A terrain-aware multi-quadruped robot model prediction and whole-body hierarchical cooperation control method

Abstract

The present application belongs to the technical field of robot control and multi-agent cooperation, and particularly relates to a multi-quadruped robot model prediction and whole-body layered cooperation control method based on terrain perception, which firstly constructs a two-layer planning and control architecture, a global motion planner generates a globally optimized trajectory of a multi-robot cooperation carrying system based on a target point by using model predictive control (MPC); secondly, each robot is equipped with a distributed motion controller, and an optimal local trajectory considering a safe foothold at a future time is generated in combination with a terrain elevation map generated by a sensor; finally, a whole-body controller (WBC) uniformly solves and allocates feedforward torque instructions of each driving joint according to motion task requirements. Compared with a single machine or a traditional cooperation strategy under a flat terrain, the present application has better terrain self-adaptive capability and carrying process stability, can stably cross uneven terrains such as steps and discrete stepping stones, and is suitable for heavy-load cooperation carrying scenes such as field exploration and industrial complex limited spaces.

Description

Technical Field

[0001] This invention belongs to the field of robot control and multi-agent cooperation technology, specifically involving a method for predicting multi-quadruped robot models and implementing whole-body hierarchical cooperative control based on terrain perception. Background Technology

[0002] In recent years, quadruped robots have been widely used in modern industry and field exploration. However, in actual working scenarios, robot systems often face unstructured, complex, and ever-changing environmental constraints. When multiple quadruped robots need to cooperate to transport heavy objects, traditional methods struggle to achieve stable transport of the manipulated object in three-dimensional space, such as in discrete uneven terrain or three-dimensional undulating terrain. These complex environments not only significantly exacerbate the motion coupling characteristics between the robot and the work object but also introduce uncertainties such as terrain disturbances and abrupt changes in contact states, greatly increasing the difficulty of collaborative control.

[0003] Current research has proposed an integrated perception-planning-control motion framework based on nonlinear model predictive control (MMC). This framework transforms complex terrain into convex polygon foot constraints through elevation map processing, significantly improving the autonomous navigation capability of single robots in unstructured environments. However, existing terrain-aware MPC solutions are only designed for single-robot scenarios and have not been adapted or extended to scenarios with multiple robots and rigid load coupling. The lack of global coordination and low-level compliant disturbance rejection mechanisms to address multi-robot coupling is a core technical bottleneck for the application of multi-quadruped robots in real-world complex environments. Therefore, there is an urgent need for a hierarchical cooperative control method that can fuse and optimize the coupling of local terrain perception and global multi-robot motion states. To this end, this invention provides a terrain-aware multi-quadruped robot model prediction and whole-body hierarchical cooperative control method. Summary of the Invention

[0004] The purpose of this invention is to provide a terrain-aware multi-quadruped robot model prediction and whole-body hierarchical cooperative control method to solve the problems of poor terrain adaptability, drastic motion coupling abrupt changes, and lack of stable operation guarantee in uneven environments in existing multi-quadruped robot cooperative systems. The invention proposes a terrain-aware multi-quadruped robot model prediction and whole-body hierarchical cooperative control method, which helps to improve the system's adaptive passage capability and cooperative load accuracy in discrete steps and undulating terrain.

[0005] The specific technical solution adopted by this invention is as follows: A method for model prediction and whole-body hierarchical cooperative control of a multi-quadruped robot based on terrain perception; including the following steps: Step 1: Construct a global motion planner, which includes a model of a multi-quadruped robot cooperative system with rigid linkages and its kinematic geometry. Step 2: In the global motion planner, a reference trajectory is generated based on the target position set by the user, and a global model predictive control (MPC) problem is constructed to optimize the expected state sequence of the payload in the future time domain. Step 3: Use robot sensors to collect point cloud data, convert the point cloud into a 2.5D elevation grid map, and perform terrain segmentation and smoothing filtering to obtain local terrain elevation features; Step 4: In the distributed motion controller, based on the elevation map and the global planned trajectory, design a local MPC to solve for the safe landing point and the optimal target trajectory of the fuselage on the discrete terrain; Step 5: Construct a whole-body controller (WBC) based on a weighted quadratic programming (QP) problem, encapsulate local trajectory tracking, contact constraints, and dynamic consistency into multi-task constraints, and solve for the reference feedforward torques of each joint of the multi-quadruped robot. Step 6: Based on the hybrid control law of error integral compensation and predictive feedforward, design a joint torque controller to drive the underlying motor to track the desired trajectory with high precision.

