identification network. After extracting the local temporal features of each joint actuator through causal temporal convolution, cross-actuator multi-head self-attention operation is carried out in the actuator dimension to model the torque redistribution coupling relationship between the N joint actuators. The preliminary estimates of the degradation parameters corresponding to the N joint actuators are output, where N is a positive integer. The coupled temporal degradation identification network is adapted to the saturated degradation physical model. Based on the preliminary estimates of the degradation parameters corresponding to the N joint actuators, the torque compensation calculation formula is obtained by mathematical inverse solution of the saturated degradation physical model. According to the torque compensation calculation formula, the expected torque output by the motion strategy is converted into the compensated command torque.

[0113] The above steps, implemented by the computer program when executed by the processor, utilize multi-joint coupling for identification, eliminating the need to modify the motion strategy, compensating for actuator degradation online, and improving the stability of robot motion control.

[0114] It should be noted that the functions or steps that the computer-readable storage medium or humanoid robot can achieve are described in the relevant descriptions of the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0115] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), Synchlink, DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0116] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0117] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A torque compensation method for a joint actuator, characterized in that, The method includes: Construct a saturated degradation physical model to characterize the parametric relationship between the joint actuator in the nominal state and the degradation state; Based on the collected temporal ontology perception signals of N joint actuators, observation tensors corresponding to the N joint actuators are constructed under a set time window. The N observation tensors are input into a coupled temporal degradation identification network. After extracting the local temporal features of each joint actuator through causal temporal convolution, cross-actuator multi-head self-attention operation is performed in the actuator dimension to model the torque redistribution coupling relationship between the N joint actuators. The preliminary estimates of the degradation parameters corresponding to the N joint actuators are output, where N is a positive integer. The coupled temporal degradation identification network is adapted to the saturated degradation physical model. Based on the preliminary estimates of the degradation parameters corresponding to the N joint actuators, the saturated degradation physical model is mathematically inversely solved to obtain the torque compensation calculation formula. According to the torque compensation calculation formula, the expected torque output by the motion strategy is converted into the compensated command torque.

2. The torque compensation method for the joint actuator according to claim 1, characterized in that, based on the preliminary estimates of the degradation parameters corresponding to the N joint actuators, the torque compensation calculation formula is obtained by mathematical inverse solution of the saturated degradation physical model, including: The preliminary estimates of the degradation parameters corresponding to the N joint actuators are smoothed to obtain the smoothed degradation parameters. The torque compensation calculation formula is obtained by performing mathematical inverse solution on the saturated degradation physical model based on the smoothed degradation parameters.

3. The torque compensation method for the joint actuator according to claim 1, characterized in that, the step of converting the desired torque output by the motion strategy into a compensated command torque according to the torque compensation calculation formula includes: By combining domain clipping and hard limiting of command amplitude for safety boundary processing, the desired torque output by the motion strategy is converted into the compensated command torque according to the torque compensation calculation formula.

4. The torque compensation method for the joint actuator according to claim 1, characterized in that the method further includes: The compensated command torque is sent to the corresponding joint actuator, and the new data of the robot's body perception is collected to update the observation tensor. The process of collecting the temporal body perception signals of all joint actuators of the humanoid robot and updating the observation tensor with the new data of the robot's body perception is executed cyclically to realize the continuous online identification and torque compensation of joint actuator degradation.

5. The torque compensation method for the joint actuator according to claim 1, characterized in that, The coupled temporal degradation identification network is trained using a torque reconstruction loss, which includes: In a simulation environment, multiple sets of different degradation parameters are configured for N joint actuators. The motion strategy is controlled to drive the robot to perform various types of motion tasks, and data pairs of command torque and feedback torque are collected. The observation data obtained from the simulation is input into the coupled temporal degradation identification network, and the reconstruction feedback torque is obtained by reconstructing the degradation parameters output by the network. The mean square error between the reconstructed feedback torque and the simulated measured feedback torque is calculated as the torque reconstruction loss. The network weights are updated by backpropagation based on the torque reconstruction loss, and the training process does not require real degradation parameter labels.

6. The torque compensation method for the joint actuator according to claim 1, characterized in that, The saturated degradation physical model constructed to characterize the parametric relationship between the joint actuator in the nominal state and the degradation state includes: The output gain is defined to characterize the degree of attenuation of the maximum torque output capability of the joint actuator, and the saturation characteristic parameter is defined to characterize the change in the shape of the torque output saturation curve. Establish a saturation function mapping relationship between the feedback torque under degenerate state and the command torque, output gain, saturation characteristic parameters, and the rated maximum torque of the joint actuator; Verify the mapping relationship. When both the output gain and the saturation characteristic parameter are at their nominal values, the saturation function mapping relationship degenerates into the torque input-output relationship of a healthy actuator.

7. The torque compensation method for the joint actuator according to claim 5, characterized in that, The method further includes: During the actual operation of the robot, real command torque and feedback torque are continuously collected; The torque reconstruction loss is reused, and the network weights of the coupled temporal degradation identification network are fine-tuned online using real-time acquired data. The online fine-tuning process keeps the network inference link uninterrupted and does not interrupt the robot's current motion task.

8. The torque compensation method for the joint actuator according to claim 2, characterized in that, The method further includes: Continuously record the smoothed degradation parameters updated over time to generate a time series sequence of degradation parameters; Monitor the changing trends of time-series degradation parameters; When the output gain is lower than the first threshold and / or the saturation characteristic parameter deviates from the nominal value to the second threshold, a joint actuator maintenance warning signal is triggered.

9. The torque compensation method for the joint actuator according to claim 6, characterized in that, The method further includes: Keep the underlying motor and reducer dynamics models of the simulation environment unchanged, and insert a degenerate transformation layer before the command torque is transmitted to the joint; The input command torque is transformed by calling a two-parameter saturated degradation physical model to obtain the transformed output torque; The transformed output torque is used as the actual driving torque of the joint in the simulation environment.

10. The torque compensation method for the joint actuator according to claim 8, characterized in that, The method further includes: Input the time series of degradation parameters into the trend prediction module to predict the evolution trajectory of degradation parameters of each joint actuator in the future time period; The remaining working time of each joint actuator is assessed based on the predicted evolution trajectory of the degradation parameters of each joint actuator. By combining the predicted evolution trajectory of the degradation parameters of each joint actuator, the trigger threshold for maintenance warnings can be adjusted in advance.

11. A torque compensation device for a joint actuator, characterized in that, The device includes: The building module is used to construct a saturated degradation physical model that characterizes the parametric relationship between the joint actuator in the nominal state and the degradation state. An estimation module is used to construct observation tensors corresponding to the N joint actuators under a set time window based on the collected temporal ontology perception signals of the N joint actuators. The N observation tensors are input into a coupled temporal degradation identification network. After extracting the local temporal features of each joint actuator through causal temporal convolution, cross-actuator multi-head self-attention operation is performed in the actuator dimension to model the torque redistribution coupling relationship between the N joint actuators. The module outputs preliminary estimates of the degradation parameters corresponding to the N joint actuators, where N is a positive integer. The coupled temporal degradation identification network is adapted to the saturated degradation physical model. The conversion module is used to perform mathematical inverse solution on the saturated degradation physical model based on the preliminary estimated values ​​of the degradation parameters corresponding to the N joint actuators to obtain the torque compensation calculation formula, and convert the expected torque output by the motion strategy into the compensated command torque according to the torque compensation calculation formula.

12. A humanoid robot, the humanoid robot comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the torque compensation method for the joint actuator as described in any one of claims 1 to 10.

13. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the torque compensation method for the joint actuator as described in any one of claims 1 to 10.

Images