Coordinated control method for impact suppression and walking motion of quadruped robot

Abstract

The present application relates to the technical field of robot motion control, in particular to a four-legged robot impact suppression and walking motion cooperative control method, the method comprising constructing a robot reinforcement learning motion controller, outputting each joint motion control torque; constructing an impact suppression controller, calculating the generalized control force required by the robot body, and then mapping to the joint space using the robot Jacobian matrix to obtain each joint impact compensation torque; weighting and fusing each joint motion control torque and each joint impact compensation torque to generate the final control torque of each joint of the robot; dynamically adjusting the weight coefficients of each joint motion control torque and each joint impact compensation torque to realize the cooperative control of the robot, which realizes the cooperative optimization of external impact suppression and walking motion control, effectively solves the posture stability problem of the robot in the dynamic running impact working condition, and improves the load action accuracy and platform survivability.

Description

Technical Field

[0001] This invention relates to the field of robot motion control technology, and in particular to a method for coordinated control of impact suppression and walking motion in quadruped robots. Background Technology

[0002] Quadruped robots possess excellent capabilities in navigating complex terrain, maneuverability, and environmental adaptability. They can perform tasks such as reconnaissance, patrol, target search, and material transport in environments where traditional wheeled and tracked platforms are difficult to traverse, and have become an important development direction for intelligent unmanned combat equipment in recent years. With the continuous development of intelligent sensing, autonomous decision-making, and motion control technologies, integrating launch payloads into quadruped robot platforms to achieve integrated reconnaissance, maneuverability, and strike capabilities has broad application prospects. The external impact load referred to in this invention refers to the transient impact force acting on the robot's body, which can originate from launch payload recoil, demolition impact, forging reaction, etc. This paper uses launch payload recoil as a typical scenario for explanation.

[0003] However, the transient recoil generated during payload launch imposes a significant impact load on the quadruped robot's body, causing pitch and roll disturbances, leading to decreased gait stability, trajectory deviation, and in severe cases, even robot instability or tipping over. Furthermore, during launch while the robot is moving dynamically, the recoil impact couples with the dynamic inertia generated by the robot's own motion, making attitude control more complex and placing higher demands on the real-time performance, stability, and coordination of the robot's motion control system. Therefore, effectively suppressing the impact of launch recoil on robot motion stability has become a key technical problem in quadruped robot applications.

[0004] Currently, control methods for external impact disturbances in quadruped robots mainly fall into two categories: mechanical buffering and active control. Mechanical buffering typically absorbs some of the impact energy by using springs, dampers, or flexible connection mechanisms, which can reduce the peak recoil force. However, this increases the platform weight and structural complexity, and the buffering effect is limited by mechanical structural parameters, making it difficult to simultaneously meet the requirements of lightweight design and high mobility. Active control, on the other hand, utilizes the robot's own joint driving capabilities to output compensating torque through methods such as proportional-derivative (PD) control, impedance control, or model predictive control (MPC) to counteract the impact of external impacts on the robot's posture. It has advantages such as fast response speed and no need to add additional hardware, but it still suffers from problems such as strong model dependence, complex parameter tuning, and insufficient real-time performance under complex working conditions.

[0005] On the other hand, reinforcement learning has become the mainstream technology paradigm for quadruped robot motion control. Reinforcement learning strategies can autonomously learn stable gait and complex terrain movement patterns through interaction with the environment, exhibiting strong environmental adaptability and robustness, and have been widely applied to the motion control of quadruped robots. However, existing reinforcement learning motion control strategies are mainly trained for normal walking conditions. Their state space typically includes information such as robot posture, body linear velocity, body angular velocity, joint states, and motion commands, lacking the perception and modeling of recoil impact information. This makes it difficult for the robot to actively generate anti-impact motion compensation actions at the moment of payload firing. Furthermore, quadruped robot recoil suppression control and walking motion control usually adopt independent or phased switching control methods. That is, reinforcement learning strategies are used for motion control during the normal movement phase, and PD control or impedance control is switched to recoil compensation during the payload firing phase. Due to the lack of a unified coordination mechanism between the two control strategies, abrupt changes in control commands can easily occur during the control mode switching process, causing robot posture oscillations, gait instability, and decreased motion continuity. Meanwhile, the reinforcement learning strategy does not fully consider the recoil impact condition and is difficult to effectively coordinate with recoil suppression control, which limits the application performance of the quadruped robot in dynamic moving shooting scenarios.

