Quadruped robot based on sliding detection and anti-falling method thereof

Abstract

The invention provides a quadruped robot based on sliding detection and an anti-falling method thereof.The method comprises the steps that a plastic deformation area of soil is analyzed and processed, and sinking depth distribution of foot ends in the soil is obtained; based on attitude data and sinking depth distribution obtained by a foot end inertial measurement unit, the slippage of the foot end relative to the soil is processed, and a motion state estimation value of the foot end in the soil is obtained; inputting the motion state estimation value into a deep reinforcement learning agent to obtain a control action parameter output by the deep reinforcement learning agent; updating an actor network of the deep reinforcement learning agent based on the control state parameters to obtain a control strategy and optimized control action parameters, and judging whether the foot end sinks or slips in the soil or not; if the foot end sinks or slides in the soil, sliding detection is started. The stability and the energy efficiency of the robot in the complex soil environment are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a quadruped robot based on sliding detection and an anti-fall method thereof. Background Art

[0002] When quadruped robots are used in slippery, muddy, and complex terrain, achieving coordinated gait optimization and sliding control is a key technical challenge. Currently, a common approach is to use reinforcement learning algorithms to adjust gait parameters in real time, continuously optimizing the gait to adapt to different terrains through interaction with the environment. However, relying solely on gait optimization is difficult to completely prevent the robot from sliding in slippery, muddy terrain, and can easily cause the robot to lose balance and fall.

[0003] Slip detection, as a real-time method for monitoring a robot's slipping state, can provide important feedback for gait optimization. However, efficiently performing slip detection and triggering appropriate control strategies without compromising gait optimization presents a pressing technical challenge. Traditional serial processing introduces a time delay between gait optimization and slip detection, impacting the system's real-time performance. Simply processing both processes in parallel can introduce additional system overhead and scheduling complexity.

[0004] Therefore, designing an efficient parallel processing framework to coordinate the operation of the reinforcement learning algorithm and the slip detection module, and achieve seamless integration of gait optimization and slip control, is a challenging technical problem. This requires in-depth research and innovation in multiple aspects, such as algorithm design, system architecture, and resource scheduling, in order to maximize the synergistic effect of reinforcement learning and slip detection while ensuring real-time performance, thereby improving the quadruped robot's environmental adaptability and motion stability in slippery and muddy terrain. Summary of the Invention

[0005] The present invention provides a quadruped robot based on sliding detection and an anti-fall method thereof, which are used to effectively improve the stability and energy efficiency of the robot in a complex soil environment.

[0006] In a first aspect, the present invention provides a quadruped robot based on sliding detection, comprising:

[0007] The six-dimensional force sensor at the foot end is used to obtain the normal and tangential force distributions of the foot end on the soil. The plastic deformation area of ​​the soil is analyzed and processed based on the shear strength, internal friction angle, and cohesion of the soil to obtain the distribution of the foot end's sinking depth in the soil.

[0008] Based on the posture data and sinking depth distribution obtained by the foot end inertial measurement unit, the slip amount of the foot end relative to the soil is processed to obtain the estimated value of the motion state of the foot end in the soil;

[0009] The motion state estimation value is input into the deep reinforcement learning agent to obtain the control action parameters output by the deep reinforcement learning agent; the control state parameters include gait parameters, joint torques and continuous body motion actions;

[0010] The actor network of the deep reinforcement learning agent is updated based on the control state parameters to obtain the control strategy and optimized control action parameters. The hardness and load-bearing capacity of the soil are processed based on the obtained contact area and shape of the foot and the soil, combined with the soil's moisture content, particle size distribution, and porosity. Based on the processed results, it is determined whether the foot will sink or slip in the soil.

[0011] If the foot sinks or slips in the soil, the sliding detection is activated, and the gait parameters are updated according to the preset sliding control strategy, sliding detection priority and multi-threaded coordination strategy to obtain the emergency adjustment amount.

[0012] In a second aspect, the present invention further provides a quadruped robot based on sliding detection, which is applied to the quadruped robot based on sliding detection in the first aspect; the quadruped robot based on sliding detection comprises:

[0013] The sinking depth processing module is used to obtain the normal force and tangential force distribution of the foot end on the soil based on the six-dimensional force sensor at the foot end. It also analyzes and processes the plastic deformation area of ​​the soil based on the shear strength, internal friction angle, and cohesion of the soil to obtain the sinking depth distribution of the foot end in the soil.

[0014] A motion state estimation module is used to process the slippage of the foot end relative to the soil based on the posture data and sinking depth distribution obtained by the foot end inertial measurement unit to obtain an estimated value of the motion state of the foot end in the soil;

[0015] A processing module is used to input the motion state estimation value into the deep reinforcement learning intelligent agent to obtain the control action parameters output by the deep reinforcement learning intelligent agent; the control state parameters include gait parameters, joint torques and continuous body movement actions;

[0016] The judgment module is used to update the actor network of the deep reinforcement learning agent based on the control state parameters to obtain the control strategy and optimized control action parameters. Based on the obtained contact area and shape of the foot end and the soil, combined with the soil's water content, particle size distribution, and porosity, the module processes the hardness and bearing capacity of the soil, and judges whether the foot end will sink or slip in the soil based on the processing results.

