Robot gait control method, system and robot

Abstract

The application discloses a robot gait control method, a system and a robot. The method comprises the following steps: constructing a model for the robot, wherein the model satisfies the mapping relationship between the normal ground reaction force of the sole, the joint position, the joint speed and the joint acceleration, and the model satisfies the physical consistency constraint; constructing a physical information neural network according to the model, wherein the physical information neural network satisfies the inverse dynamics constraint; converging to the minimum value of a joint loss function as the target, and optimizing and training the physical information neural network; when the robot is running: outputting an initial control instruction according to a preset control strategy; obtaining the predicted value of the normal ground reaction force of the sole of the robot by using the trained physical information neural network, and adding the predicted value to the total reward; and correcting the initial control instruction by using the total reward to obtain an optimal control instruction for controlling the robot. The application realizes the low-impact, low-noise and stable walking of the robot.

Description

Technical Field

[0001] This invention belongs to the field of robot control, and more specifically, relates to a robot gait control method, system, and robot. Background Technology

[0002] As humanoid robots gradually enter human-robot co-location environments such as homes, hospitals, and offices, robots not only need to perform basic locomotion tasks such as stable walking, speed tracking, and disturbance resistance, but also need to reduce the impact, vibration, and noise generated when their feet contact the ground. The ground reaction force during foot-to-ground contact is a significant source of structural vibration and foot noise; excessive instantaneous impact can degrade the user experience and potentially accelerate wear on joints, reducers, foot structures, and ground materials. Currently, reinforcement learning-based gait control methods for humanoid robots primarily use speed tracking, posture stability, energy consumption constraints, and joint smoothness as the main reward items. For quiet or low-impact walking, common methods often use kinematic proxy indicators such as foot-to-ground velocity, foot acceleration, or change in contact velocity as indirect penalties. These methods do not directly constrain the foot-to-ground contact force, easily leading to problems such as "reducing noise at the expense of gait stability" or "maintaining speed but still experiencing significant ground impact."

[0003] One approach involves obtaining ground reaction forces through plantar force sensors, force plates, pressure insoles, or external force measuring devices, and then using the measured force signals for control or training. However, these force sensors are expensive, have limited installation space, insufficient impact resistance and long-term reliability, and the sensor signals are easily affected by footwear, ground materials, temperature drift, and changes in contact position. Directly using noisy force sensor signals for reinforcement learning rewards may lead to unstable rewards, difficulties in policy convergence, or training failure.

[0004] Another approach estimates contact forces based on robot dynamics models and inverse dynamic equations. However, humanoid robots are characterized by high degrees of freedom, underactuation, frequent contact switching, and ill-conditioned Jacobian matrices. Directly solving for contact forces is easily affected by model errors, single-leg support, changes in contact points, and near-singular configurations. In particular, when the robot wears different footwear, the foot geometry, effective leg length, contact stiffness, damping, and contact area change, making it difficult for traditional fixed models to maintain stable and accurate force estimates.

[0005] In recent years, neural networks have been used to predict ground reaction forces from proprioceptive signals. However, purely data-driven models often lack physical consistency and are prone to generating non-physical outputs outside the training distribution. For example, they may predict a large normal force during the swaying phase, the contact forces of the left and right feet may not match the dynamic equilibrium relationship, and jitter may be predicted during the contact switching phase. These problems reduce the reliability of the contact force estimation and further affect gait training based on the estimation results. Therefore, there is an urgent need for a ground reaction force prediction method that does not rely on external force sensors during the deployment phase, while possessing both physical consistency and data adaptability. This method should be combined with reinforcement learning gait training to enable humanoid robots to achieve low-impact, low-noise, and stable walking. Summary of the Invention

[0006] The main objective of this invention is to provide a robot gait control method, system, and robot to overcome the shortcomings of the prior art.

