Quadruped robot energy efficiency optimization method based on series elastic drivers
Through the driving-energy consumption model based on LuGre friction model and reinforcement learning optimization, combined with model prediction control, the problems of low energy efficiency and poor stability of four-legged robots are solved, and the energy efficiency and stability are improved, the design process is simplified and the design process is improved. Energy efficiency limit.
Patent Information
- Application Number
- CN202410813502.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-21
- Publication Date
- 2024-11-15
AI Technical Summary
The problems of low energy efficiency and poor stability of four-legged robots during movement, especially the series elastic drivers reduce the control bandwidth and increase the difficulty of control, and the existing energy consumption evaluation methods are far from the actual situation.
Based on the LuGre friction model, a driving-energy consumption model is established, combined with reinforcement learning and model prediction control, a collaborative energy efficiency optimization motion planning controller is designed, and the joint parameters and control parameters of the four-legged robot are optimized through offline parameter optimization and real-time controller. Improve energy efficiency and stability.
By analyzing the energy consumption model and dynamic model of the four-legged robot, using reinforcement learning to optimize the design parameters and control parameters, the energy efficiency and stability of the four-legged robot are improved, the redesign time caused by unreasonable design is reduced, and the energy efficiency upper limit is improved .
Smart Images

Figure CN118963111A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a motion planning and control method for a quadruped robot, and in particular to a quadruped robot using a series elastic drive. A method for optimizing the energy efficiency of joint parameters and control parameters of the quadruped robot based on reinforcement learning and model predictive control is designed. Technical Background
[0002] The common interaction between quadruped robots and the ground is mostly point contact. This contact mode improves the adaptability of quadruped robots to unstructured terrain by selecting a reasonable foothold position, but also increases the difficulty of maintaining dynamic stability during movement. Traditional control algorithms are difficult to cope with the high dynamic interaction requirements between quadruped robots and the environment. Therefore, optimization-based motion planning methods are applied to quadruped robot control, among which the most representative is model predictive optimization control; reinforcement learning has strong self-learning ability, which can reduce the control difficulty and improve the adaptability to the environment. Therefore, the combination of model predictive control and reinforcement learning is widely used for motion planning control of quadruped robots.
[0003] The series elastic drive is composed of a high torque output density motor, a high reduction ratio reducer and an elastic element. Compared with the quasi-direct drive, it has stronger impact resistance, higher torque control accuracy and stronger overload capacity. However, the series elastic drive reduces the bandwidth of the quadruped robot control and increases the difficulty of control. At the same time, the movement of the quadruped robot presents a periodic characteristic. The addition of elastic elements allows energy to be recycled, but most of the existing research evaluates energy consumption based on the square of the torque, which is far from the actual energy consumption. Therefore, how to improve the recycling rate of energy in the periodic movement of the quadruped robot, how to analyze the energy consumption model in the movement, and how to improve the stability of the movement have become research problems. Summary of the invention
[0004] In order to solve the above-mentioned technical problems existing in the prior art, the present invention targets the problems of low energy efficiency and poor smoothness during the movement of a quadruped robot. A series elastic collimator drive is applied to the quadruped robot, a drive-energy consumption model is established based on the LuGre friction model, and a collaborative energy efficiency optimization motion planning controller based on reinforcement learning and model predictive control is designed, including offline parameter collaborative energy efficiency optimization based on reinforcement learning and a real-time motion controller based on model predictive optimization control, thereby providing a feasible solution for improving the energy efficiency and smoothness of the movement of the quadruped robot.
[0005] To achieve the above object, the present invention provides the following technical solutions:
[0006] A method for optimizing energy efficiency of a quadruped robot based on a series elastic actuator comprises the following steps:
[0007] Step 1: Set the design parameter set of the elastic element of the series elastic actuator: D, which includes the joint stiffness k, the parameters a, b, c of the Archimedean spiral, determine the stiffness range according to the motor output torque, and use the motor size as the constraint condition of the Archimedean spiral;
[0008] Step 2: Establish the dynamic model of the quadruped robot according to the parameters of the physical prototype, and establish a drive-energy consumption model based on the LuGre friction model; construct the constraints and performance indicators of the quadruped robot control through body trajectory tracking, joint torque tracking and foot contact force tracking, and use quadratic programming to solve the optimization in the MPC-based motion planning controller by setting weight parameters.