[0006] Preferably, the design of the global motion planner in step 2 specifically includes: Construct a state vector containing payload position, quaternion, linear velocity, and angular velocity. and an input vector containing the force and torque applied to the load. ; Optimizing global MPC: ; The cost function aims to minimize the deviation between the load state and the reference trajectory, as well as the input cost. ; The dynamic equations include the dynamics of the effective load, including the connecting rod, and involve mass. Moment of inertia And quaternion multiplication.

[0007] Preferably, the local elevation map processing and landing point planning in steps 3 and 4 specifically include: Based on the terrain height information after smoothing filtering and expected baseline height Centroid height: ; Fitting the plane equation using the least squares method Estimate local roll and pitch attitude; A heuristic formula based on velocity feedback is used to set the nominal landing point, and it is projected into the safe and feasible terrain domain within the kinematic limits.

[0008] Preferably, the design of the whole-body controller (WBC) in step 5 specifically includes: Define decision variables It includes generalized acceleration, foot contact force, and joint torque; Constructing a standard QP optimization problem: ; The following hard constraints must also be met: Condition a: Satisfies the consistency constraint of the dynamic equations: ; Condition b: Satisfy the kinematic constraints of the supporting leg foot contact: ; Condition c: Satisfy the joint torque boundary limit of the actual hardware motor and the friction cone inequality constraint of the contact surface.

[0009] Preferably, the underlying torque control law in step 6 specifically includes: For the swing leg, a joint acceleration error integral compensation term is introduced to suppress unmodeled dynamic disturbances: ; For the support leg, a hybrid control law combining feedforward torque and PD error adjustment is applied: .

[0010] The technical effects achieved by this invention are as follows: This invention comprises a two-layer control architecture: a global motion planner and a distributed motion controller. The method includes the following key steps: the system model uses two quadruped robots connected by rigid links with rotational degrees of freedom at both ends, allowing the system to adapt smoothly to terrain undulations; the global motion planner, operating at 100Hz, generates an optimized trajectory for the multi-robot load system based on the target point and terrain height constraints, using global MPC. The distributed motion controller, operating at 200Hz, combines a 2.5D elevation map with a resolution of 4cm generated from the point cloud, and uses local MPC to generate an optimal local reference sequence that takes into account both future safe landing points and fuselage pose. The whole-body controller (WBC), operating at 800Hz, models the dynamics and various contact constraints as a weighted quadratic programming (QP) problem, uniformly solving and issuing feedforward torque commands. The underlying joint controller introduces a hybrid control approach combining error integral term and predictive feedforward, effectively overcoming the instantaneous torque impact caused by sudden changes in ground contact. This invention not only effectively decouples the high-dimensional multi-machine collaborative problem in complex environments into a hierarchical optimization real-time planning problem, but also theoretically verifies the robustness of position and attitude tracking through numerous simulation experiments in scenarios such as stepping stones and continuous steps. Compared with traditional flat terrain strategies, this invention can effectively mitigate terrain disturbances, ensure contact safety, and enhance the impact resistance of control signals, demonstrating significant theoretical value and promising engineering applications. Attached Figure Description

[0011] Figure 1 A framework diagram for model prediction and whole-body hierarchical cooperative control of a terrain-aware multi-quadruped robot is described.

[0012] Figure 2 illustrates the generation of the elevation map; where Figure 2a It is a model diagram and; Figure 2b This is a schematic diagram after the elevation map has been processed.

[0013] Figure 3 The Gazebo simulation implementation of the stepping stone scene is described.

[0014] Figure 4 illustrates the temporal variation of the load's center of mass spatial motion trajectory and attitude angle under the stepping stone scenario; where... Figure 4a It is the trajectory of the load's center of mass; Figure 4b It is the attitude of the load center of mass.

[0015] Figure 5 illustrates the performance of left foreleg joint angular displacement tracking and angular velocity tracking in the stepping stone scenario; where Figure 5a It is joint position tracking; Figure 5b It's joint velocity tracking.