[0007] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: The first aspect of this invention provides a robot gait control method, including a network construction and training phase S1-S3 and a control phase S4-S5. S1 involves constructing a model for the robot that satisfies physical consistency constraints, representing the mapping relationship between the plantar normal ground reaction force, joint position, joint velocity, and joint acceleration. S2 involves constructing a physical information neural network (PEN) that satisfies inverse dynamics constraints based on the model. The inputs of the PSN are joint position, joint velocity, and joint acceleration, and the output is the plantar normal ground reaction force. S3 involves optimizing and training the PSN using pre-collected sample data, with the objective of minimizing a joint loss function comprising supervised regression error, inverse dynamics residual, swing phase zero force constraint, temporal smoothing constraint, contact phase consistency loss, impact peak weighted loss, and impulse consistency loss. S4 involves outputting initial control commands according to a preset control strategy during robot operation; obtaining a predicted value of the robot's plantar normal ground reaction force using the trained PSN, and adding the predicted value to the total reward. S5 involves correcting the initial control commands using the total reward to obtain the optimal control commands for controlling the robot.

[0008] Preferably, S1 specifically includes: transforming the robot's inverse dynamics equations using the damped least squares regularized pseudo-inverse method to obtain a first model describing the relationship between the original plantar normal reaction force to the ground and the joint position, joint velocity, and joint acceleration; projecting the original plantar normal reaction force to the ground in the first model onto a non-negative orthogonal domain to obtain a model that satisfies physical consistency constraints and characterizes the mapping relationship between the plantar normal reaction force to the ground, joint position, joint velocity, and joint acceleration.

[0009] Preferably, the contact phase consistency loss is: ; in, For the contact phase consistency loss, For the Sigmoid function, The slope coefficient, The normal ground reaction force of the foot is output by the physical information neural network. The contact force threshold, For left and right foot contact masks.

[0010] Preferably, the impact peak weighted loss is: ; ; in, The weighted loss for the peak impact. Let be the weight at the t-th sampling time. The normal ground reaction force of the foot is the output of the physical information neural network at the t-th sampling time. Let be the true value of the normal ground reaction force of the foot at the t-th sampling time. For peak weighting coefficients, For the Sigmoid function, The slope coefficient, This is the threshold for the rate of change of impact.

[0011] Preferably, the impulse consistency loss is: ; in, This refers to the impulse consistency loss. The normal ground reaction force of the foot is the output of the physical information neural network at the t-th sampling time. As a supporting phase time window, The sampling period is Let be the true value of the normal ground reaction force of the foot at the t-th sampling time.

[0012] Preferably, the total reward is: ; ; in, The total reward is as stated above. As a basic task reward, To assist in regularization rewards, The low-impact reward corresponding to the predicted value. In order to impact the penalty weight, The predicted value represents the normal ground reaction force of the left foot. The predicted value represents the normal ground reaction force of the right foot.

[0013] Preferably, the method for setting the impact penalty weight includes: setting a basic impact penalty weight, and then correcting the basic impact penalty weight online based on the robot's current impact intensity, velocity tracking error, and attitude stability error to obtain the final impact penalty weight; the basic impact penalty weight and the final impact penalty weight are respectively: ; ; ; in, Based on the impact penalty weight, For the final impact penalty weight, The minimum impact penalty weight, To maximize the impact penalty weight, For growth rate, To the number of training iterations, To mitigate the impact of rapidly increasing penalty weights in the center iteration count, For the amplitude limiting function, The first adjustment coefficient, For the Sigmoid function, To normalize the impact intensity, The desired impact strength threshold, This is the second adjustment coefficient. For speed tracking error, This is the third adjustment coefficient. For attitude stability error, For the t-th sampling time , For the t-th sampling time , For robot quality, This is the acceleration due to gravity.

[0014] Preferably, the sample data includes: samples consisting of the robot's foot normal ground reaction force, joint position, joint velocity, and joint acceleration under different terrains, and samples consisting of the robot's foot normal ground reaction force, joint position, joint velocity, and joint acceleration when wearing different footwear.