[0009] Step 3: For each design parameter instance, use reinforcement learning to jointly optimize the design parameters and weight parameters. The reward function encourages the robot to move forward with minimal energy consumption and punishes hasty and inefficient actions. By learning the basic strategy π, the current state x t ∈R 18 and the last action a t-1 ∈R 7 As input, predict the next action a t , a t is the set of all optimized parameters.
[0010] Step 4: Apply the parameters obtained after reinforcement learning optimization to the controller, input the desired gait and speed into the gait phase controller, calculate the desired foot contact timing list, and then use the timing list, weight parameters, desired gait and speed as the input of MPC optimization, and use quadratic programming and the state information fed back by the robot to solve the foot contact force, and then solve the joint torque; the joint uses feedforward control and PID control to track the desired torque to drive the robot to move.
[0011] Furthermore, the parameters of the dynamic model in step 2 include mass, moment of inertia about the joint, center of mass position of each leg and center of mass position of the trunk, joint inertia and friction coefficient.
[0012] Furthermore, in step 2, the entire robot is regarded as a rigid body, and single rigid body dynamics modeling is performed. The resultant external force at the center of gravity of the rigid body is:
[0013]
[0014] Where p b is the coordinate of each point in the rigid body, and the resultant external force is simplified to The Euler equations for a rotating rigid body are: Add gravity and transform to world coordinates:
[0015]
[0016] Where I: inertia tensor, f is : The combined force of ground support and friction received by each foot end, p gi : The coordinates of the vector from the robot's center of gravity to the foot end i in the world coordinate system {s}, R sb : Conversion matrix from the body coordinate system to the world coordinate system, g: gravitational acceleration, m: total mass.
[0017] Furthermore, the drive-energy consumption model based on the LuGre friction model in step 2 includes the following steps: given the motion state of each joint, the torque of each joint is calculated according to the Newton-Euler method:
[0018]
[0019] Where M(θ i ) is the inertia matrix, is a vector containing the Coriolis moment and the centripetal moment, G(θ i ) is a vector containing the gravity torque; according to the LuGre friction model, the friction torque of each joint can be calculated as follows:
[0020]
[0021] Where z, are the average deformation of the bristles of the high-speed shaft and the average deformation rate of the bristles, F ci is the Coulomb friction, F si is the static friction, w si is the Stribeck speed, is the static friction function; σ 0i , σ 1i , σ 2i are the bristle stiffness coefficient, bristle damping coefficient and viscous friction coefficient respectively. ;
[0022] The driving power of the i-th joint can be calculated as follows:
[0023]
[0024] where η m,i is the efficiency of the joint motor, η r,i is the efficiency of the joint reducer;
[0025] The comprehensive energy consumption during the entire exercise process is as follows:
[0026]
[0027] Furthermore, in step 3, the state space of reinforcement learning is selected as the roll angle, roll angle, pitch angle and posture of the quadruped robot body, the landing time, position and expected time and position error of the foot end, and the torque output by each leg.
[0028] Furthermore, the vector of the state space is normalized, and the normalization formula is as follows:
[0029]
[0030] Where x is the input value and x∈(x min ,x max ), x new is the normalized output, and x new ∈(-1,1).
[0031] Furthermore, in step 3, the action space of reinforcement learning is designed as the design parameter set D of the elastic element and the weight parameters P(i), S(i), U(i) and W(i) of the model predictive controller.
[0032] Furthermore, in step 3, the reward function of reinforcement learning is set as follows:
[0033]
[0034] Among them, r1 is the reward for moving in the target direction; r2 is the angle of the body that is punished for deviating from the target direction; r3 is the roll angle and pitch angle of the body that are punished; r4 is the excessive energy consumption of the motor; r5 is the undesirable angular velocity of the body that is punished; k i is the weight; p t is the position at time t; Δθ roll is the change in lateral angle; Δθ pitch is the pitch angle change value.
[0035] Furthermore, in step 4, the body posture, velocity, acceleration and ground reaction force during the movement are fed back to the MPC optimization controller and the gait timing controller through the onboard IMU sensor to update the robot's motion state information, and dynamically adjust the gait timing according to the landing information.
[0036] Furthermore, the MPC-based motion planning controller includes the following modules:
[0037] Gait planning module: used to switch gaits during the movement of the quadruped robot and dynamically adjust the timing of subsequent gaits according to the state information of the ground contact, so as to enhance the robustness of the gait;
[0038] Model predictive control module: Based on the parameters optimized by reinforcement learning and the expected gait and speed input, the foot contact force is optimized, the joint torque, speed and acceleration are solved, and the optimized model predictive control is deployed on the physical prototype;
[0039] Joint execution module: tracks the execution of the motor according to the torque, velocity, and acceleration information output by the model predictive control module;
[0040] Communication network module: connects the twelve motors of the quadruped robot and the main control board into one, reducing the communication delay between the main control board and each motor.