[0016] Figure 6 illustrates the torque variation of the left foreleg joint under the stepping stone scenario; where... Figure 6a It is the WBC feedforward torque vs. the joint output torque; Figure 6b It is the joint output torque vs. the actual feedback torque.

[0017] Figure 7 A schematic diagram of the Gazebo environment test for a stepped scene is shown.

[0018] Figure 8 illustrates the temporal variation of the load's centroid spatial motion trajectory and attitude angles in a stepped scenario; where... Figure 8a It is the trajectory of the load's center of mass; Figure 8b It is the attitude of the load center of mass.

[0019] Figure 9 illustrates the tracking performance of the left foreleg joint angular displacement and angular velocity in a stepped scenario. Figure 9a It is joint position tracking; Figure 9b It's joint velocity tracking.

[0020] Figure 10 illustrates the torque variation of the left foreleg joint in a stepped scenario. Figure 10a It is the WBC feedforward torque vs. the joint output torque; Figure 10b It is the joint output torque vs. the actual feedback torque. Detailed Implementation

[0021] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one or more specific embodiments of the invention and does not strictly limit the scope of protection specifically claimed by the invention.

[0022] like Figure 1 As shown, this invention relates to the field of robot control and multi-agent cooperation technology, specifically to a multi-quadruped robot model prediction and whole-body hierarchical cooperative control method based on terrain perception, and more particularly to a cooperative planning control method integrating environmental elevation map perception, two-level model predictive control (MPC), and weighted whole-body control (WBC), applicable to complex control scenarios such as field material transportation, industrial inspection, and disaster relief; including the following steps: Step 1: Construct a global motion planner, which includes a model of a multi-quadruped robot cooperative system with rigid linkages and its kinematic geometry. Step 2: In the global motion planner, a reference trajectory is generated based on the target position set by the user, and a global model predictive control (MPC) problem is constructed to optimize the expected state sequence of the payload in the future time domain. Step 3: Use robot sensors to collect point cloud data, convert the point cloud into a 2.5D elevation grid map, and perform terrain segmentation and smoothing filtering to obtain local terrain elevation features; Step 4: In the distributed motion controller, based on the elevation map and the global planned trajectory, design a local MPC to solve for the safe landing point and the optimal target trajectory of the fuselage on the discrete terrain; Step 5: Construct a whole-body controller (WBC) based on a weighted quadratic programming (QP) problem, encapsulate local trajectory tracking, contact constraints, and dynamic consistency into multi-task constraints, and solve for the reference feedforward torque of each joint of the multi-quadruped robot; Step 6: Based on the hybrid control law of error integral compensation and predictive feedforward, design a joint torque controller to drive the underlying motor to track the desired trajectory with high precision.

[0023] The working principle of this invention is as follows: Step a: Implementation method for global motion planner design This step aims to build a global motion planner. Its core objective is to generate a safe and efficient optimized path in unstructured, complex terrain environments, taking into account system kinematics and physical constraints, to guide a team of robots to collaboratively transport loads to a designated target location.

[0024] The global motion planner is modeled as a model predictive control (MPC) problem. First, the state vector of the system is defined. With input vector The state vector describes the pose and motion state of the payload in the global coordinate system, and is expressed as: ; in The location of the center of mass. The pose is represented by a quaternion. For linear velocity, Angular velocity. Input vector. The forces and moments exerted on the payload by the cooperative system are described as follows: ; in, To work together, This is the resultant torque.

[0025] The goal of a global motion planner is to optimize the solution so that the generated trajectory tracks the user-defined reference trajectory to the greatest extent possible. This ensures both the safety and physical feasibility of the handling process. The MPC problem is described as follows: Establish the effective load dynamic equations including the connecting rod ; The constraints are: ; in, Indicates the terminal cost. This represents the intermediate cost within the prediction time domain. This is the dynamic equation of the load system.

[0026] To achieve smooth and accurate trajectory tracking, a quadratic cost function is designed as follows. This function not only penalizes state deviations but also penalizes the control input to avoid severe torque jitter in the system: Intermediate cost: ; Terminal costs: ; in, and These are the terminal weight matrix, the state weight matrix, and the input weight matrix, respectively.

