A control method, device and equipment of a wheeled robot and a storage medium

Abstract

This application discloses a control method, device, equipment, and storage medium for a wheeled robot in the field of robot control technology. The method includes: establishing a dynamic model and a wheel control model for the wheeled robot; using the MPC algorithm to initially predict the motion state and motion control input for the next N steps based on the dynamic model; using the PPO algorithm to analyze the wheel shape and wheel control input for the next N steps based on the motion state predicted by the MPC algorithm; using the MPC algorithm to re-optimize and predict the motion state and motion control input for the next N steps based on the wheel shape predicted by the PPO algorithm; controlling the wheeled robot to execute the control input, and optimizing the PPO algorithm based on the execution result; returning to execute the initial prediction step using the MPC algorithm until the convergence condition is met, and then using the MPC algorithm and the PPO algorithm to achieve control of the wheeled robot. Therefore, this application can effectively improve the control effect of wheeled robots with variable wheel shapes in different environments.

Description

Technical Field

[0001] This application relates to the field of robot control technology, and more specifically, to a control method, device, equipment, and storage medium for a wheeled robot. Background Technology

[0002] Wheeled robots are robots that use wheels as their locomotion mechanism. They are the most common type of mobile robot and are widely used in various fields such as industry, service, and scientific research.

[0003] Traditional control methods for wheeled robots primarily rely on precise dynamic modeling. While these methods perform well in stable environments, their control effectiveness is severely limited by the stability of the external environment. When operating in complex, dynamic, or uncertain environments, model-based control methods often experience control failures or poor performance. This is especially true for systems using morphologically variable wheels, where the nonlinear characteristics of the wheels themselves make it difficult to establish accurate dynamic models for control.

[0004] In summary, existing technologies for wheeled robots with variable wheel shapes cannot guarantee ideal control performance in different environments. Summary of the Invention

[0005] The purpose of this application is to provide a control method, device, equipment, and storage medium for wheeled robots, which can achieve relatively ideal control effects in different environments for wheeled robots with variable wheel shapes.

[0006] To achieve the above objectives, this application provides the following technical solution: A control method for a wheeled robot includes: A dynamic model and a wheel control model for a wheeled robot are established; wherein, the dynamic model describes the relationship between the motion state and motion control input of the wheeled robot, and the wheel control model describes the relationship between the wheel shape and wheel control input of the wheeled robot; Using the MPC algorithm, the motion state and corresponding motion control input of the wheeled robot in the next N steps are initially predicted based on the dynamic model. Using the PPO algorithm, the wheel shape and corresponding wheel control input of the wheeled robot in the next N steps are determined based on the motion state analysis obtained from the wheel control model and the MPC algorithm prediction. Using the MPC algorithm, the wheel shape predicted by the PPO algorithm is re-optimized and the motion state and corresponding motion control input of the wheeled robot in the next N steps are predicted; The wheeled robot is controlled to execute the predicted wheel control inputs and the latest motion control inputs, and the PPO algorithm is optimized based on the execution results. Return to the initial prediction step using the MPC algorithm until the convergence condition is met, and then use the MPC algorithm and PPO algorithm to control the wheeled robot.

[0007] Preferably, optimizing the PPO algorithm based on the execution results includes: Based on the execution results, obtain the balance maintenance data, trajectory tracking data, energy consumption data, and wheel shape adjustment data of the wheeled robot during the execution process; Based on the balance maintenance data, trajectory tracking data, energy consumption data, and wheel shape adjustment data, a reward function value is calculated to evaluate the performance of the PPO algorithm. The PPO algorithm is optimized based on the reward function value.

[0008] Preferably, controlling the wheeled robot to execute the predicted wheel control input and the latest motion control input includes: The wheeled robot is controlled to execute predicted wheel control inputs and the latest motion control inputs in a preset simulation environment; wherein the preset simulation environment is a pre-constructed environment containing different terrains and load conditions.

[0009] Preferred options also include: When controlling the wheeled robot using the MPC algorithm and PPO algorithm, the motion control input output by the MPC algorithm and the wheel control input output by the PPO algorithm are detected in real time. If at least one of the control inputs exceeds the corresponding execution range, the at least one control input is adjusted to the corresponding execution range.

[0010] Preferred options also include: When using the MPC algorithm and PPO algorithm to control the wheeled robot, if the wheeled robot's motion state and / or wheel shape become abnormal during the execution of corresponding motion control inputs and / or wheel control inputs, the wheeled robot is controlled to stop an emergency action.

[0011] Preferred options also include: The current motion state of the wheeled robot is collected. If the collected motion state data is lost or abnormal, the lost or abnormal data is compensated by the historical data difference method or the state estimation method. The PPO algorithm is then used to comprehensively analyze the current motion state of the wheeled robot and the motion state predicted by the MPC algorithm to determine the wheel shape and corresponding wheel control input of the wheeled robot for the next N steps.

[0012] Preferably, the wheels of the wheeled robot are flexible wheels with variable radial stiffness and diameter, and the wheel shape includes the radial stiffness and diameter of the flexible wheel.

[0013] A control device for a wheeled robot includes: A module is established for: establishing a dynamic model and a wheel control model for a wheeled robot; wherein, the dynamic model describes the relationship between the motion state and motion control input of the wheeled robot, and the wheel control model describes the relationship between the wheel shape and wheel control input of the wheeled robot; The initial prediction module is used to: use the MPC algorithm to make an initial prediction of the motion state and corresponding motion control input of the wheeled robot for the next N steps based on the dynamic model; The analysis module is used to: use the PPO algorithm to analyze and determine the wheel shape and corresponding wheel control input of the wheeled robot for the next N steps based on the motion state predicted by the wheel control model and the MPC algorithm; The re-prediction module is used to: use the MPC algorithm to re-optimize the wheel shape predicted by the PPO algorithm and predict the motion state and corresponding motion control input of the wheeled robot for the next N steps; The execution module is used to: control the wheeled robot to execute the predicted wheel control input and the latest motion control input, and optimize the PPO algorithm based on the execution results; The control module is used to: return to the step of performing the preliminary prediction using the MPC algorithm until the convergence condition is met, and then use the MPC algorithm and PPO algorithm to control the wheeled robot.

[0014] A control device for a wheeled robot, the device comprising a memory and a processor, wherein a program stored in the memory, when executed by the processor, can implement the steps of the control method for the wheeled robot as described above.