[0015] A second aspect of the present invention provides a robot gait control system, comprising: a first construction module for constructing a model for the robot that satisfies physical consistency constraints, representing the mapping relationship between plantar normal ground reaction force, joint position, joint velocity, and joint acceleration; a second construction module for constructing a physical information neural network that satisfies inverse dynamics constraints based on the model, wherein the input of the physical information neural network is joint position, joint velocity, and joint acceleration, and the output is plantar normal ground reaction force; and an optimization training module for optimizing the training by incorporating supervised regression error, inverse dynamics residual, oscillation phase zero force constraint, temporal smoothing constraint, contact phase consistency loss, and other parameters. The joint loss function of impact peak weighted loss and impulse consistency loss converges to the minimum value as the objective. The physical information neural network is optimized and trained using pre-collected sample data. The prediction module is used to: output initial control commands according to a preset control strategy when the robot is running; obtain the predicted value of the robot's foot normal ground reaction force using the trained physical information neural network, and add the predicted value to the total reward; the correction control module is used to correct the initial control commands using the total reward to obtain the optimal control commands to control the robot; the robot gait control system is used to implement the robot gait control method described above.

[0016] A third aspect of the present invention provides a robot including a controller for performing the robot gait control method described above.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention provides a robot gait control method, system and robot, constructs a physical information neural network that satisfies inverse dynamic constraints, and combines data-driven prediction with robot dynamic balance constraints through the physical information neural network, avoiding non-physical predictions generated by pure supervised networks under contact switching, swing phase and cross-foot conditions; the optimized and trained physical information neural network is used as a reinforcement learning reward evaluator, which can provide stable, continuous and deployable impact force feedback for robot control strategy, thereby reducing foot contact impact without sacrificing basic walking ability, and realizing low impact, low noise and stable walking of humanoid robots. Attached Figure Description

[0018] 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 recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart of a robot gait control method provided in an embodiment of the present invention.

[0020] Figure 2 This is an overall framework diagram of the robot gait control method provided in an embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram of the physical information neural network provided in an embodiment of the present invention.

[0022] Figure 4 This is a block diagram of a robot gait control system provided in an embodiment of the present invention. Detailed Implementation

[0023] In view of the shortcomings of the prior art, the inventors of this invention, through long-term research and extensive practice, have proposed the technical solution of this invention. The following will further explain and illustrate this technical solution, its implementation process, and its principles.

[0024] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0025] Furthermore, in the description of this invention, it should be understood that the terms "upper," "lower," "inner," "outer," "horizontal," "vertical," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0026] In the description of this specification, the references to terms such as "an embodiment," "a particular embodiment," or "the embodiment" indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0027] Figure 1 A flowchart illustrating a robot gait control method provided in an embodiment of the present invention. (See attached document.) Figure 1 , combined Figures 2-3 The robot gait control method provided in this embodiment of the invention will be described in detail. The method includes a network construction and training phase (operations S1-S3) and a control phase (operations S4-S5), and its overall framework is as follows: Figure 2 As shown.

[0028] Operation S1 constructs a model for the robot that satisfies physical consistency constraints, representing the mapping relationship between the foot normal ground reaction force, joint position, joint velocity, and joint acceleration.

[0029] Operation S2 constructs a physical information neural network that satisfies the inverse dynamics constraint based on the model. The input of the physical information neural network is the joint position, joint velocity, and joint acceleration, and the output is the normal ground reaction force of the foot.

[0030] Operation S3 aims to minimize the joint loss function, which includes supervised regression error, inverse dynamic residual, oscillating phase zero force constraint, temporal smoothing constraint, contact phase consistency loss, impact peak weighted loss, and impulse consistency loss, and optimizes the physical information neural network using pre-collected sample data.

[0031] Operation S4: When the robot is running: output initial control commands according to the preset control strategy; use the trained physical information neural network to obtain the predicted value of the robot's foot normal ground reaction force, and add the predicted value to the total reward.

[0032] Operation S5 uses the total reward to correct the initial control command and obtain the optimal control command to control the robot.

[0033] The robot gait control method provided in this invention addresses the problem that existing low-noise gait control relies on kinematic surrogate indices or vulnerable force sensors. It designs a physical information neural network that satisfies inverse dynamic constraints as a predictor of the plantar normal ground reaction force. This predictor is frozen and embedded with reinforcement learning training rewards, enabling the robot to learn a low-impact, low-noise, and footwear-robust walking strategy. This achieves low-impact, low-noise, and stable walking for humanoid robots, and features no need for plantar force sensors during the deployment phase, physical consistency in contact force estimation, stable low-impact rewards, and strong footwear adaptability.