[0027] Considering the linkage coupling characteristics, the dynamic equation of the effective load including the linkage is expressed as follows: ; In the formula, and These are the total mass and moment of inertia of the system, respectively. This represents quaternion multiplication.

[0028] Then step b: Elevation map construction and local MPC design This step aims to ensure that a single quadruped robot can accurately track the optimized target trajectory issued by the global motion planner through local elevation sensing and local model predictive control (MPC). It calculates the safe foothold position for each leg on discrete terrain and adjusts the target trajectory to maintain the stable movement of the robot system.

[0029] The local MPC controller is constructed using a simplified floating-base dynamics model. The robot's overall body dynamics equations are expressed as follows: ; in, For generalized coordinates, For generalized speed, The inertia matrix, It is a nonlinear term. To access the Jacobian matrix, This represents the joint torque vector. Since the leg inertia is relatively small, this embodiment simplifies the process using the floating base dynamics equations: ; in Let the floating base's attitude and velocity vectors be... Let be the buoyancy component in the whole-body dynamics. Therefore, the state vector of the local MPC is defined. With input vector for: ; The optimization problem of local MPC is expressed as: ; Constraints include the dynamic equations and environmental safety and physical limits .

[0030] The elevation map is represented using a 2.5D grid with a spatial resolution of 4 cm. A GPU-accelerated method is used to convert the point cloud data acquired by the sensor into an elevation grid map in real time. For blind spots or occluded areas in the sensor detection, a minimum value filling method is used for data repair along the occlusion boundaries. After obtaining the original elevation map, terrain segmentation and smoothing filtering are performed to extract the geometric features of the ground.

[0031] Integrating the global instructions yields the two-dimensional planar trajectory. Center of mass reference height Calculated from topographic elevation information: ; in The terrain height information has been processed using a smoothing algorithm. This represents the desired fuselage reference ground clearance.

[0032] Attitude angle estimation employs a least-squares plane fitting method. The planar positions of each hip joint in the global coordinate system are calculated recursively using forward kinematics. And obtain the corresponding height from the elevation map. Construct the fitting equation Its matrix representation is as follows: ; Solve for parameters Then, the roll and pitch angles of the robot body are calculated using the plane normal vector.

[0033] The landing point planning and safety projection employ a strategy combining heuristic methods and velocity feedback capture point formulas. At the midpoint of the contact phase, the nominal landing point... The calculation is as follows: ; To ensure safety in discrete terrains, Projected onto the feasible terrain region ; ; in This is a kinematic penalty function used to penalize kinematic infeasibility caused by excessive foot elongation, ultimately obtaining the projected safe landing point. .

[0034] Then, step c: Design of weighted whole-body control (WBC) based on quadratic programming. This step aims to optimize the global / local instructions (desired foot force) generated in step b. With joint velocity This maps precise torque control commands to each joint of the robot. By constructing a weighted whole-body control (WBC) framework, the multi-task coordination problem is transformed into a standard quadratic programming (QP) problem for online solution, thereby achieving the unification of motion planning, force distribution, and torque mapping.

[0035] Define a unified decision vector ; in The generalized acceleration vector that describes changes in the body's motion state. For the set of contact reaction force vectors acting on the foot, This is the joint torque vector output by the actuator.

[0036] Construct a quadratic programming problem of the following form: ; in, Here, is the Hessian matrix, used to characterize the curvature of the objective function; The gradient vector; To constrain the Jacobian matrix; and These are the lower and upper bounds, respectively.

[0037] To ensure that the robot's motion conforms to the laws of physics, the following constraints are set as hard constraints that must be strictly satisfied: System-wide dynamic consistency constraints: ensuring that the joint driving torque, generalized acceleration, and foot contact force constitute a dynamically compatible solution. ; Foot-end friction pyramid constraint: To ensure no-slip contact, the nonlinear friction cone is linearized into a friction pyramid constraint. For the supporting leg... It must meet the following requirements: ; For swinging leg Then the constraint .