[0006] Therefore, there is an urgent need for a coordinated control method for impact suppression and walking motion of quadruped robots, so that the robot can sense recoil impact during dynamic movement and coordinate reinforcement learning motion control and recoil suppression control to improve the posture stability, motion continuity and firing accuracy of quadruped robots during dynamic shooting. Summary of the Invention

[0007] The purpose of this invention is to provide a method for coordinated control of impact suppression and walking motion in quadruped robots, aiming to solve the problem that the impact suppression control and walking motion control of quadruped robots are independent and lack coordination, resulting in abrupt changes in control commands under impact loads and the inability of reinforcement learning strategies to effectively cope with impact conditions.

[0008] To achieve the above objectives, this invention provides a method for coordinated control of impact suppression and walking motion in a quadruped robot, comprising the following steps: Construct a reinforcement learning motion controller for a quadruped robot, which outputs joint motion control torques based on the input state space; An impact suppression proportional-differential controller is constructed. Based on the error between the quadruped robot's body posture and the desired posture, the generalized control force required by the robot body is calculated. Then, the robot Jacobian matrix is ​​used to map to the joint space to obtain the impact compensation torque of each joint. The motion control torque of each joint and the impact compensation torque of each joint are weighted and fused according to a preset fusion strategy to generate the final control torque of each joint of the robot. Based on the characteristic information of external impact loads, the weighting coefficients of the motion control torque and the impact compensation torque of each joint are dynamically adjusted to achieve coordinated control of the robot in the static working stage, normal walking stage, impact preparation and impact suppression stage, and posture recovery stage.

[0009] In one embodiment, the input state space includes the robot body position, body posture, body linear velocity, body angular velocity, 12 joint angles, 12 joint angular velocities, speed commands, and external impact load characteristic information. The external impact load characteristic information includes the impact triggering state signal of the external impact load, the impact load timing, the impact force intensity value, and the impact force direction information.

[0010] In one embodiment, the reinforcement learning motion controller employs the Proximal Policy Optimization (PPO) algorithm and performs policy training based on a quadruped robot dynamics simulation environment to obtain a motion control policy suitable for continuous motion space. The motion space of the reinforcement learning motion controller adopts a continuous motion space structure. Each dimension of the motion vector corresponds to a drive joint of the quadruped robot. The reinforcement learning motion controller outputs the motion control joint torques for each joint. ; in, To enhance the joint motion control torque output by the learning motion controller, This indicates the action output by the policy network.

[0011] In one embodiment, the control law of the impact suppression proportional-derivative controller is: ; in, Desiring generalized force for the robot body, and These are the proportional gain matrix and the differential gain matrix, respectively. and These are the desired joint angle and the desired joint angular velocity, respectively. and These are the actual joint angle and the actual joint angular velocity, respectively. This is the joint damping torque vector output by the impact damping controller. It is a Jacobian matrix.

[0012] In one embodiment, the motion control torque of each joint and the impact compensation torque of each joint are weighted and fused according to a preset fusion strategy to generate the final control torque of each joint of the robot. The specific method of the weighted fusion is as follows: ; in, This is the final joint control torque vector, used to drive the motion control of each joint of the quadruped robot. To enhance the learning of the motion control torque vector output by the motion controller, reflecting the autonomous motion control actions generated by the robot in its current motion state, This is the impact suppression torque vector output by the proportional-derivative (PD) controller, used to compensate for impact disturbances caused by external impact loads. To enhance the learning of motor control weighting coefficients, for time... t The time-varying function is used to adjust the proportion of reinforcement learning control actions in the final control command. Here, represents the proportional-derivative impact suppression control weighting coefficient, and represents the time-dependent weighting coefficient. t The time-varying function is used to adjust the proportion of the impact suppression control quantity in the final control command; t To control the timing; and Satisfy the normalization constraint: , , .