[0015] A computer-readable storage medium, wherein a program stored on the computer-readable storage medium, when executed by a processor, implements the steps of the control method for a wheeled robot as described in any of the preceding claims.

[0016] This application provides a control method, device, equipment, and storage medium for a wheeled robot. The method includes: establishing a dynamic model and a wheel control model for the wheeled robot, wherein the dynamic model describes the relationship between the motion state and motion control input of the wheeled robot, and the wheel control model describes the relationship between the wheel shape and wheel control input of the wheeled robot; using the MPC algorithm, preliminarily predicting the motion state and corresponding motion control input of the wheeled robot for the next N steps based on the dynamic model; using the PPO algorithm, analyzing and determining the wheel shape and corresponding wheel control input of the wheeled robot for the next N steps based on the motion state predicted by the wheel control model and the MPC algorithm; using the MPC algorithm, re-optimizing and predicting the motion state and corresponding motion control input of the wheeled robot for the next N steps based on the wheel shape predicted by the PPO algorithm; controlling the wheeled robot to execute the predicted wheel control input and the latest motion control input, and optimizing the PPO algorithm based on the execution result; returning to the step of preliminarily predicting using the MPC algorithm until the convergence condition is met, and then using the MPC algorithm and the PPO algorithm to control the wheeled robot. As can be seen, the technical solution of this application combines the precise trajectory planning capability of the MPC algorithm with the strategy optimization characteristics of the PPO algorithm through an algorithm interaction mechanism. It retains the precise trajectory planning capability of the MPC algorithm and uses the PPO algorithm to dynamically adjust the wheel shape to adapt to terrain changes. This can effectively solve the problems of modeling difficulties when the wheel shape is variable and poor control effect in complex environments, thereby effectively improving the control effect of wheeled robots with variable wheel shape in different environments. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a control method for a wheeled robot provided in an embodiment of this application; Figure 2 A schematic diagram illustrating the interaction process between the MPC algorithm and the PPO algorithm in a control method for a wheeled robot provided in an embodiment of this application; Figure 3 An example diagram illustrating the variation of the flexible wheel diameter in a control method for a wheeled robot provided in this application embodiment; Figure 4 An example diagram illustrating the change in radial stiffness of a flexible wheel in a control method for a wheeled robot provided in this application embodiment; Figure 5A schematic diagram illustrating the implementation of each functional component in a control method for a wheeled robot provided in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of a control device for a wheeled robot provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] Please see Figure 1 The diagram illustrates a flowchart of a control method for a wheeled robot provided in an embodiment of this application, which may specifically include: S11: Establish the dynamic model and wheel control model of the wheeled robot; wherein, the dynamic model describes the relationship between the motion state of the wheeled robot and the motion control input, and the wheel control model describes the relationship between the wheel shape of the wheeled robot and the wheel control input.

[0021] It should be noted that the wheeled robots in this application are all robots with variable-shape wheels (such as bipedal robots). Based on this, by analyzing the response characteristics of the system to which the variable-shape wheels belong in complex terrain, it was found that there is a two-way influence relationship between the motion state of the wheeled robot and the wheel shape adjustment. If the two are treated separately, it will lead to the accumulation of control errors. Based on this, this application proposes to iteratively interact Model Predictive Control (MPC) and Proximal Policy Optimization (PPO) to establish a two-layer optimization architecture through dynamic parameter feedback, so that motion control and wheel shape adjustment form a closed-loop optimization.

[0022] The dynamic equations of a wheeled robot can be derived using Lagrange mechanics as its dynamic model. This model describes the relationship between the overall motion state of the wheeled robot (which may include the robot's linear velocity, joint angular velocity, and joint angles) and the control input (i.e., motion control input, the input used to control the wheeled robot to achieve the corresponding motion state). This can be expressed as: ;in, It is a quality matrix. It is the matrix of Coriolis force and centrifugal force. It is the gravity vector. q It is a state of motion. This is the motion control input, which requires precise measurement or estimation of the wheeled robot's physical parameters (such as mass, inertia, etc.) to achieve.

[0023] For wheels with variable shapes, a nonlinear relationship between different wheel shapes (wheel forms) and control inputs (i.e., wheel control inputs, which are inputs used to control the wheels of the wheeled robot to achieve the corresponding shapes) can be established through experiments or theoretical analysis, serving as a wheel control model.

[0024] After modeling is completed, the accuracy of the model can be verified through simulation or experimentation to ensure that it can truly reflect the dynamic characteristics of the wheeled robot system. Furthermore, to address model uncertainties, robust modeling methods, such as interval analysis or fuzzy theory, can be introduced to enhance the model's adaptability.

[0025] In addition, motion state, motion control input, wheel shape, and wheel control input can all be vectors, and other settings can be made according to actual needs, all of which are within the protection scope of this application.

[0026] S12: Using the MPC algorithm, the motion state and corresponding motion control input of the wheeled robot in the future N steps are initially predicted based on the dynamic model.

[0027] The Multiprocessing Control (MPC) algorithm is an optimization control method based on a system dynamic model. It predicts future system behavior by solving an optimization problem at each control time step and calculates the control input for the current time step based on these predictions. Specifically, the MPC algorithm uses the system's mathematical model to predict the state trajectory over a future period. Then, considering constraints (such as input limitations and state limitations), it minimizes the error or cost function of predicting the future state by optimizing the control action at the current time step. After each control step, the MPC algorithm recalculates the optimization problem and updates the control strategy, allowing the control strategy to be gradually optimized through continuous feedback, adapting to dynamic changes in the system. Its main advantages lie in its ability to handle complex constrained and multivariable systems, exhibiting good control performance and flexibility.

[0028] In this embodiment, the MPC algorithm is used for short-cycle motion trajectory planning based on a dynamic model. Specifically, the logic of the MPC algorithm in this step is based on model prediction and rolling optimization, optimizing the current control input by predicting the system behavior over a future period. First, based on the dynamic model, MPC predicts the future... The motion state of step (N can be set according to actual needs) ( t (moment state at any time) and motion control input ( t (motion control input at time step), and construct the cost function. J , ,in , and These are all weight matrices used to balance motion state errors and control energy consumption. t Indicates time ( t (Time), T represents transpose. This represents the predicted motion states for the next N steps (the set of future N-step motion states); then, numerical optimization methods are used to solve the optimization problem of the cost function to generate the optimal motion control input. It should be noted that, in its underlying implementation, the MPC algorithm can employ sparse matrix optimization or parallel computing techniques to accelerate the solution process for optimization problems in order to efficiently handle real-time computations. Furthermore, to improve robustness, the MPC algorithm can incorporate constraints into the optimization problem, such as restrictions on at least one of the following: robot joint angles, robot joint angular velocities, robot linear velocities, and physical constraints on wheel configurations. Moreover, to address model prediction errors in the MPC algorithm, a feedback correction mechanism can be introduced to compare the actual motion state with the predicted motion state and dynamically adjust the MPC algorithm based on the comparison results.