[0041] Compared with the existing technology, by analyzing the energy consumption model and dynamic model of the quadruped robot joint, a drive-energy consumption model is established based on the LuGre friction model, and reinforcement learning is used to optimize the design parameters of the series elastic drive and the parameters of the quadruped robot motion controller. While ensuring the stability of the quadruped robot's movement, all resources are fully utilized to improve its energy efficiency. Reinforcement learning is used in the design stage to simulate and find the optimal design parameters, which reduces the time required to redesign the elastic elements due to unreasonable design in the later stage, and at the same time improves the upper limit of the quadruped robot's energy efficiency from the hardware perspective. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 It is the spring design drawing of the present invention;
[0043] Figure 2 It is a module framework diagram of the controller of the present invention;
[0044] Figure 3 It is a flow chart of the control method of the present invention. DETAILED DESCRIPTION
[0045] The technical solution of the present invention is further described in detail below with reference to the accompanying drawings and embodiments.
[0046] The energy efficiency optimization method of a quadruped robot based on a series elastic drive of the present invention comprises the following steps:
[0047] Step 1: Figure 1 As shown, the design parameter set of the elastic element of the series elastic actuator is set: D, which includes the joint stiffness k, the parameters a, b, c of the Archimedean spiral, the stiffness range is determined according to the motor output torque, and the motor size is used as the constraint condition of the Archimedean spiral.
[0048] Step 2: The dynamic model of the quadruped robot is established according to the parameters of the physical prototype, including mass, moment of inertia of the joint, center of mass position of each leg and the center of mass position of the trunk, joint inertia and friction coefficient. At the same time, a drive-energy consumption model based on the LuGre friction model is established. The constraints and performance indicators of the quadruped robot control are constructed by tracking the body trajectory, joint torque and foot contact force. The quadratic programming (QP) is used to solve the optimization in the motion planning controller based on MPC (model predictive control) by setting weight parameters.
[0049] Step 3: If Figure 3 As shown in the figure, for each design parameter instance, reinforcement learning is used to jointly optimize the design parameters and weight parameters. The reward function encourages the robot to move forward with minimal energy consumption and punishes hasty and inefficient actions. By learning the basic strategy π, the current state x t ∈R 18 and the last action a t-1 ∈R 7 As input, predict the next action a t , a t is the set of all optimized parameters.
[0050] Step 4: Figure 3 As shown in the figure, the parameters obtained after reinforcement learning optimization are applied to the controller, the desired gait and speed are input into the gait phase controller, the desired foot contact timing list is calculated, and then the timing list, weight parameters, desired gait and speed are used as the input of MPC optimization. The foot contact force is solved by using quadratic programming and the state information fed back by the robot, and then the joint torque is solved. The joint uses feedforward control and PID control to track the desired torque and drive the robot to move. The robot uses the IMU and other sensors on the body to feed back the body posture, speed, acceleration and ground reaction force during the movement to the MPC optimization controller and gait timing controller, update the robot motion state information, and dynamically adjust the gait timing according to the landing information.
[0051] In step 2, the influence of the four legs on the dynamics is ignored, and the entire robot is regarded as a rigid body for single rigid body dynamics modeling. The resultant external force at the center of gravity of the rigid body is:
[0052]
[0053] Where p b is the coordinate of each point in the rigid body, and the resultant external force is simplified to The Euler equations for a rotating rigid body are: Add gravity and transform to world coordinates:
[0054]
[0055] Where I: inertia tensor
[0056] f is : The combined force of ground support and friction received by each foot end
[0057] p gi : The coordinates of the vector from the robot's center of gravity to the foot end i in the world coordinate system {s}
[0058] R sb : Transformation matrix from fuselage coordinate system to world coordinate system
[0059] g: acceleration due to gravity
[0060] m: total mass
[0061] The drive-energy consumption model based on the LuGre friction model in step 2 is established as follows: given the motion state of each joint (i.e., angle, velocity, and acceleration), the torque of each joint is calculated according to the Newton-Euler method:
[0062]
[0063] Where M(θ i ) is the inertia matrix, is a vector containing the Coriolis moment and the centripetal moment, G(θ i ) is a vector containing the gravity torque. According to the LuGre friction model, the friction torque of each joint can be calculated. as follows:
[0064]
[0065] Where z, are the average deformation of the bristles of the high-speed shaft and the average deformation rate of the bristles, F ci is the Coulomb friction, F si is the static friction, w si is the Stribeck speed, is the static friction function; σ 0i , σ 1i , σ 2i are the bristle stiffness coefficient, bristle damping coefficient and viscous friction coefficient respectively.