[0013] The method for determining the reinforcement learning motion control weight coefficients and the proportional-derivative impact suppression control weight coefficients is as follows: When a quadruped robot is not walking, but is performing an external impact load task , The system is in a pure proportional-derivative impact compensation control mode; When the quadruped robot walks normally and no external impact load triggers an impact. , The system is in a pure reinforcement learning motion control mode; During the impact suppression preparation stage before or at the moment of impact, when an external impact load occurs. Within a preset time window, the value decreases from 1 to a preset lower limit. The control mode is smoothly switched by synchronously increasing from 0 to a preset upper limit value. Once the impact suppression is complete Gradually recover to 1, Gradually recovering to 0, the quadruped robot re-enters the reinforcement learning motion control mode.

[0014] Specifically, when the dynamic acceleration of the quadruped robot's body in the forward-backward direction or the dynamic acceleration at any leg's hip joint exceeds a preset threshold, the impact suppression proportional-differential controller is triggered to resist the impact force.

[0015] This invention discloses a cooperative control method for impact suppression and walking motion of a quadruped robot. The method constructs a reinforcement learning motion controller for the quadruped robot, outputting joint motion control torques based on the input state space. It also constructs an impact suppression proportional-derivative controller, calculating the required generalized control force based on the error between the quadruped robot's body posture and the desired posture, and then mapping this force to the joint space using the robot's Jacobian matrix to obtain the impact compensation torques for each joint. The joint motion control torques and impact compensation torques are weighted and fused according to a preset fusion strategy to generate the final control torques for each joint. Based on external impact load characteristic information, the weight coefficients of the joint motion control torques and impact compensation torques are dynamically adjusted to achieve cooperative control of the robot during the stationary working phase, normal walking phase, impact preparation and suppression phase, and posture recovery phase. This method constructs a cooperative control architecture coupling reinforcement learning motion control strategy and proportional-derivative impact suppression control. External impact load characteristic information is introduced into the state space of the reinforcement learning strategy as a priori input. Finally, it utilizes an inertial measurement unit (IMU)... The system fuses multi-source information from an IMU (Integrated Mutual Actuation Unit) and a 3D accelerometer to detect and calculate the impact trigger state signal, impact load timing, impact force intensity, and impact force direction of the external impact load. Then, based on the characteristics of the external impact load, it dynamically adjusts the fusion weight coefficients of the reinforcement learning motion control torque and the PD impact compensation torque to achieve smooth switching of the control mode. This enables the coordinated optimization of external impact force suppression and walking motion control, effectively solving the posture stability problem of quadruped robots under dynamic walking impact conditions, and improving the load application accuracy and platform survivability.

[0016] Compared with existing technologies, this method has the following advantages: Improved control coordination: By introducing external impact load feature information into the state space of the reinforcement learning motion controller, and adopting a control architecture that weights and fuses the joint motion control torque output by reinforcement learning with the PD impact suppression torque, the coordinated optimization of impact suppression control and walking motion control is achieved, overcoming the shortcomings of the two controls being independent and lacking coordination in the existing technology.

[0017] Smooth switching of control modes: By designing time-varying weight coefficients, a smooth transition of control modes is achieved, avoiding abrupt changes in control commands and attitude oscillations caused by traditional strategy switching methods, and improving the robot's attitude stability at the moment of impact of external impact loads.

[0018] Impact Adaptability of Reinforcement Learning Strategies: By introducing external impact load feature information as prior information into the reinforcement learning state space, the policy network can perceive the impact condition during the training phase and learn to generate action strategies that are coordinated with impact suppression, thus solving the impact adaptability problem in this scenario.

[0019] Reduced hardware dependence: No additional mechanical buffer devices are needed. Shock suppression and walking motion can be coordinated and controlled simply through innovative control algorithms, without increasing the platform's weight or structural complexity. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0021] Figure 1 This is a flowchart of the method for coordinated control of impact suppression and walking motion of a quadruped robot provided by the present invention.

[0022] Figure 2 This is a flowchart of the reinforcement learning motion control algorithm for quadruped robots provided by the present invention.

[0023] Figure 3 This is a flowchart of the steps of the method for coordinated control of impact suppression and walking motion of a quadruped robot provided by the present invention.

[0024] Figure 4 This is a schematic diagram of the coordinated control system for impact suppression and walking motion of a quadruped robot provided by the present invention.

[0025] In the diagram: 1-Reinforcement learning motion control module, 2-PD impact suppression module, 3-Perception module, 4-Collaborative control fusion module, 5-Switching control module. Detailed Implementation

[0026] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, but should not be construed as limiting the present invention.