[0029] S13: Using the PPO algorithm, the wheel shape and corresponding wheel control input of the wheeled robot in the future N steps are determined by analyzing the motion state predicted by the wheel control model and MPC algorithm.

[0030] The Policy Optimization (PPO) algorithm is a policy optimization algorithm (belonging to the policy gradient method) in reinforcement learning. It aims to ensure stability during policy updates, avoiding performance instability caused by excessive policy changes. Its core idea is to limit the magnitude of policy updates at each optimization step, ensuring the new policy doesn't deviate too far from the old one, thus maintaining stability. Specifically, the PPO algorithm uses an objective function that includes importance sampling and introduces a pruning operation within this objective function to prevent policy collapse due to excessive updates. The pruned objective function limits the proportion of policy updates, making policy updates more stable and less prone to divergence.

[0031] In this embodiment, the PPO algorithm generates a wheel shape adjustment strategy to adapt to terrain changes based on the motion state prediction results. Specifically, after predicting the motion state for the next N steps, the MPC algorithm transmits the predicted motion state and the current motion state of the wheeled robot to the PPO algorithm. Based on the motion state (including the current motion state and the predicted future motion state) transmitted by the MPC algorithm and the body information of the wheeled robot (such as joint angles, joint angular velocities, robot linear velocities, etc.), the PPO algorithm outputs the wheel shape adjustment strategy through its own deep neural network (i.e., the policy network), including the wheel shape to be controlled by the wheeled robot (which can be represented by the change data of the current wheel shape) and the corresponding wheel control input.

[0032] In addition, the MPC algorithm can also include a reward function, including functions describing robot balance, trajectory tracking accuracy, etc.; the ultimate goal is to optimize the parameters of the policy network by maximizing the pruned objective function, so that it can output accurate and effective wheel shapes and wheel control inputs; the objective function is... It can be represented as ,in clip It's a cropping operation; it will... r t (θ) Cut to make r t (θ) Stay and between; r t (θ) It is the ratio of the current strategy to the old strategy, that is, at time t. t Importance sampling ratio; It is a moment t The advantage function measures the "goodness" of a particular action relative to the expected reward under the current strategy, that is, the advantage of the action compared to other actions; It is a small hyperparameter used to limit r t (θ) The range of fluctuations; This indicates the expectation value, i.e., the average value within [ ]. The PPO algorithm in this embodiment can use stochastic gradient descent to update the parameters of the optimization strategy network, and combine the experience replay mechanism and exploration strategy to improve its training efficiency and exploration ability. At the same time, it can also use GPU accelerated computing and target network delayed update mechanism to ensure the stability and convergence of training, and finally realize the dynamic adjustment of wheel shape to adapt to environmental changes, and improve the robot's adaptability and robustness.

[0033] S14: Using the MPC algorithm, the wheel shape predicted by the PPO algorithm is re-optimized and the motion state and corresponding motion control input of the wheeled robot in the next N steps are predicted.

[0034] S15: Control the wheeled robot to execute the predicted wheel control input and the latest motion control input, and optimize the PPO algorithm based on the execution results.

[0035] In this embodiment, the interaction design logic between the MPC and PPO algorithms is based on state sharing and parameter feedback. Its core objective is to achieve collaborative optimization of short-term control and long-term strategy. First, the MPC algorithm transmits the prediction results to the PPO algorithm. The PPO algorithm outputs a wheel shape adjustment strategy based on these results and feeds back the adjusted wheel shape parameters to the optimization problem of the MPC algorithm. The MPC algorithm re-optimizes the motion control input based on these parameters and executes the re-optimized motion control input according to the adjusted wheel shape. At the same time, the execution result is fed back to the PPO algorithm for strategy update (including but not limited to the optimization and update of policy network parameters). Furthermore, if a feedback correction mechanism is introduced to address the model prediction error of the MPC algorithm, the execution result can also be fed back to the MPC algorithm, enabling it to optimize its own parameters based on the execution result. Specifically, this can be achieved by comparing the actual motion state of the wheeled robot for N steps with the motion state predicted by the MPC algorithm based on the execution result, and dynamically adjusting the MPC algorithm based on the comparison result.

[0036] In the underlying implementation, the interaction between the MPC and PPO algorithms requires an efficient communication mechanism. Asynchronous communication or multithreading techniques can be used to ensure real-time data exchange. Furthermore, an anomaly detection mechanism can be introduced to handle exceptional situations. When the output of either the MPC or PPO algorithm exceeds a pre-defined reasonable range, a safety strategy is triggered, such as stopping the movement or switching to a backup controller. Moreover, to improve the robustness of their interaction, a fault-tolerance mechanism can be designed so that if one algorithm fails, the other can run independently, ensuring system continuity.

[0037] S16: Return to the step of performing the initial prediction using the MPC algorithm until the convergence condition is met, and then use the MPC algorithm and PPO algorithm to control the wheeled robot.

[0038] The convergence condition can be that the overlap between the wheeled robot's motion trajectory and the predicted N-step motion trajectory reaches a preset value (such as 90%). Of course, other settings can be made according to actual needs, all of which are within the scope of protection of this application. Meeting the convergence condition indicates that the optimization of the MPC algorithm and PPO algorithm is complete, and therefore they can be deployed to the wheeled robot to realize the control of the wheeled robot.

[0039] This application uses the MPC algorithm to predict the robot's motion trajectory based on a dynamic model, providing dynamic environmental information for wheel shape adjustment. The PPO algorithm is then used to evaluate the impact of different wheel states on terrain adaptability based on this information, generating an optimal wheel shape adjustment strategy. The adjusted wheel shape is fed back to the MPC algorithm, allowing it to consider the impact of wheel shape changes on kinematic constraints during secondary optimization. Through multiple iterative optimizations, the MPC algorithm gradually corrects motion trajectory deviations, while the PPO algorithm synchronously updates the wheel adjustment strategy, ultimately forming a collaboratively optimized control strategy. During the hardware execution phase, real-time acquired robot state data can also be used to update algorithm parameters, ensuring continuous adaptability in dynamic environments.