[0034] Before optimizing the training of the physical information neural network, it is necessary to collect body perception data (including joint position, joint velocity, and joint acceleration) during the robot's walking process and foot normal ground reaction force data during the training phase to form sample data. The specific sample data collection process is as follows.

[0035] Collect body perception data (joint positions) of the robot during its movements such as walking, lateral movement, backward movement, and turning in place. Joint velocity Joint acceleration During the training phase, the normal ground reaction force of the foot can be obtained through foot pressure insoles or force sensors. ,in, , These are the reaction forces of the left and right feet to the ground, respectively.

[0036] To capture short-term transients near contact switching, the most recent H frames of body perception data are stitched together to form an input history window. For example, H can be 6 frames with a sampling frequency of 50Hz, corresponding to approximately 0.12s of historical information.

[0037] The input features can be represented as: ; in, For the first Network input at each control moment, , This represents the number of degrees of freedom of the robot's joints.

[0038] The output is the predicted values ​​of the normal ground reaction forces of the left and right feet. : ; In a preferred embodiment, the sample data includes: samples of the robot's foot normal ground reaction force, joint position, joint velocity, and joint acceleration under different terrains, and samples of the robot's foot normal ground reaction force, joint position, joint velocity, and joint acceleration under different footwear.

[0039] In this embodiment of the invention, the terrain randomization difficulty is gradually increased during training, and multiple footwear switching events are introduced, allowing the robot to gradually transition from barefoot conditions to different foot contact conditions such as skateboard shoes, sneakers, and high heels. Furthermore, to reduce the impact of simulation asset differences on training, an integrated foot and shoe model can be constructed. Based on the robot's foot model, the footwear mesh and foot links are Boolean merged or rigidly bound, adjusting the sole thickness, heel height, toe curvature, and contact surface area. The mass and inertia are then recalculated under a unified density assumption.

[0040] In a preferred embodiment, operation S1 specifically includes the following sub-operations S11-S12.

[0041] In suboperation S11, the robot's inverse dynamics equations are transformed using the damped least squares regularized pseudo-inverse method to obtain the first model describing the relationship between the original reaction force of the foot to the ground and the joint position, joint velocity, and joint acceleration.

[0042] In sub-operation S12, the original reaction force of the foot to the ground in the first model is projected onto a non-negative orthogonal domain to obtain a model that satisfies the physical consistency constraint and represents the mapping relationship between the reaction force of the foot to the ground, joint position, joint velocity and joint acceleration.

[0043] Under normal contact conditions, the robot's joint space dynamics satisfy the following relationship (i.e., the robot's inverse dynamics equation): ; in, The inertia matrix, For inertial force, For gravity, For equivalent generalized driving force, The Jacobian matrix for normal contact of the left and right feet. The reaction forces of the left and right feet on the ground are to be predicted.

[0044] Based on the above formula, we can obtain the generalized force caused by contact. for: ; Under single-leg support, two-leg switching, or near-singular configurations, directly solving for the contact force may result in numerical instability. Therefore, this invention employs a damped least-squares regularized pseudo-inverse to obtain an estimate of the original contact force (the original reaction force from the foot to the ground): ; in, The original reaction force of the foot on the ground. The Jacobian matrix for normal contact of the left and right feet. for The transpose of the matrix, The damping coefficient is... It is the identity matrix. To solve the robot's inverse dynamics equations for the normal ground reaction force of the foot, The damping coefficient is used to improve numerical stability under pathological contact Jacobian conditions.

[0045] Furthermore, to satisfy the unilateral contact constraint, the original reaction force of the foot method to the ground is projected onto a non-negative orthogonal domain: ; Thus, the predicted result of the physically non-negative normal ground reaction force of the foot was obtained.

[0046] In a preferred embodiment, the physical information neural network includes a rigid body dynamics structure branch, a temporal generalized driving force branch, and a fusion output module. The rigid body dynamics structure branch calculates the inertia matrix and the gravity term due to potential energy based on joint positions, and calculates the inertial force term based on joint positions and joint velocities. The temporal generalized driving force branch calculates the equivalent generalized driving force based on joint positions, joint velocities, and joint accelerations within a historical window. The fusion output module calculates and outputs the plantar normal ground reaction force based on the inertia matrix, gravity term, inertial force term, and equivalent generalized driving force.