[0027] The external impact load characteristic information described in this invention refers to the timing, intensity, and direction of the transient impact force acting on the robot body, which can be acquired through an IMU and a three-dimensional accelerometer. As a typical application example, this invention can be applied to a quadruped robot system integrating a launch payload. In this case, the external impact load is the recoil impact generated during the launch payload's firing process, and its triggering time can be obtained by testing the launch payload's firing state signal. The following specific implementation method uses the launch payload's recoil impact as a typical working condition. Those skilled in the art can extend the method described in this invention to other types of external impact load conditions.

[0028] To facilitate understanding by those skilled in the art, the correspondence between the general terms described in the following specific embodiments and the typical application example is as follows: The "external impact load" in this embodiment typically refers to the recoil impact generated during the firing process of the launch load; the "external impact load characteristic information" corresponds to the recoil characteristic information of the launch load, including the recoil timing, peak value, and direction of action; the "impact force" corresponds to the launch recoil; the "impact trigger" corresponds to the firing of the launch load; the "before impact trigger" corresponds to before the launch load is fired; the "impact trigger preparation" corresponds to launch preparation; the "not triggered" corresponds to not firing; the "static working phase" corresponds to the static launch phase; the "dynamic moving impact condition" corresponds to the dynamic moving firing condition; the "load action accuracy" corresponds to the launch load's strike accuracy; and the "integration of perception, maneuver, and operation" corresponds to the integration of reconnaissance, maneuver, and strike. The above correspondence is only used to illustrate the specific performance of the present invention in a particular application scenario and does not constitute a limitation on the scope of protection of the present invention.

[0029] Please see Figures 1 to 3 In a first aspect, the present invention provides a method for coordinated control of impact suppression and walking motion in a quadruped robot, comprising the following steps: S1 constructs a reinforcement learning motion controller for a quadruped robot, which outputs joint motion control torques based on the input state space. In this embodiment of the invention, the input state space includes the robot's body position, body posture, body linear velocity, body angular velocity, 12 joint angles, 12 joint angular velocities, velocity commands, and external impact load characteristic information. The external impact load characteristic information is detected and acquired by an inertial measurement unit installed on the robot's body and an accelerometer installed at the hip joint of the leg. The external impact load characteristic information includes the impact trigger state signal of the external impact load, the impact load timing, the impact force intensity value, and the impact force direction information. Specifically, a dynamic simulation model is first established based on the physical parameters of the quadruped robot, and a state space including body posture, body velocity, joint state, motion commands, and external impact load characteristic information is constructed. A core task objective is established with joint control torque as the motion space, motion command tracking, body posture stability, dynamic balance maintenance under impact loads, and control energy consumption optimization. At the same time, reward shaping mechanisms such as joint motion penalty, motion smoothing penalty, foot contact reward, and abnormal contact penalty are introduced to establish a reinforcement learning training environment. The PPO algorithm employs an Actor-Critic network structure (Actor network and Critic network). The Actor network outputs continuous control torques for each joint based on the current state, while the Critic network evaluates the value of the current state and calculates the advantage function to guide the policy network update. During training, the PPO algorithm improves the stability of the policy iteration process by limiting the update amplitude between the old and new policies, preventing excessive policy updates from causing training divergence. Furthermore, during reinforcement learning training, external impact load characteristics are introduced as prior experience into the reinforcement learning state space, enabling the policy network to learn the motion compensation laws under external impact conditions and generate motion control torques coordinated with PD impact compensation control. Through extensive interactive training in simulated environments, the reinforcement learning policy acquires autonomous motion control and impact resistance capabilities in complex dynamic environments. The trained policy is then directly deployed on a quadruped robot platform to achieve coordinated control of walking motion and impact compensation during dynamic walking impact conditions.

[0030] By introducing external impact load feature information into the reinforcement learning state space, the policy network can perceive the extreme working condition of impact load during the training phase and learn to generate motion control actions that are coordinated with impact suppression, thus fundamentally solving the problem of impact adaptability in this scenario.