[0040] In one specific implementation, the interaction process between the MPC algorithm and the PPO algorithm in this application embodiment can be as follows: Figure 2 As shown, "robot" is short for "wheeled robot," "state" is short for "motion state," "control input" is short for "motion control input," and "flexible wheel parameters" are parameters representing the shape of the flexible wheel. Specifically, the robot state, control parameters, and policy network parameters can be initialized first. The robot state refers to the robot's motion state, the control parameters are the parameters that control the robot's execution input, and the policy network parameters are the parameters of the policy network in the PPO algorithm. Initializing these items means setting them to preset initial values. Then, the MPC algorithm predicts the robot's motion state and motion control input for the next N steps and passes the predicted motion state and the robot's current motion state to the PPO algorithm. The PPO algorithm outputs an adjustment strategy for the flexible wheel parameters through its own policy network, calculates the reward function, updates the policy network parameters, and feeds back the adjusted flexible wheel parameters to the MPC algorithm. The MPC algorithm re-optimizes and generates new motion control input, executes the new motion control input, and feeds back the execution result to the PPO algorithm for further optimization and updating. Finally, it determines whether the target has converged (i.e., whether the convergence condition is met). If it has, the training is considered complete; otherwise, the MPC algorithm is executed to predict the motion state and motion control input for the next N steps.

[0041] As can be seen, the technical solution of this application combines the precise trajectory planning capability of the MPC algorithm with the strategy optimization characteristics of the PPO algorithm through an algorithm interaction mechanism. It retains the precise trajectory planning capability of the MPC algorithm and uses the PPO algorithm to dynamically adjust the wheel shape to adapt to terrain changes. This can effectively solve the problems of modeling difficulties when the wheel shape is variable and poor control effect in complex environments, thereby effectively improving the control effect of wheeled robots with variable wheel shape in different environments.

[0042] This application provides a control method for a wheeled robot. The wheels of the wheeled robot are flexible wheels with variable radial stiffness and diameter. The wheel shape includes the radial stiffness and diameter of the flexible wheel.

[0043] It should be noted that a flexible wheel with adjustable radial stiffness and diameter refers to a wheel structure with adjustable mechanical properties. This can be achieved using shape memory alloys, pneumatic actuators, or motor-driven telescopic structures. The stiffness and diameter are dynamically adjusted by changing the geometry or material properties of the internal support structure. Radial stiffness refers to the wheel's ability to resist deformation under radial loads, which can be achieved by adjusting the tension of the spokes or the arrangement of the support structure. For example, a servo motor can be used to drive a spoke telescopic mechanism to change the wheel's support stiffness. Diameter refers to the outer contour dimension of the wheel, which can be continuously adjusted using a telescopic or folding mechanism of the rim. For example, a worm gear mechanism can be used to drive the rim to unfold in segments. An example of how the robot wheel's diameter changes when encountering obstacles is as follows: Figure 3 As shown, an example of radial stiffness variation can be found as follows: Figure 4 As shown, the diameter and radial stiffness can vary in different ways under different environments or requirements, and all of these are within the scope of protection of this application.

[0044] In this embodiment, the wheels of the wheeled robot can be configured as flexible wheels that can change both radial stiffness and diameter to adapt to complex circular terrain. Specifically, for flexible wheels, their radial stiffness can be established through experiments or theoretical analysis. and diameter With adjustment angle and Nonlinear relationship between and Here, the adjustment angle refers to the rotation angle of the corresponding adjustment mechanism (such as a motor). Different motors can be set to control the radial stiffness and diameter respectively. The corresponding wheel shape adjustment strategy can be expressed as follows: , , This indicates the adjusted wheel shape. This represents the change in radial stiffness. This indicates the amount of change in diameter.

[0045] In one specific implementation, the radial stiffness adjustment mechanism and diameter adjustment mechanism of the flexible wheel are respectively configured inside the wheel body. When the wheeled robot travels on soft ground, the motor rotation angle can be controlled to reduce the radial stiffness to increase the ground contact area, while simultaneously increasing the diameter to improve ground clearance. When encountering hard obstacles, the radial stiffness can be increased to improve load-bearing capacity, while the diameter can be reduced to decrease the risk of collision. During the control process, the motor rotation angle is input to the wheel body actuator as a direct control quantity, and is converted into the actual adjustment amount of radial stiffness and diameter through a preset nonlinear mapping relationship. This adjustment process is embedded in the collaborative framework of the MPC algorithm and the PPO algorithm, enabling the flexible wheel parameters to respond in real time to the motion state changes predicted by the MPC algorithm, while the PPO algorithm optimizes the long-term adjustment strategy through a reward mechanism.

[0046] This solution combines the physically adjustable characteristics of flexible wheels with an MPC-PPO hybrid control framework, retaining the real-time advantages of the MPC algorithm while utilizing the PPO algorithm to optimize parameter adjustment strategies in complex terrains, thus avoiding the limitations of a single control method. It is evident that this application solves the problem of poor adaptability of traditional rigid wheels in complex terrains, achieving a control method that actively adapts to ground characteristics through physical adjustment. The morphological parameter adjustment capability of the flexible wheel enables the wheeled robot to automatically adjust its support characteristics on surfaces of varying hardness, avoiding motion risks in obstacle environments through size changes. Simultaneously, the direct control of the motor rotation angle ensures parameter adjustment accuracy and response speed. This design eliminates the need for a precise global dynamic model of the flexible wheel in the control algorithm, reducing the impact of modeling errors on control performance and minimizing the training time cost required by relying solely on learning algorithms.

[0047] This application provides a control method for a wheeled robot, which optimizes the PPO algorithm based on the execution result, and may include: Based on the execution results, obtain balance maintenance data, trajectory tracking data, energy consumption data, and wheel shape adjustment data of the wheeled robot during the execution process; The reward function value used to evaluate the performance of the PPO algorithm is calculated based on balance maintenance data, trajectory tracking data, energy consumption data, and wheel shape adjustment data. The PPO algorithm is optimized based on the reward function value.