[0047] See Figure 3 This illustrates the rigid body dynamics structure branch (rigid body network) and the temporal generalized driving force branch (temporal network) of a physical information neural network. The rigid body dynamics structure branch is based on joint positions. Estimating the inertia matrix Potential energy and the gravitational term caused by potential energy. and according to , and Calculate the Coriolis / eccentric term (i.e., the inertial force term). The time-series generalized driving force branch is used by the history window. Estimating equivalent generalized driving force To absorb joint friction, actuator dynamics, and unmodeled non-contact terms.

[0048] To ensure the physical rationality of the inertial matrix, It can be constructed as a symmetric positive definite matrix using decomposition, for example: ; in, This is the lower triangular matrix output by the physical information neural network. It is a positive number. It is an identity matrix. The potential energy branch outputs a scalar. Obtained by automatic differentiation or analytical differentiation .

[0049] In another preferred embodiment, to improve the prediction accuracy of the plantar reaction force, especially the prediction stability during the contact switching phase, across shoe conditions, and under different ground conditions, the robot's inertial state, foot kinematic state, contact phase state, and shoe-ground contact parameters can be further incorporated into the mapping model. The new mapping model can be expressed as: ; in, It indicates the robot's inertial state, including its posture, angular velocity, and linear acceleration; It indicates the kinematic state of the foot, including foot height, foot velocity, foot acceleration, and foot position relative to the fuselage. This is a contact phase state or contact phase estimate, which can be considered as a contact phase tag. This is a vector of footwear / ground contact parameters, which may include sole thickness, heel height, sole equivalent stiffness, equivalent damping, contact area, ground type, or ground equivalent stiffness, etc.

[0050] Accordingly, the physical information neural network can add contact parameter encoding and contact phase encoding branches to the existing rigid body dynamics structure branch and temporal generalized driving force branch. The contact parameter encoding branch encodes the footwear / ground contact parameters as latent variables, while the contact phase encoding branch identifies or represents the support phase, swing phase, and dual-foot support switching phase. The fusion output module then combines the dynamics structure branch, temporal generalized driving force branch, contact parameter encoding branch, and contact phase encoding branch to output the predicted values ​​of the normal ground reaction forces of the left and right feet.

[0051] In this preferred embodiment, the predictor can not only rely on joint motion state, but also sense the fuselage impact response, foot contact motion state, and contact dynamics changes caused by footwear / ground, thereby improving the accuracy and generalization ability of predicting plantar normal ground reaction force under contact switching, high impact ground contact, cross-footwear and cross-ground conditions.

[0052] When training a physical information neural network, a joint loss function that includes supervised regression and physical consistency is used. : ; in, , , , , , , These are supervised regression error, inverse dynamic residual, oscillating phase zero force constraint, time series smoothing constraint, contact phase consistency loss, impact peak weighted loss, and impulse consistency loss, respectively. , , , , , , These are the corresponding weighting coefficients. This is used to constrain the predicted normal force to be consistent with the foot contact phase state, so as to avoid the network outputting a large normal force in the swing phase or an abnormally small normal force in the support phase. This is used to assign higher weights to the moment of contact with the ground and the rapid rise of the normal force, so that the network focuses on learning the peak of the contact impact. This is used to ensure that the predicted normal force integral within a support phase period remains consistent with the actual or dynamically estimated impulse.

[0053] Preferred, , , , , , , Specifically, it can be expressed as: ; ; ; ; ; ; ; in, For zero-force constraint of the oscillating phase, For left and right foot contact masks, This represents element-wise multiplication. for The normalized value, The normal ground reaction force of the foot is output by the physical information neural network. The value of the zero normal force after normalization. This represents the square of the second norm of a vector. for The normalized value; The training label is obtained from the plantar force sensor, i.e., the true value of the plantar normal ground reaction force; the symbol "~" indicates the force value normalized using the mean and standard deviation of the training set. , This is the length of the history window. For the Sigmoid function, The slope coefficient, The contact force threshold. The weight at the t-th sampling time. Adaptively determined based on the rate of change of normal force or ground contact event: ;in, For peak weighting coefficients, The slope coefficient, This is the threshold for the rate of change of impact. This makes the network prioritize peak prediction error when a ground impact occurs, rather than just optimizing the overall average error. As a supporting phase time window, The sampling period. This is used to ensure that the overall impulse of the predicted normal force within a support phase is consistent with the actual impulse or the dynamically estimated impulse, thereby improving the prediction stability on the gait period scale.

