A Running Control Method and System for a Two-Wheeled Legged Amphibious Robot

Abstract

The present invention discloses a method and system for operating and controlling a two-wheeled legged land-air amphibious robot, which relates to the technical field of robot control and optimization. The method includes: inputting the robot attitude and load data into a nonlinear dynamics model, and dynamically calculating the optimal torque ratio of the robot's rotors and legs by using the gradient descent method; based on the optimal torque ratio, outputting control parameters through a genetic algorithm; based on the control parameters, constructing and training a DQN model to learn the optimal energy distribution scheme, and outputting an energy distribution rule and a power setting value; converting the energy distribution rule and the power setting value into control signals, and transmitting them to each drive unit through PWM signals to complete balance adjustment and energy distribution, so as to control the operation of the two-wheeled legged land-air amphibious robot. By using the Lagrangian mechanics method to establish a nonlinear dynamics model, the accurate description of the complex dynamic characteristics of the two-wheeled legged land-air amphibious robot is realized, and the accuracy and stability of the operation control are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of robot control and optimization, and particularly to a running control method and system for a two-wheeled legged land-air amphibious robot. Background Art

[0002] With the rapid development of robot technology, two-wheeled legged land-air amphibious robots have gradually become a hot research field in academia and industry due to their unique advantages in complex terrain adaptability and versatility. Such robots combine the high-efficiency mobility of traditional wheeled robots and the flexibility of legged robots, and are also equipped with a rotor module to achieve aerial flight functions, enabling seamless switching between the ground, air, and transitional states. However, to achieve stable operation and efficient energy management, a series of complex dynamics modeling and control problems need to be solved. Current research mainly focuses on robot attitude control, energy distribution optimization, and multi-modal motion coordination. For example, in terms of attitude control, existing research has improved the stability of robots in different motion modes by introducing nonlinear control theory; while in terms of energy management, machine learning algorithms are used to predict and optimize the energy consumption of each drive unit of the robot. Although these technologies have made significant progress, there are still many deficiencies in the existing methods.

[0003] The main limitations of the existing technology are reflected in two aspects: one is the insufficient accuracy of the dynamics model, which is difficult to accurately describe the nonlinear dynamic characteristics of two-wheeled legged land-air amphibious robots in complex environments. Traditional dynamics modeling methods are usually based on simplified assumptions, ignoring the coupling effects during the switching of different motion modes of the robot, resulting in a large deviation between the model prediction results and the actual operation. The other is the lack of adaptability of the energy distribution strategy, which cannot dynamically adjust the optimal torque ratio between the rotor and the legs according to real-time load changes. This not only reduces the overall efficiency of the system but also may cause the robot to become unstable under certain working conditions. Therefore, how to construct a more accurate dynamics model and design an efficient energy distribution scheme has become a key technical problem in this field. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a running control method for a two-wheeled legged land-air amphibious robot to solve the problems of insufficient accuracy of the dynamics model and lack of adaptability of the energy distribution strategy in the existing technology.

[0006] To solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a method for controlling the operation of a two-wheeled legged land-air amphibious robot, which includes: real-time collecting robot attitude and load data through sensors; establishing a nonlinear dynamic model using the Lagrangian mechanics method; inputting the robot attitude and load data into the nonlinear dynamic model, and dynamically calculating the optimal torque ratio of the robot's rotor and legs using the gradient descent method; based on the optimal torque ratio, outputting control parameters through a genetic algorithm; based on the control parameters, constructing and training a DQN model to learn the optimal energy distribution scheme, and outputting an energy distribution rule and a power setting value; converting the energy distribution rule and the power setting value into control signals, and transmitting them to each drive unit through PWM signals to complete balance adjustment and energy distribution, and controlling the operation of the two-wheeled legged land-air amphibious robot.

[0008] As a preferred embodiment of the method for controlling the operation of the two-wheeled legged land-air amphibious robot according to the present invention, wherein: the robot attitude data includes angle, angular velocity, and acceleration;

[0009] The robot load data includes total load mass, load mass distribution, and center of gravity position.

[0010] As a preferred embodiment of the method for controlling the operation of the two-wheeled legged land-air amphibious robot according to the present invention, wherein: the specific steps of establishing the nonlinear dynamic model using the Lagrangian mechanics method are as follows.

[0011] Define the basic parameters of the robot through the data collected by the sensors;

[0012] Construct the kinetic energy function and potential energy function of the robot according to the basic parameters;

[0013] Combine the kinetic energy function and the potential energy function to construct the Lagrangian function;

[0014] Obtain the nonlinear dynamic model by solving the Lagrangian function.

[0015] As a preferred embodiment of the method for controlling the operation of the two-wheeled legged land-air amphibious robot according to the present invention, wherein: the specific steps of inputting the robot attitude and load data into the nonlinear dynamic model and dynamically calculating the optimal torque ratio of the robot's rotor and legs using the gradient descent method are as follows.

[0016] Substitute the attitude data and the load data as variables into the established nonlinear dynamic model;

[0017] Construct an optimal torque ratio objective function according to the nonlinear dynamic model;

[0018] Initialize the ratio parameters, and set the torque contribution ratios of the rotor and the legs to be the same;

[0019] Use the gradient descent method to iteratively update the torque ratio of the rotor and the legs.