[0031] The state representation of a reinforcement learning motion controller is as follows: ; in, Indicates the position of the robot's body. Indicates the robot's body posture. Indicates the linear velocity of the fuselage. Indicates the fuselage angular velocity. This represents 12 joint angles. This represents the angular velocity of 12 joints. Indicates speed command, This indicates the characteristic information of external impact loads.

[0032] The speed command can be written as: ; in, Indicates the desired forward speed. Indicates the desired lateral velocity. This represents the desired yaw rate.

[0033] The motion space of the reinforcement learning motion controller adopts a continuous motion space structure. Each dimension of the motion vector corresponds to a drive joint of the quadruped robot. The reinforcement learning motion controller outputs the motion control joint torques for each joint. ; in, To enhance the joint motion control torque output by the learning motion controller, This indicates the action output by the policy network.

[0034] The reward function for reinforcement learning uses a multinomial weighted summation, and the total reward function can be expressed as: ; in, For the first Weighting coefficients for each reward / penalty item; For the first A specific reward or penalty function.

[0035] Specifically, the reward function mainly consists of velocity tracking reward, angular velocity tracking reward, fuselage stability penalty, joint motion penalty, energy consumption penalty, motion smoothness penalty, foot contact reward and abnormal contact penalty, and external impact load characteristic information penalty. Based on the current project configuration, the total reward function can be written as: ; in, Represents the reward function, This represents a linear velocity tracking reward, used to encourage the robot to track the desired direction. , Directional velocity, This represents the angular velocity tracking reward, used to encourage the robot to track the desired yaw angular velocity. This indicates a vertical linear velocity penalty, used to suppress vertical bounce of the aircraft. This is represented as a roll and pitch angular velocity penalty, used to suppress fuselage attitude sway. This is represented as a joint angular velocity penalty. This is expressed as a joint angular acceleration penalty. This is represented as a joint torque penalty. This is represented as a penalty for the rate of change of action. This is to reduce the penalty for sudden changes in action between adjacent time steps, making the robot's motion smoother. This is represented as an energy consumption penalty. This is represented as an attitude stability penalty, used to constrain the fuselage to remain level. This is represented as a joint posture penalty, used to limit the joint angle from deviating too much from the default posture. This is represented as a reward for the time the foot is off the ground, used to encourage the robot to develop a reasonable gait. This is represented as the variance penalty for foot airtime, used to constrain the gait rhythm of all four legs. This is referred to as a foot slip penalty, used to reduce horizontal slippage of the supporting foot on the ground. This refers to the foot contact force penalty, used to prevent excessive impact on the foot. This refers to unintended contact penalties, used to prevent collisions between the fuselage, hips, thighs, calves, and other non-foot parts of the aircraft and the ground. This represents the penalty for the characteristic information of external impact loads. Right now to , which is the reward / punishment coefficient.

[0036] Please see Figure 2 This is the training and execution process of a reinforcement learning motion controller: Data Acquisition (Interaction Phase): Policy Network (Actor, parameters are...) ) Receive the current state Output action (i.e., the control torque of each joint movement).

[0037] After this action is applied to the robot's environment, the environment provides a reward. And move to the next state. At the same time, a termination flag is generated. Interactive data It is stored in the trajectory buffer.

[0038] Network update (learning phase): Batch data is sampled from the trajectory buffer and used to update the policy network (parameters). ) and value network (Critic, parameter The Proximal Policy Optimization (PPO) algorithm optimizes the policy towards higher rewards while limiting the update magnitude; it evaluates the merits of the current policy to guide the update direction of the policy network.

[0039] Iterative convergence: The above acquisition and update process is repeated until the policy network converges, obtaining a motion control policy applicable to the continuous action space, which can be directly deployed to the robot's output joint torque.

[0040] Specifically, the reinforcement learning motion controller employs a two-stage training method: Phase 1: Pre-training in the Simulation Environment. A reinforcement learning training environment is constructed based on the quadruped robot's dynamics simulation model. This environment includes robot body dynamics, leg dynamics, joint actuation models, and ground contact models, simulating different terrains, different movement speeds, and conditions such as single and continuous impacts. The reinforcement learning motion controller continuously interacts with the simulation environment using the PPO algorithm, obtaining an initial motion control strategy through iterative optimization. During training, external impact load characteristics are input as prior information into the policy network, enabling it to learn the motion compensation laws under external impact force conditions.