[0054] Through this joint loss function, the physical information neural network not only fits the training data but is also constrained to satisfy dynamic equilibrium, zero contact force in the oscillating phase, and smooth contact switching. The contact phase consistency loss improves the ability to identify the boundary between the support and oscillating phases; the impact peak weighted loss improves the accuracy of peak force prediction at ground contact; and the impulse consistency loss improves the overall physical consistency of the predicted force within a gait cycle. The combination of these three factors further improves the training performance and deployment reliability of the physical information neural network.

[0055] After training, the network parameters of the physical information neural network are frozen and used as the impact reward evaluator in reinforcement learning training. A preset control strategy (e.g., a low-impact walking strategy πθ) receives proprioceptive historical information and outputs the joint target position increment in each control cycle. The controller generates joint torque commands through PD tracking: ; ; in, and These represent the current joint position and the current joint velocity, respectively. , These are the stiffness matrix and the damping matrix, respectively.

[0056] In a preferred embodiment, the total reward is: ; ; in, For the total reward, As a basic task reward, To assist in regularization rewards, The low-impact reward corresponding to the predicted value. In order to impact the penalty weight, The predicted normal ground reaction force of the left foot. This represents the right foot's ground reaction force in the predicted value. The basic task rewards include speed tracking, posture stability, and fall penalty; the auxiliary regularization rewards include energy consumption, motion smoothness, and joint limitation. By adjusting... By penalizing the square term of the predicted normal force, the strategy can proactively reduce the transient impact at ground contact while preserving the goals of velocity tracking and gait stability.

[0057] In a preferred embodiment, the method for setting the impact penalty weight includes: setting a basic impact penalty weight, and then correcting the basic impact penalty weight online based on the robot's current impact intensity, speed tracking error, and attitude stability error to obtain the final impact penalty weight.

[0058] Preferably, the basic impact penalty weight and the final impact penalty weight are as follows: ; ; ; in, Based on the impact penalty weight, For the final impact penalty weight, The minimum impact penalty weight, To maximize the impact penalty weight, For growth rate, To the number of training iterations, To mitigate the impact of rapidly increasing penalty weights in the center iteration count, For the amplitude limiting function, The first adjustment coefficient, For the Sigmoid function, To normalize the impact intensity, The desired impact strength threshold, This is the second adjustment coefficient. For speed tracking error, This is the third adjustment coefficient. For attitude stability error, For the t-th sampling time , For the t-th sampling time , For robot quality, This is the acceleration due to gravity.

[0059] The above-described method of setting the basic impact penalty weights allows the strategy to prioritize learning stable walking in the early stages of training, and gradually enhance low-impact constraints after acquiring basic gait capabilities. The meaning of the final impact penalty weight setting method described above is: when predicting impact intensity... When the impact exceeds a reference threshold, the impact penalty weight is appropriately increased to strengthen the strategy's suppression of foot impact. When the velocity tracking error or posture error is large, the impact penalty weight is appropriately reduced to avoid the strategy sacrificing basic walking stability and velocity tracking ability, or exhibiting degenerate behaviors such as dragging or extremely slow walking, in order to reduce impact. This method can achieve an adaptive balance between stability, velocity tracking, and low impact, thereby improving the stability of reinforcement learning strategy training and the final gait control effect.

[0060] This method eliminates the need for foot force sensors during the actual deployment phase. The deployment control system only uses body sensors such as joint encoders and inertial measurement units (IMUs) to acquire data. , , The system uses a historical window to estimate the ground reaction forces of the left and right feet based on frozen physical information neural networks, and outputs the target joint position increments based on the trained control strategy. For scenarios requiring online monitoring, the predicted ground reaction forces can also serve as auxiliary signals for impact alarms, foot protection, or gait mode switching.