[0020] Compare the change in the optimal torque ratio objective function between two consecutive iterations;

[0021] Output the optimal torque ratio that satisfies the stable state of the rotor and legs.

[0022] As a preferred solution of the operation control method for the two-wheeled legged land-air amphibious robot described in the present invention, wherein: based on the optimal torque ratio, control parameters are output through a genetic algorithm, and the specific steps are as follows.

[0023] Use the optimal torque ratio obtained by the gradient descent method as the constraint condition for the control parameters;

[0024] Initialize the population of the genetic algorithm according to the actual operating range of the robot;

[0025] Use the reciprocal method to calculate the fitness value of each control parameter vector in the population;

[0026] Adopt the roulette wheel selection method to select individuals from the population for crossover and mutation operations;

[0027] Perform single-point crossover on the selected individuals to generate a new generation of candidate solutions;

[0028] Randomly adjust some control parameter values to increase the population diversity;

[0029] Replace the low-fitness individuals through the elite retention strategy method for population iterative evolution;

[0030] Compare the population fitness differences to select the optimal torque ratio in the stable state;

[0031] Extract the optimal individuals that meet the termination conditions as control parameters.

[0032] As a preferred solution of the operation control method for the two-wheeled legged land-air amphibious robot described in the present invention, wherein: based on the control parameters, a DQN model is constructed and trained to learn the optimal energy allocation scheme, and the energy allocation rule and power setting value are output, and the specific steps are as follows.

[0033] Define the state space through the attitude and load data collected by the sensor, define the action space for the robot's rotor and leg actions, and then define the reward function for the robot's energy consumption and motion performance;

[0034] Construct a DQN model through the defined state space, action space, and reward function, including an input layer, multiple hidden layers, and an output layer;

[0035] Randomly initialize the weight parameters of the DQN model, and set an experience replay buffer to store training samples and perform subsequent updates;

[0036] Run the robot in the simulation session, select actions through the greedy policy, and record the sample data of the state, action, reward, and next state;

[0037] Extract samples from the experience replay buffer and update the DQN model parameters by minimizing the loss function;

[0038] Synchronize the main network parameters to the target network every fixed number of steps to maintain the stability of the target network and improve the training effect;

[0039] Evaluate the optimal actions of the rotor and legs according to the change amount of the DQN model parameters. If no output is obtained, continue iterative training until the termination criterion is reached;

[0040] Use the DQN model trained based on the control parameters to output the optimal actions of the rotor and legs, and extract the optimal energy distribution rule and power setting value.

[0041] As a preferred solution of the running control method for the two-wheeled legged land-air amphibious robot described in the present invention, wherein: converting the energy distribution rule and power setting value into a control signal, and transmitting it to each drive unit through a PWM signal to complete balance adjustment and energy distribution, and controlling the running of the two-wheeled legged land-air amphibious robot. The specific steps are as follows.

[0042] Combine the energy distribution rule and power setting value with the nonlinear dynamics model to calculate the specific control parameters of the torque contribution ratio of the rotor and legs;

[0043] Convert the control parameters of the torque contribution ratio of the rotor and legs into a PWM signal applicable to the drive unit by using a smooth cosine function;

[0044] Adopt a fuzzy logic controller, combine the angular velocity characteristics of the rotor and legs collected by the sensor in real time, and reasonably distribute the PWM signal to each drive unit to control the running of the two-wheeled legged land-air amphibious robot.

[0045] Second aspect, the present invention provides an operation control system for a two-wheeled legged amphibious robot, including a data acquisition module, a dynamics modeling module, a torque optimization module, a parameter generation module, an intelligent learning module, and an execution control module; the data acquisition module is used to collect the robot attitude and load data in real time through sensors; the dynamics modeling module is used to establish a nonlinear dynamics model using Lagrangian mechanics method; the torque optimization module is used to input the robot attitude and load data into the nonlinear dynamics model, and dynamically calculate the optimal torque ratio of the robot's rotor and legs using the gradient descent method; the parameter generation module is used to output control parameters based on the optimal torque ratio through the genetic algorithm; the intelligent learning module is used to construct and train a DQN model based on the control parameters to learn the optimal energy distribution scheme, and output the energy distribution rule and power setting value; the execution control module is used to convert the energy distribution rule and power setting value into control signals, and transmit them to each drive unit through PWM signals to complete balance adjustment and energy distribution, and control the operation of the two-wheeled legged amphibious robot.

[0046] Third aspect, the present invention provides a computer device, including a memory and a processor, where the memory stores a computer program, and: when the computer program is executed by the processor, it realizes any step of the operation control method for the two-wheeled legged amphibious robot as described in the first aspect of the present invention.

[0047] Fourth aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and: when the computer program is executed by the processor, it realizes any step of the operation control method for the two-wheeled legged amphibious robot as described in the first aspect of the present invention.

[0048] The beneficial effects of the present invention are as follows: By using the Lagrangian mechanics method to establish a nonlinear dynamics model, the accurate description of the complex dynamic characteristics of the two-wheeled legged amphibious robot is realized, and the accuracy and stability of operation control are improved; at the same time, based on the control parameters, a DQN model is constructed and trained to learn the optimal energy distribution scheme, realizing the intelligent optimization of the energy consumption of the rotor and legs, significantly reducing the energy consumption while ensuring the motion performance, and extending the endurance time. The combination of the two not only enhances the adaptability of the robot in complex environments, but also greatly improves its operation efficiency and sustainability, providing an innovative solution for the efficient and intelligent control of amphibious robots. Description of the Drawings

[0049] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0050] Figure 1 It is a flowchart of the operation control method for the two-wheeled legged land-air amphibious robot in Embodiment 1.