[0048] Among these, balance maintenance data refers to the state parameters that maintain the robot's stability during movement. Specifically, this can be achieved by collecting the robot's tilt angle, angular velocity, and center of mass offset using attitude sensors, and is used to quantify the robot's anti-interference capability in dynamic environments. Trajectory tracking data refers to the deviation between the actual trajectory and the planned path (i.e., the target trajectory). This can be calculated using the coordinate difference between the real-time positioning system and the planned path, and is used to evaluate path following accuracy. Energy consumption data refers to the energy consumption values ​​of the drive motor and flexible wheel adjustment mechanism, which can be calculated using motor current, voltage, and the integral of running time, and is used to measure the energy efficiency of the control strategy. Wheel shape adjustment data refers to the changes in the radial stiffness and diameter parameters of the flexible wheel, which can be achieved by measuring the wheel structure deformation using encoders or displacement sensors, and is used to reflect the flexible wheel's ability to dynamically adapt to terrain. The reward function value is a quantitative evaluation result that integrates multiple performance indicators. Specifically, it can be achieved by linearly combining balance reward, tracking reward, energy reward, and flexible wheel adjustment reward using a weighted summation method, with the weight coefficients dynamically adjusted according to task requirements.

[0049] Specifically, during the execution of the wheeled robot, sensors acquire data on balance maintenance, trajectory tracking, energy consumption, and wheel configuration adjustment. Balance maintenance data is collected by attitude sensors; trajectory tracking data is obtained by comparing the planned path with real-time position data; energy consumption data is calculated by monitoring motor current and voltage; and wheel configuration adjustment data is recorded by an encoder showing changes in flexible wheel parameters. This data is transmitted to the PPO algorithm. Based on pre-set weight coefficients, the balance reward, tracking reward, energy reward, and flexible wheel adjustment reward are weighted and summed to generate a total reward function value. This reward function value serves as the optimization objective of the policy network. The network parameters are updated using a gradient ascent algorithm, allowing the policy network to progressively optimize the balance between multiple objectives in subsequent decisions. During the iteration process, when trajectory tracking accuracy improves but energy consumption increases, the reward function automatically adjusts the weights of each component, prioritizing a reasonable allocation between stability and energy consumption.

[0050] In one specific implementation, this application embodiment can design a reward function to evaluate the merits of the strategy, specifically including: balancing rewards. (Encourage the robot to maintain balance, punish tilting or falling), tracking rewards. (Encourage the robot to follow the set target trajectory, and punish deviations) Energy rewards (Encourage low energy consumption, penalize high energy input) and adjust rewards using a flexible wheel. (Encourage reasonable adjustments to the wheel shape, penalize excessively large or small adjustments), the total reward function is: ;in, , , , The preset coefficients, For the robot's tilt angle, For the robot's actual movement trajectory, For the target trajectory, To consume energy.

[0051] This application establishes a multi-dimensional evaluation system encompassing balance maintenance, trajectory tracking, energy efficiency, and mechanical state. This system enables the PPO algorithm to simultaneously coordinate conflicts among different performance indicators during optimization. For example, it prioritizes maintaining balance in complex terrain while focusing on reducing energy consumption on flat surfaces. Clearly, this application effectively solves the problem of unreasonable indicator weight allocation in multi-objective optimization, allowing the wheeled robot to maintain motion stability and accurately track the planned path while executing control commands, simultaneously considering energy utilization efficiency and the durability of the mechanical system. Furthermore, by dynamically adjusting the weight coefficients of each component in the reward function, it can adapt to the needs of different task scenarios. For instance, it increases the balance reward weight to enhance anti-tipping capability during emergency obstacle avoidance, while increasing the energy reward weight to extend endurance during long-distance cruising.

[0052] This application provides a control method for a wheeled robot, which controls the wheeled robot to execute predicted wheel control inputs and the latest motion control inputs, and may include: The wheeled robot is controlled to execute the predicted wheel control inputs and the latest motion control inputs in a preset simulation environment; the preset simulation environment is a pre-constructed environment that includes different terrains and load conditions.

[0053] The preset simulation environment refers to a virtual testing platform that simulates a real physical environment using a computer. Specifically, this can be achieved using multibody dynamics simulation software combined with terrain modeling tools. For example, different friction coefficients, slope parameters, or obstacle layouts can be set in the simulation environment to simulate actual terrain. Different terrain and load conditions refer to covering terrain types such as flat, rugged, and sloping terrain, as well as external loads of different masses or distributions. This can be achieved by adjusting the physical parameters of the simulation model or loading preset load configuration files. The execution verification of wheel control input and motion control input refers to converting the control signals output by the algorithm into action commands for the virtual actuator in the simulation environment. Specifically, this can be achieved by mapping the control input to the rotation angle or torque output of the wheel motor through the actuator interface in the simulation engine.

[0054] Specifically, test scenarios with various terrain and load conditions can be pre-constructed in the simulation environment, such as gravel ground, a 15-degree inclined road surface, and a 5 kg additional load configuration. After receiving motion control input optimized by the MPC algorithm and wheel control input generated by the PPO algorithm, the wheeled robot converts these input signals into motion commands for virtual motors through the simulation engine, driving the flexible wheels to adjust their radial stiffness and diameter before moving. The simulation environment calculates the contact force between the wheels and the terrain, the robot's center of mass displacement, and energy consumption data in real time, and feeds the results back to the various algorithms. By repeatedly executing the control input under different terrain and load conditions, the algorithm's adaptability in dynamic environments can be verified. For example, the impact of flexible wheel diameter adjustment on obstacle-crossing performance can be verified in rugged terrain, or the stability of motion control input can be evaluated under additional load conditions.

[0055] The embodiments of this application can be simulated and verified in a preset simulation environment. The simulation and verification design logic is based on performance evaluation and anomaly handling, with the core objective of verifying the effectiveness and robustness of the algorithm and providing support for practical deployment. Specifically, simulation scenarios with different terrains (such as flat, rugged, and sloping terrain) and load conditions can be constructed to evaluate the robot's balance performance, trajectory tracking accuracy, and energy consumption. In the underlying implementation, the simulation environment needs to accurately simulate the robot's dynamic behavior and the characteristics of the flexible wheels. To ensure the reliability of the simulation results, the simulation model needs to be calibrated to make it as consistent as possible with the actual physical system.