[0038] Kinematic constraints for foot contact on supporting leg: Based on a hard contact model, ensure that the position, velocity, and acceleration of the foot contact point are all zero in the global coordinate system. ; Joint torque limiting: Ensure that the output torque of each drive joint motor does not exceed its rated output limit. : ; Based on hard constraints, three tracking tasks are introduced, and differentiated configurations are made through weight coefficients: Swing leg foot acceleration tracking: generating acceleration commands through proportional-derivative control laws. And map it to a generalized coordinate space, with the following constraints: ; Airframe acceleration tracking: The aim is to minimize the deviation between the desired and actual acceleration at the center of mass. The constraint form is as follows: ; Foot contact force tracking: Requires WBC to calculate the contact force. The expected contact force predicted by NMPC Numerical values ​​are equal: ; Finally; simulation experiments verify and underlying torque execution. This step performs final execution control on the torque command output by WBC, and builds a simulation platform to comprehensively evaluate the effectiveness, robustness, and adaptability of the proposed algorithm.

[0039] Figure 1 This paper demonstrates the overall architecture and flow of the terrain-aware-based multi-quadruped robot model prediction and whole-body hierarchical cooperative control method proposed in this invention. The entire control chain achieves fine-grained hierarchical cooperation on both spatial and temporal scales, and the joint torque commands output by the WBC (Weighted Bracket Control Unit) are shown. As the input to the joint torque controller, a force-position hybrid control framework is used for driving at the lowest level.

[0040] To improve trajectory tracking accuracy, the swing leg control employs the following strategy: suppressing unmodeled dynamics and external disturbances, a compensation term based on the joint acceleration error integral is introduced. This term is equipped with a gain coefficient K>0, defined as follows: ; in This represents the initial moment of the oscillation phase. Mathematically, this construction utilizes the velocity difference term. Instead of actual joint acceleration Direct measurement avoids the noise amplification problem caused by high-frequency differential measurement of sensor signals in engineering. High-precision acceleration layer error compensation can be calculated solely based on the difference between the current and initial velocities. To address integral saturation, output limiting is implemented through a feedback term.

[0041] To balance tracking accuracy and disturbance rejection rigidity under high loads, the support leg control strategy employs a hybrid control law that combines feedforward torque prediction and PD error adjustment: ; in These are the gain parameters, which, in physical essence, inject virtual stiffness and virtual damping into the supporting legs, respectively, in conjunction with the WBC's dynamic feedforward torque. This greatly enhances the system's positional stiffness and impact-resistant convergence speed during heavy-duty collaborative handling.

[0042] The experimental verification platform was built on the Gazebo simulator and ROS Noetic environment. The frequency allocation for each module in the simulation was as follows: global MPC 100Hz, local MPC 200Hz, and WBC 800Hz. Prediction time domain settings: global MPC 5 seconds, local MPC 3 seconds, making the global trajectory slightly ahead of the local planner. The Pinocchio library was used to perform online dynamics calculations. Safe zones were extracted through point cloud processing and online segmentation. The overall online construction and data flow of the perception system are shown in Figure 2.

[0043] Scene 1: Stepping Stone Scene Setup: The platform contains three discrete stepping stones, each 60cm wide and 45cm apart. Figure 3 Figure 4 shows the Rviz visual perception interface and Gazebo physical simulation environment of the bipedal quadruped robot performing heavy-load collaborative transport in this discrete terrain. The two robots need to accurately plan their footing points based on the online elevation map to traverse the cliff-like terrain.

[0044] Results analysis: Position and Attitude: Analysis of the three-dimensional spatial motion trajectory of the load centroid and the temporal variation of attitude angles shown in Figures 5(a) and 5(b) reveals that the load centroid trajectory is smooth, without deviation, and the motion is stable. Yaw, pitch, and roll angles are suppressed within a very small range, indicating good attitude stability. This strongly demonstrates that the global-local dual-layer MPC architecture designed in this invention can perfectly decouple external load interference and ensure the macroscopic stability of the multi-machine collaborative system.

[0045] Tracking performance: Analysis of the joint position and velocity tracking curves shown in Figure 6 reveals that the joint position tracking is good, with minor local deviations; in velocity tracking, the sudden contact change at the moment the foot touches the ground / leaves the ground causes an instantaneous impact, resulting in a peak in the hip joint velocity.