[0051] Figure 2 It is a schematic diagram of the operation control system for the two-wheeled legged land-air amphibious robot in Embodiment 1. Specific Embodiments

[0052] To make the above objects, features, and advantages of the present invention more obvious and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings of the specification.

[0053] In the following description, many specific details are set forth to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar generalizations without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0054] Secondly, the so-called "one embodiment" or "embodiment" herein refers to specific features, structures, or characteristics that can be included in at least one implementation manner of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a separate or selectively exclusive embodiment from other embodiments.

[0055] Embodiment 1, referring to Figure 1 and Figure 2 , is the first embodiment of the present invention. This embodiment provides an operation control method for a two-wheeled legged land-air amphibious robot, including the following steps:

[0056] S1. Real-time collect the robot's attitude and load data through sensors.

[0057] S1.1. The robot's attitude data includes angles, angular velocities, and accelerations.

[0058] It should be noted that in the operation control method of the two-wheeled legged land-air amphibious robot, the sensor attitude data collected in the first step includes the pitch angle, roll angle, and yaw angle of the fuselage and their change rates. These data are obtained through the accelerometer and gyroscope in the inertial measurement unit and processed by a filtering algorithm to improve the accuracy. Among them, the pitch angle represents the rotation angle of the fuselage around the transverse axis, the roll angle represents the rotation angle around the longitudinal axis, and the yaw angle represents the rotation angle around the vertical axis. At the same time, the accelerometer also provides linear acceleration information to assist in attitude estimation. Finally, these attitude data will be important inputs for solving the optimal control parameters in the non-linear dynamics model in the subsequent steps.

[0059] S1.2. The robot's load data includes the total load mass, load mass distribution, and center of gravity position.

[0060] It should be noted that in the operation control method of the two-wheel legged land-air amphibious robot, the load data includes the total load mass, load mass distribution, center of gravity position, dynamic load change, and load type identification. Among them, the total load mass is calculated by integrating the pressure sensor. The mass distribution and center of gravity position are used to adjust the robot's posture to ensure stability. The dynamic load change monitors the data change rate of the pressure sensor in real time to adapt to the load fluctuations during the handling process. The load type identification combines the accelerometer and machine learning algorithms to analyze the load characteristics. These data jointly provide key inputs for subsequent non-linear dynamics modeling and control parameter optimization, ensuring the efficient operation of the robot under different working conditions.

[0061] S2. Establish a non-linear dynamics model using the Lagrangian mechanics method.

[0062] S2.1. Define the basic parameters of the robot through the data collected by the sensors.

[0063] It should be noted that defining the basic parameters of the robot through the data collected by the sensors specifically means calculating and determining the key attributes of the robot, such as the inertial parameters, center of mass position, motion range limits, and maximum output torque capabilities of each joint, based on the information of angles, angular velocities, accelerations, total load mass, load mass distribution, and center of gravity position obtained by the sensors in real time, combined with the structural characteristics of the robot itself.

[0064] S2.2. Construct the kinetic energy function and potential energy function of the robot according to the basic parameters.

[0065] It should be noted that according to the basic parameters of the robot, including information such as mass, length, center of gravity position, and motion state, combined with the principles of classical mechanics, the sum of the products of the squares of the motion velocities of each component of the robot and their masses is expressed as the kinetic energy function, and at the same time, the sum of the products of the heights of each component of the robot under the action of gravity and their masses is expressed as the potential energy function, so as to comprehensively reflect the energy distribution characteristics of the robot in different postures and motion states.

[0066] S2.3. Combine the kinetic energy function and potential energy function to construct the Lagrangian function.

[0067] It should be noted that by combining the kinetic energy function and potential energy function of the robot, specifically, the Lagrangian function is constructed by subtracting the potential energy function from the kinetic energy function. This process fully considers the motion characteristics of the robot under different postures and load conditions. Among them, the kinetic energy function reflects the total kinetic energy of all moving components in the system, while the potential energy function represents the total potential energy generated by the system due to the action of gravity or other conservative forces. The Lagrangian function formed by combining the two can comprehensively describe the dynamic behavior of the robot.

[0068] S2.4. Obtain the nonlinear dynamic model by solving the Lagrangian function.

[0069] It should be noted that specifically, based on the kinetic energy and potential energy expressions of the robot system, combined with the derivative form of the Lagrangian equation with respect to time, expand and simplify, take the angles, angular velocities and their change rates of each joint in the system as variables, comprehensively consider the influence of external torques and constraint conditions, and finally derive a dynamic model that describes the complex nonlinear relationship between the robot's motion state and control input.

[0070] S3. Input the robot attitude and load data into the nonlinear dynamic model, and dynamically calculate the optimal torque ratio of the robot's rotors and legs using the gradient descent method.

[0071] S3.1. Substitute the attitude data and load data as variables into the established nonlinear dynamic model.

[0072] It should be noted that the current attitude data of the robot such as angles, angular velocities, accelerations, and load data such as total load mass, mass distribution and center of gravity position are collected in real time by sensors and transmitted to the calculation module. In the nonlinear dynamic model, these data are substituted as key variables to describe the dynamic characteristics of the robot.