[0041] Phase Two: Physical Prototype Deployment and Strategy Verification. The pre-trained reinforcement learning strategy is deployed to the quadruped robot platform, and its control performance is verified and parameters are fine-tuned under actual walking and launching conditions. When the robot performs a launching mission, the impact suppression proportional-differential controller outputs an impact compensation torque based on the robot's posture error, and this torque is weighted and fused with the motion control torque output by the reinforcement learning strategy to achieve coordinated optimization of external impact force compensation and motion control. When the robot's posture deviation exceeds a preset safety threshold, the system increases the PD impact compensation control weight to enhance the robot's posture stability and ensure the safety and robustness of the control process.

[0042] S2 constructs an impact suppression proportional-differential controller. Based on the error between the quadruped robot's body posture and the desired posture, it calculates the generalized control force required by the robot body and then uses the robot Jacobian matrix to map to the joint space to obtain the impact compensation torque of each joint. In this embodiment of the invention, the desired state is defined as the robot body being horizontal and its position fixed, while the actual state is collected in real time by the onboard inertial measurement unit and joint encoder.

[0043] The control law of the shock suppression proportional-derivative controller is: ; in, Desiring generalized force for the robot body, and These are the proportional gain matrix and the differential gain matrix, respectively. and These are the desired joint angle and the desired joint angular velocity, respectively. and These are the actual joint angle and the actual joint angular velocity, respectively. This refers to the joint damping torque vector output by the impact damping controller. It is a Jacobian matrix.

[0044] S3 performs weighted fusion of the motion control torque of each joint and the impact compensation torque of each joint according to a preset fusion strategy to generate the final control torque of each joint of the robot. In this embodiment of the invention, the collaborative control fusion module reinforces the motion control torque vector output by the motion controller. Impact suppression torque vector output by the proportional-derivative controller for impact suppression Weighted fusion is performed to generate the final joint control torque vector: ; in, This is the final joint control torque vector, used to drive the motion control of each joint of the quadruped robot. To enhance the learning of the motion control torque vector output by the motion controller, reflecting the autonomous motion control actions generated by the robot in its current motion state, This is the impact suppression torque vector output by the proportional-differential controller for impact suppression, used to compensate for impact disturbances generated during launch. To enhance the learning of motor control weighting coefficients, for time... t The time-varying function is used to adjust the proportion of reinforcement learning control actions in the final control command; Here, represents the PD shock suppression control weighting coefficient, and represents the time-dependent weighting coefficient. t The time-varying function is used to adjust the proportion of the impact suppression control quantity in the final control command; t To control the timing; and Satisfy the normalization constraint: , , .in, The lower limit of the value is denoted as the preset lower limit value. , The upper limit of the value is denoted as the preset upper limit value. Both satisfy .

[0045] The method for determining its time-varying weighting coefficients is as follows: when the quadruped robot is not walking but performing an external impact load task, , The system is in pure PD impact compensation control mode; when the quadruped robot walks normally and the external impact load does not trigger an impact. , The system is in a pure reinforcement learning motion control mode; during the impact suppression preparation phase before the external impact load occurs or at the instant of the impact, ... Decrease from 1 to a preset lower limit value within a preset time window. , Synchronously increment from 0 to the preset upper limit value This enables smooth switching between control modes; once the impact suppression is complete... Gradually recover to 1, Gradually returning to 0, the quadruped robot re-enters the reinforcement learning motion control mode. The smooth transition of time-varying weight coefficients avoids the abrupt changes in control commands caused by traditional strategy switching methods.

[0046] Based on the characteristic information of external impact loads, S4 dynamically adjusts the weighting coefficients of the motion control torque of each joint and the impact compensation torque of each joint to achieve coordinated control of the robot in the static working stage, normal walking stage, impact preparation and impact suppression stage, and posture recovery stage. In this embodiment of the invention, the weighting coefficient is dynamically adjusted by the switching control module based on the characteristics of the external impact load. The specific workflow of the switching control module is as follows: (1) Static working phase: When the quadruped robot is not walking but performing a launch mission, the system enters the pure PD impact compensation control mode. , Ultimately, the joint control torque is entirely output by the shock suppression proportional-differential controller. The shock suppression proportional-differential controller obtains the shock compensation torque for each joint based on the error between the quadruped robot's body posture and the desired posture, in order to maintain the robot's posture stability during launch.