[0061] The robot gait control method provided in this invention has the following advantages: First, it eliminates the need for foot force sensors, force plates, or external force measuring devices during deployment, reducing hardware costs and maintenance complexity. Second, it improves the reliability of contact force estimation by limiting network output through inverse dynamic residuals, non-negative projection of normal force, and zero-force constraints in the swing phase. Third, the frozen physical information neural network provides low-noise, differentiable, or computable reward signals for reinforcement learning, avoiding training instability caused by directly using noisy force sensor signals. Fourth, the reward function directly penalizes the predicted contact force amplitude or its squared term, enabling the strategy to proactively reduce transient impact upon ground contact. Fifth, through integrated foot and shoe modeling, multi-shoe switching, and curriculum training, the strategy adapts to different contact conditions such as barefoot, athletic shoes, skate shoes, and high heels. Sixth, this method can be integrated with various existing robot model formats and mainstream reinforcement learning training frameworks, facilitating deployment from simulation to real-world operation.

[0062] Based on the same inventive concept, and corresponding to any of the methods described above, this invention also provides a robot gait control system. (See also...) Figure 4 The robot gait control system 400 includes a first building module 410, a second building module 420, an optimization training module 430, a prediction module 440, and a correction control module 450.

[0063] The first building module 410, for example, performs operation S1 to build a model for the robot that satisfies physical consistency constraints and represents the mapping relationship between the foot normal ground reaction force, joint position, joint velocity and joint acceleration.

[0064] The second building module 420, for example, performs operation S2 to build a physical information neural network that satisfies the inverse dynamics constraint based on the model. The input of the physical information neural network is joint position, joint velocity and joint acceleration, and the output is the normal ground reaction force of the foot.

[0065] The optimization training module 430, for example, performs operation S3 to optimize the physical information neural network using pre-collected sample data, with the objective of minimizing the joint loss function that includes supervised regression error, inverse dynamic residual, oscillating phase zero force constraint, temporal smoothing constraint, contact phase consistency loss, impact peak weighted loss and impulse consistency loss.

[0066] For example, the prediction module 440 performs operation S4, which is used to: output initial control commands according to the preset control strategy when the robot is running; obtain the predicted value of the robot's foot normal ground reaction force using the trained physical information neural network, and add the predicted value to the total reward.

[0067] The correction control module 450, for example, performs operation S5 to correct the initial control command using the total reward, thereby obtaining the optimal control command to control the robot.

[0068] The robot gait control system 400 is used to implement the above-mentioned robot gait control method.

[0069] Based on the same inventive concept, corresponding to any of the above embodiments, the present invention also provides a robot, including a controller, which is used to execute the above-described robot gait control method.

[0070] It should be understood that the above embodiments are merely illustrative of the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A robot gait control method, characterized in that, This includes the network construction and training phases S1-S3 and the control phases S4-S5; S1 is a model for the robot that satisfies physical consistency constraints and represents the mapping relationship between the foot normal ground reaction force, joint position, joint velocity and joint acceleration. S2, Construct a physical information neural network that satisfies the inverse dynamics constraint based on the model. The input of the physical information neural network is the joint position, joint velocity, and joint acceleration, and the output is the normal ground reaction force of the foot. S3, with the goal of minimizing the joint loss function including supervised regression error, inverse dynamic residual, oscillating phase zero force constraint, temporal smoothing constraint, contact phase consistency loss, impact peak weighted loss and impulse consistency loss, the physical information neural network is optimized and trained using pre-collected sample data; S4, During robot operation: Output initial control commands according to the preset control strategy; Use the trained physical information neural network to obtain the predicted value of the robot's foot normal ground reaction force and add the predicted value to the total reward; S5, the initial control command is corrected using the total reward to obtain the optimal control command to control the robot.