[0073] S3.2. Construct the optimal torque ratio objective function according to the nonlinear dynamic model.

[0074] It should be noted that to construct the optimal torque ratio objective function according to the nonlinear dynamic model, specifically, by analyzing the dynamic characteristics of the robot under different attitudes and load conditions, quantify the influence of the rotor and leg torques on the system balance stability as a comprehensive performance index, and combine the principles of minimizing energy consumption and optimizing motion performance, define a multi-objective optimization problem with the torque ratio as the variable, and transform it into a single objective function form to obtain the optimal torque ratio objective function.

[0075] S3.3. Initialize the ratio parameters and set the torque contribution ratios of the rotors and legs to be the same.

[0076] It should be noted that initializing the ratio parameters and setting the torque contribution ratios of the rotors and legs to be the same is specifically achieved by assuming that the contributions of the rotors and legs to the overall balance are equal when the control system starts. Based on this premise, set the initial torque values of both to the same reference level to ensure that they are in an equilibrium state before the algorithm optimization starts, thus providing a fair and stable starting point for subsequent dynamic adjustment and avoiding problems such as decreased control performance or system instability caused by initial condition deviations.

[0077] S3.4. Use the gradient descent method to iteratively update the torque ratio of the rotors and legs.

[0078] It should be noted that the gradient descent method is used to iteratively update the ratio of the rotor torque to the leg torque. Specifically, the gradient value of the optimal torque ratio objective function with respect to the current torque ratio is calculated, and the ratio parameters of the rotor and leg torques are adjusted according to this gradient direction. In each iteration, the value of the objective function is gradually decreased according to a preset learning rate, so that the torque ratio gradually approaches the optimal solution. At the same time, to ensure the convergence stability, the algorithm dynamically monitors the change amount of the objective function after each iteration. If the change amount is less than the set threshold, it is considered that the stable state has been reached and the iteration is stopped. This process makes full use of the optimization characteristics of the gradient descent method, while ensuring the calculation efficiency, realizes the precise adjustment of the ratio of the rotor torque to the leg torque, and finally updates the ratio of the rotor torque to the leg torque that meets the balance requirements of the robot under different postures and load conditions.

[0079] S3.5. Compare the change amounts of the optimal torque ratio objective function in two consecutive iterations.

[0080] It should be noted that comparing the change amounts of the optimal torque ratio objective function in two consecutive iterations is specifically achieved by calculating the difference between the current objective function value and the objective function value of the previous iteration after each iteration, and using this difference as an index to measure the convergence of the optimization process. To ensure the accuracy of the comparison, the system stores the objective function values of the two iterations in temporary variables and calculates the difference value after each update. If this change amount is less than the preset threshold, it indicates that the objective function value has tended to be stable and the optimization process is close to the optimal solution, and the iteration can be stopped; otherwise, if the change amount is large, continue to execute the gradient descent method to further adjust the ratio of the rotor torque to the leg torque until the convergence condition is met.

[0081] S3.6. Output the optimal torque ratio that satisfies the stable state of the rotor and the leg.

[0082] It should be noted that outputting the optimal torque ratio that satisfies the stable state of the rotor and the leg is specifically achieved by continuously monitoring the dynamic response of the robot under different torque ratios. When the torque distribution of the rotor and the leg makes the overall posture deviation, angular velocity change and center of gravity stability of the robot all within the preset threshold range, it is determined that the current torque ratio has reached the stable state. This process depends on the real-time calculation of the nonlinear dynamics model, combined with the optimization results of the gradient descent method, to ensure that the torque ratio can not only meet the requirement of minimizing energy consumption, but also ensure the optimal motion performance, and use the torque ratio that meets the above stable conditions as the output result.

[0083] S4. Based on the optimal torque ratio, use the genetic algorithm to output the control parameters.

[0084] S4.1. Take the optimal torque ratio obtained by the gradient descent method as the constraint condition of the control parameters;

[0085] It should be noted that the optimal torque ratio is restricted within a certain range to ensure that the control parameter vector generated by the genetic algorithm meets the actual operation requirements of the robot. When initializing the population, the torque ratio-related parameters of all individuals need to satisfy this constraint to avoid generating unrealistic solutions. During the subsequent iterative process, any individual that violates this constraint condition will be eliminated or corrected, thus ensuring that the population evolution direction always revolves around the optimal torque ratio.

[0086] S4.2. Initialize the population of the genetic algorithm according to the actual operation range of the robot.

[0087] It should be noted that initializing the population of the genetic algorithm according to the actual operation range of the robot is specifically achieved by analyzing parameters such as the torque output range of the rotors and legs, the angle change interval, and the load capacity of the robot in different working states to determine the feasible value range of the control parameters. On this basis, a group of initial solutions that meet these constraint conditions are randomly generated as population individuals to ensure that each individual represents a possible control strategy. At the same time, to improve the population diversity and avoid falling into local optima, a certain amount of random perturbation is introduced during the initialization process to make the individual distribution evenly cover the entire search space. This process provides a reasonable starting point for the subsequent iterative optimization of the genetic algorithm and helps to more efficiently explore the optimal control parameter combination.