[0047] (2) Normal walking phase: When the quadruped robot is walking normally and no external impact load triggers an impact, the system is in a pure reinforcement learning motion control mode. , The final joint control torque is The reinforcement learning motion controller outputs motion control torques for each joint based on the robot's current motion state, enabling the robot to walk stably and track motion commands.

[0048] (3) Impact preparation and impact suppression phase: The sensing module detects the launch status based on the external impact load characteristic information and predicts the timing and characteristics of the impact load. Within a preset time window before the impact load occurs, the switching control module controls the launch. Gradually decrease from 1 to the preset lower limit value , Synchronously increment from 0 to the preset upper limit value , This enables a smooth transition of the system from a pure reinforcement learning motion control mode to a cooperative control mode. Upon triggering an external impact load, the impact suppression proportional-differential controller outputs joint impact compensation torque based on the robot's body and leg posture errors. Simultaneously, the reinforcement learning motion controller outputs motion control torque coordinated with the impact load. The two are then weighted and fused according to their respective weights to generate the final joint control torque. This effectively suppresses the impact load generated during launch, maintaining the robot's stable posture. During the impact suppression phase, the weighting coefficient remains at [value missing]. , .

[0049] (4) Attitude recovery phase: After the impact suppression is completed, the switching control module gradually restores the control weight coefficients to the original state based on the external impact load characteristics and the robot's attitude recovery status. , The system then re-enters the reinforcement learning motion control mode. The weight recovery process adaptively adjusts based on the robot's posture stability, enabling the robot to smoothly return to normal walking mode and avoiding posture oscillations caused by control mode switching.

[0050] (5) When the dynamic acceleration of the quadruped robot body in the front-back direction or the dynamic acceleration at the hip joint of any leg exceeds the preset threshold, the impact suppression proportional differential controller is triggered to resist the impact force.

[0051] In this embodiment of the invention, the robot's posture information is evaluated in real time based on the robot's body IMU and the three-dimensional acceleration of the hip joints of the legs. When the dynamic acceleration of the quadruped robot's body in the forward-backward direction or the dynamic acceleration at any leg hip joint exceeds a preset threshold, the impact suppression proportional-derivative controller is triggered to resist the impact force.

[0052] Please see Figure 4 Secondly, the present invention also provides a cooperative control system for impact suppression and walking motion of a quadruped robot, which is applied to the cooperative control method for impact suppression and walking motion of a quadruped robot as described in the first aspect above, including a reinforcement learning motion control module 1, a PD impact suppression module 2 (PD controller for suppressing recoil of launch payload), a sensing module 3, a cooperative control fusion module 4, and a switching control module 5. The reinforcement learning motion control module 1 (quadruped robot reinforcement learning motion controller) outputs motion control actions based on state space information; The PD impact suppression module 2 calculates the generalized control force required by the robot body based on the error between the quadruped robot body posture and the desired posture, and then uses the robot Jacobian matrix to map to the joint space to obtain the impact compensation torque of each joint. The sensing module 3 detects impact load events generated by external impact loads based on the IMU and a three-dimensional accelerometer, calculates and obtains the impact triggering state signal, impact load timing, impact force intensity value, and impact force direction information of the external impact load, and generates external impact load characteristic information. It should be noted that the IMU data also serves the motion control of the quadruped robot; this module is designed to accommodate the collaborative needs of both. The collaborative control fusion module 4 is used to weight and fuse the motion control torque of each joint and the impact compensation torque of each joint according to a preset fusion strategy to generate the final control torque of each joint of the robot. The switching control module 5 dynamically adjusts the weighting coefficients of the motion control torque and the impact compensation torque of each joint based on the external impact load characteristic information, so as to realize the coordinated control of the robot in the static working stage, normal walking stage, impact preparation and impact suppression stage, and posture recovery stage.

[0053] Furthermore, the reinforcement learning motion control module 1 is constructed using the PPO algorithm, including a policy network and a value network. The policy network is used to output the motion control torque of each joint, and the value network is used to evaluate the value of the current state and guide the policy update.