2. The robot gait control method according to claim 1, characterized in that, S1 specifically includes: The robot's inverse dynamics equations are transformed using the damped least squares regularized pseudo-inverse method to obtain a first model describing the relationship between the original reaction force of the foot to the ground and the joint position, joint velocity, and joint acceleration. By projecting the original plantar normal reaction force to the ground in the first model onto a non-negative orthogonal domain, a model is obtained that satisfies the physical consistency constraint and represents the mapping relationship between the plantar normal reaction force to the ground, joint position, joint velocity, and joint acceleration.

3. The robot gait control method according to claim 1, characterized in that, The loss of contact phase consistency is: ； in, For the contact phase consistency loss, For the Sigmoid function, The slope coefficient, The normal ground reaction force of the foot is output by the physical information neural network. The contact force threshold, For left and right foot contact masks.

4. The robot gait control method according to claim 1, characterized in that, The peak impact weighted loss is: ； ； in, The weighted loss for the peak impact. Let be the weight at the t-th sampling time. The normal ground reaction force of the foot is the output of the physical information neural network at the t-th sampling time. Let be the true value of the normal ground reaction force of the foot at the t-th sampling time. For peak weighting coefficients, For the Sigmoid function, The slope coefficient, This is the threshold for the rate of change of impact.

5. The robot gait control method according to claim 1, characterized in that, The impulse consistency loss is: ； in, This refers to the impulse consistency loss. The normal ground reaction force of the foot is the output of the physical information neural network at the t-th sampling time. As a supporting phase time window, The sampling period is Let be the true value of the normal ground reaction force of the foot at the t-th sampling time.

6. The robot gait control method according to claim 1, characterized in that, The total reward is: ； ； in, The total reward is as stated above. As a basic task reward, To assist in regularization rewards, The low-impact reward corresponding to the predicted value. In order to impact the penalty weight, The predicted value represents the normal ground reaction force of the left foot. The predicted value represents the normal ground reaction force of the right foot.

7. The robot gait control method according to claim 6, characterized in that, The method for setting the impact penalty weight includes: setting a basic impact penalty weight, and then correcting the basic impact penalty weight online based on the robot's current impact intensity, speed tracking error, and attitude stability error to obtain the final impact penalty weight. The base impact penalty weight and the final impact penalty weight are as follows: ； ； ； in, Based on the impact penalty weight, For the final impact penalty weight, The minimum impact penalty weight, To maximize the impact penalty weight, For growth rate, To the number of training iterations, To mitigate the impact of rapidly increasing penalty weights in the center iteration count, For the amplitude limiting function, The first adjustment coefficient, For the Sigmoid function, To normalize the impact intensity, The desired impact strength threshold, This is the second adjustment coefficient. For speed tracking error, This is the third adjustment coefficient. For attitude stability error, For the t-th sampling time , For the t-th sampling time , For robot quality, This is the acceleration due to gravity.

8. The robot gait control method according to claim 1, characterized in that, The sample data includes: samples of the robot's foot normal ground reaction force, joint position, joint velocity, and joint acceleration under different terrains, and samples of the robot's foot normal ground reaction force, joint position, joint velocity, and joint acceleration under different footwear.

9. A robot gait control system, characterized in that, include: The first building module is used to build a model for the robot that satisfies physical consistency constraints and represents the mapping relationship between the plantar normal ground reaction force, joint position, joint velocity and joint acceleration. The second construction module is used to construct a physical information neural network that satisfies the inverse dynamics constraint based on the model. The input of the physical information neural network is joint position, joint velocity and joint acceleration, and the output is the normal ground reaction force of the foot. The optimization training module is used to optimize the training of the physical information neural network using pre-collected sample data, with the goal of minimizing the joint loss function, which includes supervised regression error, inverse dynamic residual, oscillating phase zero force constraint, temporal smoothing constraint, contact phase consistency loss, impact peak weighted loss and impulse consistency loss. The prediction module is used during robot operation to: output initial control commands according to a preset control strategy; obtain the predicted value of the robot's foot reaction force to the ground using a trained physical information neural network, and add the predicted value to the total reward; A calibration control module is used to calibrate the initial control command using the total reward to obtain the optimal control command to control the robot; The robot gait control system is used to implement the robot gait control method as described in any one of claims 1-8.

10. A robot, characterized in that, Includes a controller for performing the robot gait control method as described in any one of claims 1-8.

Images