[0088] S4.3. Calculate the fitness value of each control parameter vector in the population using the reciprocal method.

[0089] It should be noted that calculating the fitness value of each control parameter vector in the population using the reciprocal method is specifically achieved by substituting the control parameter vector into a pre-defined objective function to evaluate the magnitude of its corresponding objective value, and then taking the reciprocal value of this objective value as the fitness value. The core of this method is that the smaller the objective value, the closer the control parameter is to the optimal solution, and the reciprocal method can transform this relationship into a higher fitness value, thus facilitating the genetic algorithm to select better individuals. In actual operation, first calculate the objective function for each control parameter vector in the population to obtain its performance index, and then take the reciprocal of these indices to generate the fitness value.

[0090] S4.4. Adopt the roulette wheel selection method to select individuals from the population for crossover and mutation operations.

[0091] It's important to note that roulette wheel selection is a proportional selection method based on individual fitness. In a genetic algorithm, this method calculates the probability of selection for each individual, giving individuals with higher fitness a greater chance of participating in crossover and mutation operations while also ensuring that individuals with lower fitness are not completely ignored. This method, like a roulette wheel, allows individuals with higher fitness to occupy a larger area, increasing their likelihood of selection. This method retains high-quality individuals while maintaining population diversity, helping the algorithm better search the solution space and find optimal solutions. This approach mimics the process of natural selection, achieving survival of the fittest while balancing exploration and exploitation.

[0092] S4.5. Perform single-point crossover on the selected individuals to generate a new generation of candidate solutions.

[0093] It's important to note that single-point crossover is a genetic recombination operation within a genetic algorithm. This involves exchanging gene segments at a randomly determined position between two selected individuals to generate a new generation of candidate solutions. The process involves randomly pairing individuals from the population. Then, at a randomly selected crossover point in their genetic coding sequences, the gene segments of the two individuals are swapped, resulting in two new offspring individuals. This operation combines the advantageous characteristics of the parent individuals, increasing population diversity while preserving the potential for optimal gene combinations. This generates a new generation of candidate solutions that inherit some of the characteristics of the parent while also potentially possessing new optimization potential.

[0094] S4.6. Randomly adjust some control parameter values to increase population diversity.

[0095] It is important to note that by introducing a random perturbation mechanism into the genetic algorithm's evolutionary process, the control parameter values of some individuals in the population are randomly adjusted within a small range to increase population diversity. Specifically, at each iteration, individuals are randomly selected with a certain probability and their parameter values are subjected to small random changes. These changes can be random numbers with a uniform or normal distribution. This random adjustment helps to break the phenomenon of premature convergence that may occur in the population and prevent the algorithm from being trapped in a local optimal solution. At the same time, by introducing new parameter combinations, fresh "genes" are injected into the population, enhancing the algorithm's exploration capabilities, thereby increasing the possibility of finding a global optimal solution and ensuring that the genetic algorithm has better adaptability and robustness in complex optimization problems.

[0096] S4.7. Replace low-fitness individuals through the elite retention strategy to perform iterative population evolution.

[0097] It should be noted that the population iterative evolution is achieved by replacing low-fitness individuals through the elitist retention strategy. Specifically, during the evolution process of each generation of the population, first, the high-quality individuals with the highest fitness in the current population are identified and directly retained in the next generation of the population. Subsequently, the remaining low-fitness individuals are eliminated and replaced with new individuals generated through crossover and mutation operations to ensure the continuous improvement of the overall quality of the population.

[0098] S4.8. Compare the fitness differences of the population to select the optimal torque ratio in the stable state.

[0099] It should be noted that the optimal torque ratio in the stable state is judged by calculating the fitness differences of individuals in the population. Specifically, the distribution of the fitness values of individuals within the population is analyzed. When the gap between the maximum fitness value and the average fitness value in the population for multiple consecutive generations tends to be stable or is less than the set threshold, it indicates that the algorithm has converged to an approximate optimal solution. In addition, the change trend of the population fitness needs to be monitored. If the change rate continuously drops below a certain preset standard, it is further confirmed that the system has entered the stable state, and the corresponding rotor and leg torque ratio at this time is the optimal ratio.

[0100] S4.9. Extract the optimal individuals that meet the termination conditions as control parameters.

[0101] It should be noted that by comparing the fitness values of each individual in the population, the individual with the highest fitness is selected. When the termination conditions such as the number of iterations or the fitness change tending to be stable are met, the parameter combination corresponding to this individual is extracted as the final control parameter.

[0102] S5. Based on the control parameters, construct and train a DQN model to learn the optimal energy allocation scheme, and output the energy allocation rule and the power setting value.

[0103] S5.1. Define the state space through the attitude and load data collected by the sensors, define the action space for the robot's rotor and leg movements, and then define the reward function based on the energy consumption and motion performance of the robot.

[0104] It should be noted that the attitude data (such as angles, angular velocities, accelerations) and load data (such as total mass, mass distribution, and center of gravity position) of the robot are collected in real time through sensors to comprehensively describe the current operating state of the robot, thereby defining its state space. At the same time, the action ranges of the robot's rotor and legs and their possible output torques are used as the action space to ensure that all feasible operation options are covered. On this basis, a reward function is formulated in combination with the energy consumption situation and motion performance of the robot during operation to encourage the system to achieve stable and flexible motion control while ensuring efficient energy utilization.