[0066] Based on the above formula, the driving power of the i-th joint can be calculated as follows:
[0067]
[0068] where η m,i is the efficiency of the joint motor, η r,i is the efficiency of the joint reducer.
[0069] In summary, the comprehensive energy consumption during the entire exercise process is as follows:
[0070]
[0071] In step 3, the state space of reinforcement learning includes the observations of the robot during the training process, but not all of them can be selected. Too many observations will cause the reinforcement learning model to be over-coupled during the training process, making it difficult to find the optimal solution. Too many observations will contain a lot of useless information, which will slow down the training. In the present invention, the state space is selected as the roll angle, roll angle, pitch angle and posture of the quadruped robot body, the foot end landing time, position and expected time, position error, and the torque output by each leg, totaling 21 states, as shown in Table 1 below.
[0072] Dimensions Feedback Variable Name 1-3 IMU Fuselage roll, roll and pitch angles 4-6 IMU Displacement of the center of mass of the fuselage in the x, y, and z directions 7-8 Joint encoder Errors between foot landing time, position and expected value 9-21 Joint encoder Four-leg joint torque
[0073] Table 1. State space composition table.
[0074] The above state information is continuous value, which is represented by vector: S = [s1, s2, ... s 17 ]. To facilitate network learning, the vectors in the state space are normalized to remove the effects of different value ranges. The normalization formula is as follows:
[0075]
[0076] Where x is the input value and x∈(x min ,x max ), x new is the normalized output, and x new ∈(-1,1).
[0077] The action space of reinforcement learning is designed as the design parameter set D of the elastic element and the weight parameters P(i), S(i), U(i) and W(i) of the model predictive controller.
[0078] In reinforcement learning, the reward function is set as follows:
[0079]
[0080] Among them, r1 is the reward for moving in the target direction; r2 is the angle of the body that is punished for deviating from the target direction; r3 is the roll angle and pitch angle of the body that are punished; r4 is the excessive energy consumption of the motor; r5 is the undesirable angular velocity of the body that is punished; k i is the weight; p t is the position at time t; Δθ roll is the change in lateral angle; Δθ pitch is the pitch angle change value.
[0081] The basic strategy π is implemented as a multi-layer perceptron, i.e., an artificial neural network. The basic strategy π is trained based on the reference trajectory, and the objective function g(π) is:
[0082]
[0083] Where τ={(x0,a0,r0),(x1,a1,r1)...} is the trajectory of the execution strategy, P(τ|π) represents the probability of the trajectory τ appearing under the strategy π, and the termination parameter The optimal action space parameters are solved by continuous iteration. Due to the limited manufacturing precision, the design parameters are changed to dynamically change with the optimization parameters, with a maximum change value of 10%, to improve the matching of the spring model.
[0084] like Figure 2 As shown, the motion planning controller based on MPC (model predictive control) includes:
[0085] Gait planning module: used to switch gaits during the movement of the quadruped robot and dynamically adjust the timing of subsequent gaits according to the state information of the ground contact, so as to enhance the robustness of the gait;
[0086] Model predictive control module: Based on the parameters optimized by reinforcement learning and the expected gait and speed input, the foot contact force is optimized, the joint torque, speed and acceleration are solved, and the optimized model predictive control is deployed on the physical prototype;
[0087] Joint execution module: tracks the execution of the motor according to the torque, velocity, and acceleration information output by the model prediction control module to ensure the accuracy of motor tracking;
[0088] Communication network module: connects the twelve motors of the quadruped robot and the main control board into one, reduces the communication delay between the main control board and each motor, and ensures timely processing of signals.