[0046] Torque performance: Further comparison Figure 7 The response curves of the feedforward torque and the actual feedback torque in the WBC (Wide-Body-Controlled Charge) revealed that at the moment of foot contact with the ground, the expected velocity abruptly changes, causing the compensating torque to increase instantaneously, resulting in a spike. Torque tracking was excellent during the non-contact phase. This high-frequency torque control mechanism reproduced the expected dynamic distribution in real time, demonstrating the effectiveness of the WBC.

[0047] Scene 2: Stairs Setup: Four steps, 13cm high, 30cm wide, with a slope of approximately 25°. Figure 8 shows a real-time climbing test scenario of the multi-quadruped robot facing a terrain with continuous height variations, testing the system's gait adaptive adjustment capability under continuous terrain disturbances.

[0048] Results analysis: Airframe adaptability: Referring to Figures 9(a) and 9(b), the spatial motion trajectory and attitude timing response of the load under the stepped terrain, the pitch angle (approximately -0.4 rad) is consistent with the slope of the steps, indicating that the airframe has good adaptability to the inclination of the stairs, and the system as a whole did not experience any excessive oscillation or instability and collapse caused by sudden changes in altitude.

[0049] Joint tracking: According to the details of the position and velocity tracking of each joint of the robot's left front leg given in Figure 10, the joint position tracking effect is good, and the hip joint velocity spikes due to the impact when it touches the ground.

[0050] Torque Analysis: According to Figure 10(a), at the moment the foot touches the ground, since the desired velocity is always zero, there is a large deviation between the desired velocity obtained by inverse solution and the actual velocity, which causes a sudden surge (peak) in the compensation torque in the supporting leg control law formula. Nevertheless, combined with the torque comparison diagram in Figure 10(b), it can be clearly seen that the actual output torque does not produce obvious lag or instability, and the real-time reproduction capability is excellent, proving the robustness of this control strategy under terrain with continuous height undulations.

[0051] Combination Figures 1 to 1 Based on system-level topology and experimental curve analysis, the control method presented in this invention has the following significant advantages: Active stability of multi-machine heavy-load collaboration: Through the deep cascading of the dual-layer MPC framework and the bottom WBC, the system can still lock the attitude fluctuations of the fuselage and load within a very small robust range when transporting nonlinear energy loads up to 6kg (as shown in Figures 5 and 9), avoiding the phase decoupling and tension instability that are very easy to occur in traditional multi-machine systems.

[0052] Excellent high-frequency disturbance rejection and unmodeled dynamic suppression: The joint acceleration error integral compensation term introduced by the swing leg works in conjunction with the force-position hybrid control law of the supporting leg to rapidly absorb and compensate for link pulling, nonlinear static friction and contact model uncertainty without introducing high-frequency noise (such as the system's agile performance at 800Hz), significantly reducing the physical impact at the moment the foot touches the ground.

[0053] High terrain adaptability and traversability: Utilizing online fine-grained segmentation of 2.5D elevation maps and multi-gait rolling optimization, the robot can smoothly and autonomously traverse discrete stepping stones and steep stairs (such as... Figure 3 As shown in Figure 8, this expands the industrial application prospects of multi-quadruped robot systems in unstructured confined spaces such as field exploration and complex industrial heavy-duty collaborative handling.

[0054] The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained in this invention are implemented according to conventional methods in the art unless otherwise specified or limited.

Claims

1. A method for model prediction and whole-body hierarchical cooperative control of a multi-quadruped robot based on terrain perception, characterized in that: Includes the following steps: Step 1: Construct a global motion planner, which includes a model of a multi-quadruped robot cooperative system with rigid linkages and its kinematic geometry. Step 2: In the global motion planner, a reference trajectory is generated based on the target position set by the user, and a global model predictive control (MPC) problem is constructed to optimize the expected state sequence of the payload in the future time domain. Step 3: Use robot sensors to collect point cloud data, convert the point cloud into a 2.5D elevation grid map, and perform terrain segmentation and smoothing filtering to obtain local terrain elevation features; Step 4: In the distributed motion controller, based on the elevation map and the global planned trajectory, design a local MPC to solve for the safe landing point and the optimal target trajectory of the fuselage on the discrete terrain; Step 5: Construct a whole-body controller (WBC) based on a weighted quadratic programming (QP) problem, encapsulate local trajectory tracking, contact constraints, and dynamic consistency into multi-task constraints, and solve for the reference feedforward torques of each joint of the multi-quadruped robot. Step 6: Based on the hybrid control law of error integral compensation and predictive feedforward, design a joint torque controller to drive the underlying motor to track the desired trajectory with high precision.