[0056] Through the above technical solutions, this application can safely verify the control input execution effect of wheeled robots under complex terrain and dynamic load in a virtual environment, avoiding mechanical damage caused by sudden terrain changes or overload in actual testing; it can quickly build a variety of test scenarios through simulation environment, significantly reducing the equipment configuration and debugging costs required for physical experiments; and it can generate multi-dimensional data using repeatable simulation tests, providing a quantitative basis for optimizing the adaptability of algorithms in dynamic environments.

[0057] The control method for a wheeled robot provided in this application embodiment may further include: When using the MPC algorithm and PPO algorithm to control a wheeled robot, the motion control input output by the MPC algorithm and the wheel control input output by the PPO algorithm are detected in real time. If at least one of the control inputs exceeds the corresponding execution range, the at least one control input is adjusted to the corresponding execution range.

[0058] In the process of implementing optimized MPC and PPO algorithms for controlling a wheeled robot, this application performs real-time detection of the motion control input output by the MPC algorithm and the wheel control input output by the PPO algorithm. When at least one control input exceeds the execution range, it is limited to the corresponding execution range through constraint adjustment. The execution range can be preset according to actual needs.

[0059] It should be noted that detecting motion control input and wheel control input refers to continuously acquiring the control signal values ​​output by the two algorithms through embedded sensors and the controller's built-in real-time monitoring module. For example, a current sensor is used to measure the amplitude of the motor drive signal, or an encoder is used to verify the steering angle command range. The purpose of this step is to capture the deviation between the algorithm output and the physical limitations of the actuator in real time, avoiding overload of mechanical components due to commands exceeding the limits. Adjusting to the execution range means using a dynamic limiter or constraint mapping algorithm to truncate out-of-range signals. For example, when the motor rotation angle of the wheel control input exceeds the actuator's maximum rotation angle, an angle limiter corrects the command value to the maximum allowable value, ensuring that the actuator always operates under safe conditions.

[0060] Within the control cycle where the MPC and PPO algorithms work in tandem, the calculation results of motion control input and wheel control input are acquired separately through parallel data channels. When any control input is detected to exceed the preset execution range, a constraint adjustment mechanism is immediately triggered: for continuous variables, the boundary projection method is used to map the values ​​to the effective interval; for discrete commands, the nearest feasible value is selected for replacement. This dynamic adjustment process maintains the direction of algorithm optimization while matching theoretical calculation results with actual execution capabilities. For example, in complex terrain, when the wheel diameter adjustment amount suggested by the PPO algorithm exceeds the hydraulic system's stroke limit, the system automatically corrects the adjustment amount to the maximum stroke value, preserving the core intent of the optimization strategy while ensuring the executability of the mechanical structure.

[0061] This application effectively prevents actuator overload damage or motion instability caused by control commands exceeding limits, ensuring the physical feasibility of control commands for the variable-structure flexible wheel system in complex terrain. Through a dynamic constraint adjustment mechanism, while maintaining the optimization effects of model predictive control and reinforcement learning algorithms, it achieves a precise match between theoretical calculation results and actual execution capabilities, significantly improving the system's robustness and safety in dynamic environments.

[0062] The control method for a wheeled robot provided in this application embodiment may further include: When using the MPC and PPO algorithms to control a wheeled robot, if the wheeled robot experiences an abnormal motion state and / or wheel shape while executing the corresponding motion control input and / or wheel control input, the wheeled robot will be controlled to stop an emergency action.

[0063] When using the MPC and PPO algorithms to control a wheeled robot, if the wheeled robot experiences abnormal motion or wheel shape during the execution of corresponding motion control inputs and wheel control inputs, the application will control the wheeled robot to stop urgently.

[0064] It should be noted that abnormal motion refers to the wheeled robot's attitude angle, acceleration, or displacement exceeding a preset safety threshold range. This can be achieved by using an inertial measurement unit to collect attitude data in real time and comparing it with kinematic equation predictions to identify the risk of motion instability caused by sudden terrain changes or load variations. Abnormal wheel shape refers to the rate of change of the flexible wheel's radial stiffness exceeding the material's deformation limit or the diameter adjustment exceeding the mechanical structure's allowable range. This can be achieved by using strain sensors to measure the wheel's deformation and combining it with the rotation angle feedback from the motor encoder for dynamic evaluation, preventing structural damage due to excessive wheel deformation. Emergency stop refers to interrupting motion by cutting off the drive motor power and activating the mechanical braking device. This can be achieved using a multi-stage braking strategy, such as first reducing the motor speed before triggering the electromagnetic brake to lock the axle, to reduce mechanical impact while ensuring braking efficiency.

[0065] During the MPC-PPO joint control process, the real-time monitoring system synchronously receives feedback data from motion state sensors and wheel shape sensors. When an attitude angle deviation exceeding a safety threshold is detected (e.g., pitch angle exceeding 15 degrees), or a sudden change in wheel diameter exceeding 5 mm within a single control cycle, the abnormal state determination module generates an interrupt signal. This signal is transmitted to the control execution unit, immediately cutting off the power supply to the drive motor and simultaneously activating the braking device in stages according to the current speed. During this process, both the MPC algorithm's rolling optimization process and the PPO algorithm's strategy update process are paused until manual intervention or system self-check confirms that the status has returned to normal.

[0066] Through the above technical solution, this application can detect in real time the coordination anomalies between motion state and wheel shape when a wheeled robot executes optimized control commands, and quickly eliminate safety hazards through a graded braking mechanism. For example, when crossing obstacles, if the robot's center of gravity shifts beyond the safe range due to a sudden change in the radial stiffness of the flexible wheel, the system can complete the state determination and trigger braking within 200 milliseconds to prevent rollover accidents caused by attitude loss of control. This mechanism effectively compensates for the defects of abnormal response delay in traditional algorithm control and improves operational safety in dynamic environments.

[0067] The control method for a wheeled robot provided in this application embodiment may further include: The current motion state of the wheeled robot is collected. If the collected motion state data is lost or abnormal, the historical data difference method or state estimation method is used to compensate for the lost or abnormal data. The PPO algorithm is then used to comprehensively analyze the motion state of the wheeled robot based on the current motion state of the wheeled robot and the motion state predicted by the MPC algorithm to determine the wheel shape and corresponding wheel control input of the wheeled robot for the next N steps.

[0068] If data loss or anomalies occur when collecting the current motion state of a wheeled robot, this application uses historical data difference or state estimation methods to compensate for the data, which is then used by the PPO algorithm to analyze and determine the wheel shape and wheel control input for the next N steps in conjunction with the current motion state and MPC prediction results.