[0089] The control purpose of the model predictive control module is to enable the robot to achieve stable posture trajectory following. In each iteration, according to the robot dynamics equation and constraints, the optimal control input sequence and the corresponding state trajectory sequence within a certain length of time series are solved. Since each control cycle will recalculate the control sequence after the next expected control quantity, it is only necessary to use the first control input quantity of the control input sequence as the controller output to control the expected input quantity of the supporting leg in the next control cycle. In the next control cycle, the model predictive controller repeats the above solution and output process to achieve rolling optimization of the control quantity. According to the dynamics equation in step 2, define the matrix A and vector b d :
[0090]
[0091] Therefore, the kinetic equation is abbreviated as Af-b d = 0, construct the cost function J in step 4. We hope that the foot end force is as small as possible and the sudden change of the foot end force f is not too large, so the cost function is set as follows:
[0092]
[0093] In the formula, f prev is the foot end force calculated in the previous round, S, P, U, and W are the corresponding weight matrices, i.e., the weight parameters optimized in step 3.
[0094] Assume that the friction coefficient between the ground and the foot end is μ, and the reaction force of the ground on the foot end is f is , then the constraint when the foot touches the ground is:
[0095]
[0096] The constraint when the foot is in the air is If is =0, integrate the constraints corresponding to the four legs together, and establish corresponding equality constraints and inequality constraints.
[0097] The present invention utilizes a spring to store and release energy during movement, thereby reducing the joint torque required for movement. At the same time, the characteristics of the spring are utilized to filter tiny disturbances, thereby enhancing the stability of movement. Meanwhile, only the angular displacement of the output end of the elastic body needs to be measured, thereby having better feedback driving performance.
[0098] Therefore, the present invention is divided into an offline reinforcement learning process and a real-time model prediction motion planning controller. Offline reinforcement learning observes the rollover angle, roll angle, pitch angle and posture of the quadruped robot body, the foot end landing time, position and expected time, position error, and the torque output by each leg for a total of 21 states, and uses the LuGre friction model to construct an energy consumption evaluation function. In the design of the reward function, energy consumption is used as one of the evaluation indicators. For the design parameter set D of each elastic element, the optimal controller weight parameters P(i), S(i), U(i) and W(i) are found. In a simulation environment, the real-time model prediction motion planning controller sets the cost function according to the weight parameters of reinforcement learning and the dynamic equation to control the quadruped robot to move smoothly. While ensuring the robustness of the quadruped robot's motion tracking performance, the energy consumption during the movement process is evaluated, and reinforcement learning-simulation energy consumption analysis iterations are continuously performed to select the optimal set of design parameters, and real machine verification is performed.
[0099] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.
Claims
1. A method for optimizing the energy efficiency of a quadruped robot based on a series elastic actuator, comprising the following steps: Step 1: Set the design parameter set of the elastic element of the series elastic actuator: D, which includes the joint stiffness k, the parameters a, b, c of the Archimedean spiral, determine the stiffness range according to the motor output torque, and use the motor size as the constraint condition of the Archimedean spiral; Step 2: Establish the dynamic model of the quadruped robot according to the parameters of the physical prototype, and establish the drive-energy consumption model based on the LuGre friction model; construct the constraints and performance indicators of the quadruped robot control through body trajectory tracking, joint torque tracking and foot contact force tracking, and use quadratic programming to solve the optimization in the MPC-based motion planning controller by setting weight parameters; Step 3: For each design parameter instance, reinforcement learning is used to jointly optimize the design parameters and weight parameters. The reward function encourages the robot to move forward with minimal energy consumption and punishes hasty and inefficient actions. By learning the basic strategy π, the current state x t ∈R 18 and the last action a t-1 ∈R 7 As input, predict the next action a t , a t is the set of all optimized parameters; Step 4: Apply the parameters obtained after reinforcement learning optimization to the controller, input the desired gait and speed into the gait phase controller, calculate the desired foot contact timing list, and then use the timing list, weight parameters, desired gait and speed as the input of MPC optimization, and use quadratic programming and the state information fed back by the robot to solve the foot contact force, and then solve the joint torque; the joint uses feedforward control and PID control to track the desired torque to drive the robot to move.
2. The method for optimizing the energy efficiency of a quadruped robot based on a series elastic actuator according to claim 1, characterized in that: The parameters of the dynamic model in step 2 include mass, moment of inertia about the joint, center of mass position of each leg and the center of mass position of the trunk, joint inertia, and friction coefficient.