[0105] S5.2. Construct a DQN model through the defined state space, action space, and reward function, which includes an input layer, multiple hidden layers, and an output layer.

[0106] It should be noted that based on the defined state space, action space, and reward function, a deep neural network structure is designed as the DQN model. First, map the state space to the dimension of the input layer to ensure that all possible state information can be received; then, perform non-linear feature extraction and complex relationship modeling through multiple hidden layers to capture the deep associations between states and actions; finally, set the dimension of the output layer to correspond to the size of the action space and output the expected value of each action, thus realizing the intelligent decision-making process from the input state to the optimal action selection.

[0107] S5.3. Randomly initialize the weight parameters of the DQN model, and set up an experience replay buffer to store training samples and perform subsequent updates.

[0108] It should be noted that when constructing the deep Q-network model, first randomly set the weight values of the connections between neurons in the network as initial parameters to break symmetry and ensure that the model can be effectively trained. Subsequently, create an experience replay buffer to store the training samples generated by the robot's interaction in the simulated environment, including information such as the current state, the actions taken, the rewards obtained, and the next state. This buffer not only supports subsequent extraction of batch samples from historical data for model updates but also avoids interference from related sample sequences to the training process, thereby improving the stability and convergence speed of the algorithm.

[0109] S5.4. Run the robot in the simulation session, select actions through the greedy strategy, and record the sample data of the state, action, reward, and next state.

[0110] It should be noted that when running the robot in the simulated environment, first initialize its initial state, and then select the action with the highest expected reward in the current state according to the greedy strategy to execute. Then record the actual results generated by this action, including the current state, the selected action, the obtained reward value, and the next state data transferred thereby, forming a complete sample record for subsequent training. This process is repeated to ensure that enough diverse data is collected to support model learning and parameter adjustment.

[0111] S5.5. Extract samples from the experience replay buffer and update the DQN model parameters by minimizing the loss function.

[0112] It should be noted that when sampling from the experience replay buffer, a batch of previously stored combinations of states, actions, rewards, and next states are randomly selected. Then, the gap between the current network's predicted value and the target network's true value is calculated using these samples and defined as the loss function. The model parameters are adjusted through an optimization algorithm to gradually reduce the value of this loss function, making the predicted value gradually approach the true value. This process is iterated repeatedly until the DQN model parameters are updated, improving the accuracy and stability of its optimal action selection.

[0113] S5.6. Synchronize the main network parameters to the target network every fixed number of steps to maintain the stability of the target network and improve the training effect.

[0114] It should be noted that during the deep reinforcement learning training process, by setting a fixed step interval, whenever this interval is reached, the updated parameter values in the main network are completely copied to the target network. This synchronization mechanism ensures that the target network always maintains a relatively stable state, avoiding excessive fluctuations in the target value caused by frequent updates of the main network parameters, thereby enhancing the stability of the training process. At the same time, since the target network provides a more stable reference value, it helps the main network calculate the loss function more accurately and optimize the parameters, ultimately improving the overall training effect.

[0115] S5.7. Evaluate the optimal actions of the rotor and legs according to the change amount of the DQN model parameters. If no output is generated, continue the iterative training until the termination criterion is reached.

[0116] It should be noted that during the training process, the convergence state of the model is evaluated by calculating the continuous change amount of the DQN model parameters. If the change amount of the parameters is less than the preset threshold, the model is considered stable, and the corresponding optimal actions of the rotor and legs at this time are output; if the change amount is still large, continue the iterative training, continuously update the model parameters until the termination condition is met.

[0117] S5.8. Use the DQN model trained based on the control parameters to output the optimal actions of the rotor and legs, and extract the optimal energy distribution rule and power setting value.

[0118] It should be noted that by using the trained deep reinforcement learning model, the current state information of the robot is input into the DQN model. Based on the learned experience and optimized strategy, the model outputs the optimal actions of the rotor and legs corresponding to this state. This action combination not only considers the energy consumption efficiency of the robot but also takes into account the stability of the motion performance. Subsequently, according to the results of the optimal actions, the matching energy distribution rule and power setting value are further extracted.

[0119] S6. Convert the energy distribution rule and the power setting value into a control signal, and transmit it to each drive unit through the PWM signal to complete the balance adjustment and energy distribution, and control the operation of the two-legged land-air amphibious robot.

[0120] S6.1. Combine the energy distribution rule and the power setting value with the non-linear dynamics model to calculate the specific control parameters of the torque contribution ratio of the rotor and the leg.

[0121] It should be noted that by integrating the energy distribution rule and the power setting value into the non-linear dynamics model, the relationship between the energy consumption and the motion performance of the robot under different postures and load conditions is analyzed, and the torque contribution ratio required by the rotor and the leg is calculated using the dynamics equation. The specific process is to determine the power distribution of each component according to the energy distribution rule, and combine the relationship between the torque and the motion state described in the non-linear dynamics model to deduce the torque ratio of the rotor and the leg that meets the current operation requirements.

[0122] S6.2. Use the smooth cosine function to convert the control parameters of the torque contribution ratio of the rotor and the leg into the PWM signal applicable to the drive unit.