2. The method for multi-quadruped robot model prediction and whole-body hierarchical cooperative control based on terrain perception according to claim 1, characterized in that: The design of the global motion planner in step 2 specifically includes: Construct a state vector containing payload position, quaternion, linear velocity, and angular velocity. ,in The location of the center of mass. The pose is represented by a quaternion. For linear velocity, The angular velocity; and the input vector containing the force and torque applied to the load. ;in, To work together, The resultant torque; Optimizing global MPC: ； Among them, among them, Indicates the terminal cost. This represents the intermediate cost within the prediction time domain. This indicates the payload state at the predicted time-domain terminal moment. Given the dynamic equations of the load system, the cost function aims to minimize the deviation between the load state and the reference trajectory, as well as the input cost: ； in, Payload status, It is a reference trajectory deviation. This indicates the reference trajectory deviation at the predicted time-domain terminal moment. and These are the terminal weight matrix, state weight matrix, and input weight matrix, respectively; the dynamic equations include the dynamics of the effective load, including the linkage, and involve mass. Moment of inertia And quaternion multiplication.

3. The method for multi-quadruped robot model prediction and whole-body hierarchical cooperative control based on terrain perception according to claim 1, characterized in that: The local elevation map processing and landing point planning in steps 3 and 4 specifically include: Based on the terrain height information after smoothing filtering and expected baseline height Centroid height: ； in; Provides a pose reference for discrete state sequences; fits plane equations using the least squares method. Estimate local roll and pitch attitude, among which and The coefficients of the plane equation to be identified; A heuristic formula based on velocity feedback is used to set the nominal landing point, and it is projected into the safe and feasible terrain domain within the kinematic limits.

4. The method for multi-quadruped robot model prediction and whole-body hierarchical cooperative control based on terrain perception according to claim 1, characterized in that: The design of the whole-body controller (WBC) in step 5 specifically includes: Define decision variables , A generalized acceleration vector that describes changes in the body's motion state; This represents the set of contact reaction force vectors acting on the foot. This is the sequence of joint torques required to be output by the actuator. Constructing a standard QP optimization problem: ； Among them; among them, Here, is the Hessian matrix, used to characterize the curvature of the objective function; The gradient vector; To constrain the Jacobian matrix; and These represent the lower and upper bounds, respectively. The generalized acceleration vector that describes changes in the body's motion state. For the set of contact reaction force vectors acting on the foot, This refers to the joint torque vector output by the actuator. The following hard constraints must also be met: Condition a: Satisfies the consistency constraint of the dynamic equations: ,in, The inertia matrix, For Jacobian matrices, To select a matrix, For the decision variables that need to be solved, It is a nonlinear term; Condition b: Satisfy the kinematic constraints of the supporting leg foot contact: ,in, To access the Jacobian matrix, For the current generalized acceleration; Condition c: Satisfy the joint torque boundary limit of the actual hardware motor and the friction cone inequality constraint of the contact surface.

5. The method for multi-quadruped robot model prediction and whole-body hierarchical cooperative control based on terrain perception according to claim 1, characterized in that: The underlying torque control law in step 6 specifically includes: For the swing leg, a joint acceleration error integral compensation term is introduced to suppress unmodeled dynamic disturbances: ； in, Meaning the first The total torque command for each joint motor driver. This is the WBC feedforward torque command. This is the integral compensation gain coefficient. The starting time of the oscillation phase, This represents the actual acceleration during the current shutdown. For the desired joint acceleration in WBC, This represents the actual shutdown speed at the moment. Just as I lifted my leg The actual speed of the joints recorded at that instant; For the support leg, a hybrid control law combining feedforward torque and PD error adjustment is applied: ； in, Meaning the first The total torque command for each joint motor driver. This is the joint proportional gain coefficient. For joint position tracking error, The joint differential gain coefficient, For joint velocity tracking.

Images