[0069] Historical data interpolation refers to interpolation calculations using valid data from adjacent time points, specifically linear interpolation or spline interpolation, to fill in missing data points while preserving motion trend information. State estimation methods involve deducing the current state through system dynamics models or observer algorithms, specifically using Kalman filtering or extended Kalman filtering, and combining sensor noise characteristics and system dynamic characteristics for state reconstruction.

[0070] When motion state data collected by sensors is lost due to signal interference or hardware failure, the system first detects the time and location of the abnormal data points. If adjacent historical data is complete, an interpolation function is calculated based on the valid data points in the time series, such as using cubic spline interpolation to generate compensation data with a smooth transition. If the historical data is insufficient to support interpolation or the abnormal duration is long, the system switches to state estimation mode, using the dynamic model in the MPC algorithm to calculate the predicted values ​​of state variables in real time, and then using the observer algorithm to fuse and correct them with the remaining valid sensor data. The compensated motion state data is input into the PPO algorithm, and together with the motion state predicted by the MPC algorithm for the next N steps, it serves as the basis for decision-making, ensuring that wheel control inputs that conform to the system's dynamic constraints can still be generated during periods of data abnormality.

[0071] This application can maintain the integrity of the input data of the control algorithm when the sensor data is abnormal, avoid the failure of the PPO algorithm decision due to data loss, ensure that the wheeled robot can still continuously execute the optimized control commands in complex environments, and improve the robustness of the system under non-ideal data conditions.

[0072] Furthermore, this application embodiment can incorporate special exception handling mechanisms during algorithm deployment to address uncertainties in practical applications. For example, when data is lost or abnormal, historical data interpolation or state estimation methods are used for compensation to ensure the reliability of the control algorithm's input data; when the control input exceeds the actuator's range, a limiting strategy is used to control the input to prevent actuator damage or system malfunction; when the environment undergoes sudden changes, such as the sudden appearance of obstacles or terrain changes, an emergency stop is triggered to ensure the robot's safety; when the MPC algorithm or PPO algorithm fails, a preset rule-based control strategy is switched to ensure the robot can continue operating; when communication between the MPC algorithm and the PPO algorithm is delayed or interrupted, a caching mechanism or prediction mechanism is used to ensure the continuity and stability of the control system. Furthermore, various test scenarios can be designed in simulation verification, including extreme conditions such as high load, high speed movement, and complex terrain, to comprehensively evaluate the performance of the control method. By analyzing the simulation results, control parameters and strategies can be optimized to improve the robustness and adaptability of the system. By monitoring the wheeled robot's status and control output in real time, the algorithm's stable operation can be ensured, and a user interface can be provided to facilitate users in monitoring the status, adjusting parameters, and viewing logs. Ultimately, this enables the mobile platform to move efficiently and stably in diverse environments and effectively handle various abnormal situations, ensuring safe operation.

[0073] In one specific implementation method, such as Figure 5 As shown, the control method for a wheeled robot provided in this application embodiment can include five parts, namely: System modeling: Establish a dynamic model of the wheeled robot to describe the relationship between motion state variables such as joint angles, angular velocities, and linear velocities and motion control inputs; establish the nonlinear relationship between wheel shape variables such as radial stiffness and diameter of the flexible wheel and the adjustment angle. MPC Short-Term Control: Based on the dynamic model, MPC predicts the motion state and motion control input for a period of time in the future. By minimizing the cost function, it generates short-term motion control input to ensure the robot's balance and trajectory tracking, and then applies the optimized motion control input to the robot. PPO long-term strategy optimization: Obtain prediction results and robot body information from MPC to construct the state space of PPO; use PPO to update the policy network and output the wheel shape adjustment strategy; feed the adjusted flexible wheel shape back into the optimization problem of MPC; MPC interacts with PPO: MPC passes the prediction results to PPO, PPO outputs the wheel state adjustment strategy of the flexible wheel, MPC combines the adjusted wheel state to generate the final motion control input and executes it, and at the same time feeds the execution result back to PPO for strategy update; Simulation and Deployment: Construct simulation environments with different terrains (flat, rugged, sloping, etc.) and load conditions to evaluate the robot's balance performance, trajectory tracking accuracy, and energy consumption. Verify the effectiveness and robustness of the control method through simulation experiments.

[0074] This application also provides a control device for a wheeled robot, such as... Figure 6 As shown, it can specifically include: Module 11 is established to: establish the dynamic model and wheel control model of the wheeled robot; wherein, the dynamic model describes the relationship between the motion state and motion control input of the wheeled robot, and the wheel control model describes the relationship between the wheel shape and wheel control input of the wheeled robot; Preliminary prediction module 12 is used to: use the MPC algorithm to make preliminary predictions of the motion state and corresponding motion control inputs of the wheeled robot for the next N steps based on the dynamic model; Analysis module 13 is used to: use the PPO algorithm to analyze and determine the wheel shape and corresponding wheel control input of the wheeled robot in the future N steps based on the motion state predicted by the wheel control model and MPC algorithm; The re-prediction module 14 is used to: control the MPC algorithm to re-optimize and predict the motion state and corresponding motion control input of the wheeled robot in the next N steps based on the wheel shape predicted by the PPO algorithm; The execution module 15 is used to: control the wheeled robot to execute the predicted wheel control input and the latest motion control input, and optimize the PPO algorithm based on the execution results; Control module 16 is used to: return to the step of performing the preliminary prediction using the MPC algorithm until the convergence condition is met, and then use the MPC algorithm and PPO algorithm to control the wheeled robot.

[0075] This application provides a control device for a wheeled robot, the execution module of which may include: The optimization unit is used to: acquire balance maintenance data, trajectory tracking data, energy consumption data, and wheel shape adjustment data during the execution process of the wheeled robot based on the execution results; calculate the reward function value for evaluating the performance of the PPO algorithm based on the balance maintenance data, trajectory tracking data, energy consumption data, and wheel shape adjustment data; and optimize the PPO algorithm based on the reward function value.

[0076] This application provides a control device for a wheeled robot, the execution module of which may include: The execution unit is used to control the wheeled robot to execute the predicted wheel control input and the latest motion control input in a preset simulation environment; wherein the preset simulation environment is a pre-constructed environment containing different terrains and load conditions.