3. The method for optimizing the energy efficiency of a quadruped robot based on a series elastic actuator according to claim 1, characterized in that: In step 2, the entire robot is regarded as a rigid body, and single rigid body dynamics modeling is performed. The resultant external force at the center of gravity of the rigid body is: Where p b is the coordinate of each point in the rigid body, and the resultant external force is simplified to The Euler equations for a rotating rigid body are: Add gravity and transform to world coordinates: Where I: inertia tensor, f is : The combined force of ground support and friction received by each foot end, p gi : The coordinates of the vector from the robot's center of gravity to the foot end i in the world coordinate system {s}, R sb : Conversion matrix from the body coordinate system to the world coordinate system, g: gravitational acceleration, m: total mass.
4. The method for optimizing the energy efficiency of a quadruped robot based on a series elastic actuator according to claim 1, characterized in that: The drive-energy consumption model based on the LuGre friction model in step 2 includes the following steps: Given the motion state of each joint, the torque of each joint is calculated according to the Newton-Euler method: Where M(θ i ) is the inertia matrix, is a vector containing the Coriolis moment and the centripetal moment, G(θ i ) is a vector containing the gravity torque; according to the LuGre friction model, the friction torque of each joint can be calculated as follows: Where z, are the average deformation of the bristles of the high-speed shaft and the average deformation rate of the bristles, F ci is the Coulomb friction, F si is the static friction, w si is the Stribeck speed, is the static friction function; σ 0i , σ 1i , σ 2i are the bristle stiffness coefficient, bristle damping coefficient and viscous friction coefficient respectively; The driving power of the i-th joint can be calculated as follows: where η m,i is the efficiency of the joint motor, η r,i is the efficiency of the joint reducer; The comprehensive energy consumption during the entire exercise process is as follows:
5. The method for optimizing the energy efficiency of a quadruped robot based on a series elastic actuator according to claim 1, characterized in that: In step 3, the state space selected for reinforcement learning is the roll angle, roll angle, pitch angle and posture of the quadruped robot body, the landing time, position and expected time of the foot end, the error of the position, and the torque output by each leg.
6. The method for optimizing the energy efficiency of a quadruped robot based on a series elastic actuator according to claim 5, characterized in that: Normalize the vector in the state space. The normalization formula is as follows: Where x is the input value and x∈(x min ,x max ), x new is the normalized output, and x new ∈(-1,1).
7. The method for optimizing the energy efficiency of a quadruped robot based on a series elastic actuator according to claim 1, characterized in that: In step 3, the action space of reinforcement learning is designed as the design parameter set D of the elastic element and the weight parameters P(i), S(i), U(i) and W(i) of the model predictive controller.
8. The method for optimizing energy efficiency of a quadruped robot based on a series elastic actuator according to claim 1, characterized in that: In step 3, the reward function for reinforcement learning is set as follows: Among them, r1 is the reward for moving in the target direction; r2 is the angle of the body that is punished for deviating from the target direction; r3 is the roll angle and pitch angle of the body that are punished; r4 is the penalty for excessive motor energy consumption; r5 is the penalty for the body's undesirable angular velocity; k i is the weight; p t is the position at time t; Δθ roll is the change in lateral angle; Δθ pitch is the pitch angle change value.
9. The method for optimizing energy efficiency of a quadruped robot based on a series elastic actuator according to claim 1, characterized in that: In step 4, the body posture, velocity, acceleration and ground reaction force during the movement are fed back to the MPC optimization controller and gait timing controller through the onboard IMU sensor to update the robot's motion state information and dynamically adjust the gait timing according to the landing information.
10. The method for optimizing energy efficiency of a quadruped robot based on a series elastic actuator according to claim 1, characterized in that: The MPC-based motion planning controller includes the following modules: Gait planning module: used to switch gaits during the movement of the quadruped robot and dynamically adjust the timing of subsequent gaits according to the state information of the ground contact, so as to enhance the robustness of the gait; Model predictive control module: Based on the parameters optimized by reinforcement learning and the expected gait and speed input, the foot contact force is optimized, the joint torque, speed and acceleration are solved, and the optimized model predictive control is deployed on the physical prototype; Joint execution module: tracks the execution of the motor according to the torque, velocity, and acceleration information output by the model predictive control module; Communication network module: connects the twelve motors of the quadruped robot and the main control board into one, reducing the communication delay between the main control board and each motor.
Citation Information
Cited By
Robot energy consumption optimization method based on improved particle swarm optimization
CN119347852A
Multi-legged robot slope marching control method and system based on double-peak Gaussian fusion
CN120066104A
Quadruped robot reinforcement learning gait optimization method and system
CN120370711A
Reinforcement learning gait optimization method and system for quadruped robots
CN120370711B
Robot motion control method, electronic equipment, storage medium and program product
CN122195016A