[0054] The above-disclosed embodiments are merely preferred embodiments of the collaborative control method for impact suppression and walking motion of the quadruped robot of this application, and should not be construed as limiting the scope of this application. Those skilled in the art can understand that all or part of the processes of the above embodiments can be implemented, and equivalent changes made in accordance with the claims of this application still fall within the scope of this application.

Claims

1. A method for coordinated control of impact suppression and walking motion in a quadruped robot, characterized in that, Includes the following steps: A reinforcement learning motion controller for a quadruped robot is constructed, which outputs motion control torques for each joint based on the input state space. The input state space includes the robot's body position, body posture, body linear velocity, body angular velocity, 12 joint angles, 12 joint angular velocities, velocity commands, and external impact load characteristic information. The external impact load characteristic information includes the impact trigger state signal of the external impact load, the impact load timing, the impact force intensity value, and the impact force direction information. An impact suppression proportional-differential controller is constructed. Based on the error between the quadruped robot's body posture and the desired posture, the generalized control force required by the robot body is calculated. Then, the robot Jacobian matrix is ​​used to map to the joint space to obtain the impact compensation torque of each joint. The motion control torque of each joint and the impact compensation torque of each joint are weighted and fused according to a preset fusion strategy to generate the final control torque of each joint of the robot. Based on the characteristic information of external impact loads, the weighting coefficients of the motion control torque and the impact compensation torque of each joint are dynamically adjusted to achieve coordinated control of the robot in the static working stage, normal walking stage, impact preparation and impact suppression stage, and posture recovery stage.

2. The method for coordinated control of impact suppression and walking motion in a quadruped robot as described in claim 1, characterized in that, The reinforcement learning motion controller employs a proximal policy optimization algorithm and performs policy training based on a quadruped robot dynamics simulation environment to obtain a motion control policy suitable for continuous motion space. The motion space of the reinforcement learning motion controller adopts a continuous motion space structure. Each dimension of the motion vector corresponds to a drive joint of the quadruped robot. The reinforcement learning motion controller outputs the motion control joint torques for each joint. ； in, To enhance the joint motion control torque output by the learning motion controller, This indicates the action output by the policy network.

3. The method for coordinated control of impact suppression and walking motion in a quadruped robot as described in claim 1, characterized in that, The motion control torques and impact compensation torques of each joint are weighted and fused according to a preset fusion strategy to generate the final control torques of each joint of the robot. The specific method of the weighted fusion is as follows: ; in, This is the final joint control torque vector, used to drive the motion control of each joint of the quadruped robot. To enhance the learning of the motion control torque vector output by the motion controller, reflecting the autonomous motion control actions generated by the robot in its current motion state, This is the impact suppression torque vector output by the proportional-derivative (PD) controller, used to compensate for impact disturbances caused by external impact loads. To enhance the learning of motor control weighting coefficients, for time... t The time-varying function is used to adjust the proportion of reinforcement learning control actions in the final control command. Here, represents the proportional-derivative impact suppression control weighting coefficient, and represents the time-dependent weighting coefficient. t The time-varying function is used to adjust the proportion of the impact suppression control quantity in the final control command; t To control the timing; and Satisfy the normalization constraints: , , .

4. The method for coordinated control of impact suppression and walking motion in a quadruped robot as described in claim 3, characterized in that, The method for determining the reinforcement learning motion control weight coefficients and the proportional-derivative impact suppression control weight coefficients is as follows: When a quadruped robot is not walking, but is performing an external impact load task , The system is in a pure proportional-derivative impact compensation control mode; When the quadruped robot walks normally and no external impact load triggers an impact. , The system is in a pure reinforcement learning motion control mode; During the impact suppression preparation stage before an external impact load occurs, or at the instant of the impact. Within a preset time window, the value decreases from 1 to a preset lower limit. The control mode is smoothly switched by synchronously increasing from 0 to a preset upper limit value. Once the impact suppression is complete Gradually recover to 1, Gradually recovering to 0, the quadruped robot re-enters the reinforcement learning motion control mode.

5. The method for coordinated control of impact suppression and walking motion in a quadruped robot as described in claim 1, characterized in that, Also includes: When the dynamic acceleration of the quadruped robot's body in the forward-backward direction or the dynamic acceleration at the hip joint of any leg exceeds a preset threshold, the impact suppression proportional-derivative controller is triggered to resist the impact force.

Images