[0077] The control device for a wheeled robot provided in this application embodiment may further include: The detection module is used to: detect the motion control input output by the MPC algorithm and the wheel control input output by the PPO algorithm in real time when controlling the wheeled robot using the MPC algorithm and the PPO algorithm. If at least one of the control inputs exceeds the corresponding execution range, the at least one control input is adjusted to the corresponding execution range.

[0078] The control device for a wheeled robot provided in this application embodiment may further include: The exception handling module is used to: when controlling the wheeled robot using the MPC algorithm and PPO algorithm, if the motion state and / or wheel shape of the wheeled robot are abnormal during the execution of the corresponding motion control input and / or wheel control input, control the wheeled robot to stop the action immediately.

[0079] The control device for a wheeled robot provided in this application embodiment may further include: The data processing module is used to: collect the current motion state of the wheeled robot; if the collected motion state data is lost or abnormal, it uses the historical data difference method or the state estimation method to compensate for the lost or abnormal data, so that the PPO algorithm can comprehensively analyze the motion state predicted by the MPC algorithm based on the current motion state of the wheeled robot and determine the wheel shape and corresponding wheel control input of the wheeled robot for the next N steps.

[0080] This application provides a control device for a wheeled robot. The wheels of the wheeled robot are flexible wheels with variable radial stiffness and diameter. The wheel shape includes the radial stiffness and diameter of the flexible wheel.

[0081] This application also provides a control device for a wheeled robot. The device includes a memory and a processor. The program stored in the memory can implement the steps of any of the above-mentioned wheeled robot control methods when the processor runs.

[0082] This application also provides a computer-readable storage medium, on which a program stored is executed by a processor to implement the steps of any of the above-mentioned wheeled robot control methods.

[0083] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A control method of a wheeled robot, characterized by, include: A dynamic model and a wheel control model of a wheeled robot are established. The dynamic model describes the relationship between the motion state of the wheeled robot and the motion control input, and the wheel control model describes the relationship between the wheel shape of the wheeled robot and the wheel control input. The wheels of the wheeled robot are flexible wheels with variable radial stiffness and diameter, and the wheel shape includes the radial stiffness and diameter of the flexible wheel. Using the MPC algorithm, the motion state and corresponding motion control input of the wheeled robot in the next N steps are initially predicted based on the dynamic model. Using the PPO algorithm, the wheel shape and corresponding wheel control input of the wheeled robot in the next N steps are determined based on the motion state analysis obtained from the wheel control model and the MPC algorithm prediction. Using the MPC algorithm, the wheel shape predicted by the PPO algorithm is re-optimized and the motion state and corresponding motion control input of the wheeled robot in the next N steps are predicted; The wheeled robot is controlled to execute the predicted wheel control inputs and the latest motion control inputs, and the PPO algorithm is optimized based on the execution results. Return to the initial prediction step using the MPC algorithm until the convergence condition is met, and then use the MPC algorithm and PPO algorithm to control the wheeled robot.

2. The control method of the wheeled robot according to claim 1, characterized by, Optimize the PPO algorithm based on the execution results, including: Based on the execution results, obtain the balance maintenance data, trajectory tracking data, energy consumption data, and wheel shape adjustment data of the wheeled robot during the execution process; Based on the balance maintenance data, trajectory tracking data, energy consumption data, and wheel shape adjustment data, a reward function value is calculated to evaluate the performance of the PPO algorithm. The PPO algorithm is optimized based on the reward function value.

3. The control method of the wheeled robot according to claim 2, characterized in that, Controlling the wheeled robot to execute the predicted wheel control inputs and the latest motion control inputs includes: The wheeled robot is controlled to execute predicted wheel control inputs and the latest motion control inputs in a preset simulation environment; wherein the preset simulation environment is a pre-constructed environment containing different terrains and load conditions.

4. The control method for a wheeled robot according to claim 3, characterized in that, Also includes: When controlling the wheeled robot using the MPC algorithm and PPO algorithm, the motion control input output by the MPC algorithm and the wheel control input output by the PPO algorithm are detected in real time. If at least one of the control inputs exceeds the corresponding execution range, the at least one control input is adjusted to the corresponding execution range.

5. The control method for a wheeled robot according to claim 4, characterized in that, Also includes: When using the MPC algorithm and PPO algorithm to control the wheeled robot, if the wheeled robot's motion state and / or wheel shape become abnormal during the execution of corresponding motion control inputs and / or wheel control inputs, the wheeled robot is controlled to stop an emergency action.

6. The control method for a wheeled robot according to claim 5, characterized in that, Also includes: The current motion state of the wheeled robot is collected. If the collected motion state data is lost or abnormal, the lost or abnormal data is compensated by the historical data difference method or the state estimation method. The PPO algorithm is then used to comprehensively analyze the current motion state of the wheeled robot and the motion state predicted by the MPC algorithm to determine the wheel shape and corresponding wheel control input of the wheeled robot for the next N steps.

7. A control device for a wheeled robot, characterized in that, include: A module is established for: establishing a dynamic model and a wheel control model of a wheeled robot; wherein, the dynamic model describes the relationship between the motion state of the wheeled robot and the motion control input, the wheel control model describes the relationship between the wheel shape of the wheeled robot and the wheel control input, the wheels of the wheeled robot are flexible wheels with variable radial stiffness and diameter, and the wheel shape includes the radial stiffness and diameter of the flexible wheel; The initial prediction module is used to: use the MPC algorithm to make an initial prediction of the motion state and corresponding motion control input of the wheeled robot for the next N steps based on the dynamic model; The analysis module is used to: use the PPO algorithm to analyze and determine the wheel shape and corresponding wheel control input of the wheeled robot for the next N steps based on the motion state predicted by the wheel control model and the MPC algorithm; The re-prediction module is used to: use the MPC algorithm to re-optimize the wheel shape predicted by the PPO algorithm and predict the motion state and corresponding motion control input of the wheeled robot for the next N steps; The execution module is used to: control the wheeled robot to execute the predicted wheel control input and the latest motion control input, and optimize the PPO algorithm based on the execution results; The control module is used to: return to the step of performing the preliminary prediction using the MPC algorithm until the convergence condition is met, and then use the MPC algorithm and PPO algorithm to control the wheeled robot.

8. A control device for a wheeled robot, characterized in that, The device includes a memory and a processor, wherein the program stored in the memory, when executed by the processor, can implement the steps of the control method for the wheeled robot as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, When the program stored on the computer-readable storage medium is executed by a processor, it implements the steps of the control method for the wheeled robot as described in any one of claims 1 to 6.

Images