[0123] It should be noted that by taking the control parameters of the torque contribution ratio of the rotor and the leg as the input, and using the periodic and continuous characteristics of the smooth cosine function, it is mapped to the PWM signal range applicable to the drive unit. The specific method is to design a conversion formula based on the cosine function according to the magnitude of the torque ratio parameter, combined with the preset signal period and duty cycle range. By adjusting the amplitude and phase of the cosine function, the output PWM signal is ensured to change smoothly and meet the hardware requirements of the drive unit, so as to achieve the accurate conversion from the torque ratio to the actual control signal.

[0124] S6.3. Adopt a fuzzy logic controller, combine the angular velocity characteristics of the rotor and the leg collected by the sensor in real time, and reasonably distribute the PWM signal to each drive unit to control the operation of the two-legged land-air amphibious robot.

[0125] It should be noted that by collecting the angular velocity data of the rotor and the leg in real time through the sensor, the fuzzy logic controller analyzes the current motion state according to these characteristics, and makes refined adjustments in combination with the calculated PWM signal. In specific implementation, first convert the energy distribution rule into the torque contribution ratio parameters of the rotor and the leg, and then use the smooth function to convert it into the PWM signal suitable for the drive unit. The controller judges whether the robot is running stably according to the angular velocity characteristics. If there is a deviation, it dynamically adjusts the intensity and distribution of the PWM signal to ensure the coordinated work of each drive unit. Finally, the reasonably distributed PWM signal is transmitted to the rotor and leg drive units to achieve accurate energy distribution and balance adjustment, so as to effectively control the stable operation of the two-legged land-air amphibious robot.

[0126] This embodiment also provides an operation control system for a two-wheeled legged land-air amphibious robot, including: a data acquisition module, a dynamics modeling module, a torque optimization module, a parameter generation module, an intelligent learning module, and an execution control module;

[0127] The data acquisition module is used to collect the robot's attitude and load data in real time through sensors;

[0128] The dynamics modeling module is used to establish a nonlinear dynamics model using Lagrangian mechanics;

[0129] The torque optimization module is used to input the robot's attitude and load data into the nonlinear dynamics model, and dynamically calculate the optimal torque ratio of the robot's rotors and legs using the gradient descent method;

[0130] The parameter generation module is used to output control parameters based on the optimal torque ratio through a genetic algorithm;

[0131] The intelligent learning module is used to construct and train a DQN model based on the control parameters to learn the optimal energy distribution scheme, and output the energy distribution rule and power setting value;

[0132] The execution control module is used to convert the energy distribution rule and power setting value into control signals, and transmit them to each drive unit through PWM signals to complete balance adjustment and energy distribution, and control the operation of the two-wheeled legged land-air amphibious robot.

[0133] This embodiment also provides a computer device, which is applicable to the situation of the operation control method of a two-wheeled legged land-air amphibious robot, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the operation control method of the two-wheeled legged land-air amphibious robot proposed in the above embodiment.

[0134] This computer device may be a terminal. This computer device includes a processor, a memory, a communication interface, a display screen, and an input device connected through a system bus. Among them, the processor of this computer device is used to provide computing and control capabilities. The memory of this computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of this computer device is used to communicate with an external terminal in a wired or wireless manner. The wireless manner can be achieved through WIFI, a carrier network, NFC (Near Field Communication), or other technologies. The display screen of this computer device may be a liquid crystal display screen or an electronic ink display screen. The input device of this computer device may be a touch layer covered on the display screen, or a button, a trackball, or a touchpad set on the housing of the computer device, or an external keyboard, touchpad, or mouse, etc.

[0135] This embodiment also provides a storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the method for controlling the operation of a two-wheeled legged land-air amphibious robot as proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, abbreviated as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, abbreviated as EEPROM), erasable programmable read-only memory (Erasable Programmable Read Only Memory, abbreviated as EPROM), programmable read-only memory (Programmable Red-Only Memory, abbreviated as PROM), read-only memory (Read-Only Memory, abbreviated as ROM), magnetic memory, flash memory, magnetic disk or optical disc.

[0136] In summary, through the following steps: by using the Lagrangian mechanics method to establish a nonlinear dynamic model, the present invention accurately describes the complex dynamic characteristics of the two-wheeled legged land-air amphibious robot, improving the accuracy and stability of operation control; at the same time, based on the control parameters, a DQN model is constructed and trained to learn the optimal energy allocation scheme, realizing the intelligent optimization of the energy consumption of the rotor and the legs. While ensuring the motion performance, the energy consumption is significantly reduced and the endurance time is extended. The combination of the two not only enhances the adaptability of the robot in complex environments, but also greatly improves its operation efficiency and sustainability, providing an innovative solution for the efficient and intelligent control of land-air amphibious robots.

[0137] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A running control method for a two-wheeled legged land-air amphibious robot, characterized in that: including, real-time collecting the robot's attitude and load data through sensors; establishing a nonlinear dynamic model using Lagrangian mechanics method; inputting the robot's attitude and load data into the nonlinear dynamic model, and dynamically calculating the optimal torque ratio of the robot's rotors and legs using the gradient descent method; based on the optimal torque ratio, outputting control parameters through the genetic algorithm. The specific steps are as follows: taking the optimal torque ratio obtained by the gradient descent method as the constraint condition of the control parameters; initializing the population of the genetic algorithm according to the actual operating range of the robot; calculating the fitness value of each control parameter vector in the population using the reciprocal method; adopting the roulette wheel selection method to select individuals from the population for crossover and mutation operations; performing single-point crossover on the selected individuals to generate a new generation of candidate solutions; randomly adjusting some control parameter values to increase the population diversity; replacing low-fitness individuals through the elitist retention strategy method to perform population iterative evolution; comparing the population fitness differences to select the optimal torque ratio in the stable state; extracting the optimal individual that meets the termination condition as the control parameter; based on the control parameter, constructing and training a DQN model to learn the optimal energy allocation scheme, and outputting the energy allocation rule and power setting value; converting the energy allocation rule and power setting value into a control signal, and transmitting it to each drive unit through a PWM signal to complete balance adjustment and energy allocation, and controlling the operation of the two-wheeled legged amphibious robot.

2. The running control method of the two-wheeled and legged land-air amphibious robot according to claim 1, wherein: The robot attitude data includes angle, angular velocity, and acceleration; The robot load data includes total load mass, load mass distribution, and center of gravity position.

3. The running control method of the two-wheeled and legged land-air amphibious robot according to claim 2, characterized in that: The steps of establishing the nonlinear dynamic model using the Lagrangian mechanics method are as follows: defining the basic parameters of the robot through the data collected by sensors; constructing the kinetic energy function and potential energy function of the robot according to the basic parameters; combining the kinetic energy function and potential energy function to construct the Lagrangian function; obtaining the nonlinear dynamic model by solving the Lagrangian function.

4. The running control method of the two-wheeled and legged land-air amphibious robot according to claim 3, characterized in that: The steps of inputting the robot's attitude and load data into the nonlinear dynamic model and dynamically calculating the optimal torque ratio of the robot's rotors and legs using the gradient descent method are as follows: substituting the attitude data and load data as variables into the established nonlinear dynamic model; constructing the optimal torque ratio objective function according to the nonlinear dynamic model; initializing the ratio parameter and setting the torque contribution ratios of the rotors and legs to be the same; iteratively updating the torque ratios of the rotors and legs using the gradient descent method; comparing the change amounts of the optimal torque ratio objective function in two consecutive iterations; outputting the optimal torque ratio that satisfies the stable state of the rotors and legs.

5. The running control method of the two-wheeled and legged land-air amphibious robot according to claim 1, characterized in that: The steps of constructing and training a DQN model based on the control parameter to learn the optimal energy allocation scheme and output the energy allocation rule and power setting value are as follows: defining the state space through the attitude and load data collected by sensors, defining the action space for the robot's rotors and legs, and then defining the reward function for the robot's energy consumption and motion performance; constructing a DQN model through the defined state space, action space, and reward function, including an input layer, multiple hidden layers, and an output layer; Randomly initialize the weight parameters of the DQN model, and set up an experience replay buffer to store training samples and perform subsequent updates; Run the robot in the simulation session, select actions through the greedy policy, and record the sample data of the state, action, reward, and next state; Extract samples from the experience replay buffer and update the DQN model parameters by minimizing the loss function; Synchronize the main network parameters to the target network every fixed number of steps to maintain the stability of the target network and improve the training effect; Evaluate the best actions of the rotor and legs according to the change amount of the DQN model parameters. If no output is obtained, continue iterative training until the termination criterion is reached; Use the DQN model trained based on the control parameters to output the best actions of the rotor and legs, and extract the best energy distribution rule and power setting value.

6. The running control method of the two-wheeled and legged land-air amphibious robot according to claim 1, wherein: Convert the energy distribution rule and power setting value into control signals, and transmit them to each drive unit through PWM signals to complete balance adjustment and energy distribution, and control the operation of the two-wheeled legged amphibious robot. The specific steps are as follows: Combine the energy distribution rule and power setting value with the nonlinear dynamics model to calculate the specific control parameters of the torque contribution ratio of the rotor and legs; Convert the torque contribution ratio control parameters of the rotor and legs into PWM signals suitable for the drive unit by using the smooth cosine function; Adopt a fuzzy logic controller, combine the angular velocity characteristics of the rotor and legs collected in real time by the sensor, and reasonably distribute the PWM signals to each drive unit to control the operation of the two-wheeled legged amphibious robot.

7. A running control system for a two-wheeled legged land-air amphibious robot, based on the running control method of the two-wheeled legged land-air amphibious robot according to any one of claims 1 to 6, characterized in that: Including a data acquisition module, a dynamics modeling module, a torque optimization module, a parameter generation module, an intelligent learning module, and an execution control module; The data acquisition module is used to collect the robot attitude and load data in real time through sensors; The dynamics modeling module is used to establish a nonlinear dynamics model using the Lagrangian mechanics method; The torque optimization module is used to input the robot attitude and load data into the nonlinear dynamics model, and dynamically calculate the best torque ratio of the robot rotor and legs using the gradient descent method; The parameter generation module is used to output control parameters based on the best torque ratio through the genetic algorithm; The intelligent learning module is used to construct and train a DQN model based on the control parameters to learn the best energy distribution scheme, and output the energy distribution rule and power setting value; The execution control module is used to convert the energy distribution rule and power setting value into control signals, and transmit them to each drive unit through PWM signals to complete balance adjustment and energy distribution, and control the operation of the two-wheeled legged amphibious robot.

8. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the two-wheeled legged amphibious robot operation control method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the two-wheeled legged amphibious robot operation control method according to any one of claims 1 to 6.

Images