[0017] The emergency adjustment module is used to start sliding detection if the foot sinks or slips in the soil, and to update the gait parameters according to the preset sliding control strategy, sliding detection priority and multi-threaded coordination strategy to obtain the emergency adjustment amount.

[0018] In a third aspect, the present invention further provides an electronic device comprising: a memory for storing a computer software program; a processor for reading and executing the computer software program, thereby realizing any of the above-mentioned quadruped robots based on sliding detection.

[0019] In a fourth aspect, the present invention further provides a non-transitory computer-readable storage medium, in which a computer software program is stored. When the computer software program is executed by a processor, any of the above-mentioned quadruped robots based on sliding detection is implemented.

[0020] In a fifth aspect, the present invention further provides a computer program product, comprising a computer program, which, when executed by a processor, implements any of the above-mentioned quadruped robots based on sliding detection.

[0021] The embodiment of the present invention provides a quadruped robot based on sliding detection. Traditional robots often face the challenge of sinking and sliding of the foot end in soil, which leads to unstable movement and low energy efficiency. By obtaining the force information of the foot end and the posture data of the inertial measurement unit through the six-dimensional force sensor, and combining the deep reinforcement learning intelligent agent to analyze the physical properties of the soil, this method accurately calculates the plastic deformation and motion state of the foot end, and optimizes the gait parameters and joint torque. The strategy is adjusted in real time to cope with changes in soil hardness and carrying capacity, effectively improving the stability and energy efficiency of the robot in complex soil environments, extending its service life, and promoting the further development of robot technology in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 1 is a flow chart of a quadruped robot anti-fall method based on sliding detection provided by an embodiment of the present invention;

[0023] Figure 2 1 is a schematic structural diagram of a quadruped robot based on sliding detection provided by an embodiment of the present invention;

[0024] Figure 3 An embodiment diagram of an electronic device provided by an embodiment of the present invention;

[0025] Figure 4 An embodiment diagram of a computer-readable storage medium provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0026] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0027] In the description of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of technical features indicated. Therefore, features specified as "first" or "second" may explicitly or implicitly include one or more features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0028] In the description of the present invention, the term "for example" is used to mean "used as an example, illustration or explanation". Any embodiment of the present invention described as "for example" is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is given to enable any person skilled in the art to implement and use the present invention. In the following description, details are listed for the purpose of explanation. It should be understood that a person of ordinary skill in the art can recognize that the present invention can be implemented without using these specific details. In other examples, well-known structures and processes are not elaborated in detail to avoid obscuring the description of the present invention with unnecessary details. Therefore, the present invention is not intended to be limited to the embodiments shown, but is consistent with the widest scope consistent with the principles and features disclosed herein.

[0029] See Figure 1 , Figure 1 : is a flow chart of the anti-fall method of a quadruped robot based on sliding detection provided by the present invention. In the embodiment of the present invention, the execution subject of the quadruped robot based on sliding detection is a quadruped robot. Therefore, the anti-fall method of the quadruped robot based on sliding detection includes:

[0030] Step 10: Obtain the normal force and tangential force distribution of the foot end acting on the soil based on the six-dimensional force sensor at the foot end, and analyze and process the plastic deformation area of ​​the soil based on the shear strength, internal friction angle, and cohesion of the soil to obtain the sinking depth distribution of the foot end in the soil.

[0031] Optionally, it includes: obtaining force data of the foot end acting on the soil based on the six-dimensional force sensor at the foot end, and analyzing and extracting the force data to obtain normal force distribution and tangential force distribution; determining a preliminary foot end sinking depth distribution based on a pre-established force-deformation relationship model; inputting the obtained shear strength, internal friction angle and cohesion of the soil into the Drucker-Prager yield criterion to obtain the yield stress threshold of the soil; comparing the stress of the combination of the normal force distribution and the tangential force distribution with the yield stress threshold to obtain the plastic deformation area of ​​the soil; analyzing the preliminary foot end sinking depth distribution and the plastic deformation area, and simulating the sinking process of the foot end in the soil to obtain the sinking depth distribution result of the foot end in the soil.

[0032] Furthermore, a six-dimensional force sensor at the foot end collects force data from the foot end acting on the soil (or soft, easily dented, rough terrain). The normal and tangential force distributions are extracted from the force data, and a preliminary foot end sinking depth distribution is calculated based on a pre-established force-deformation relationship model. This force-deformation relationship model is fitted by experimental data and accounts for the influence of different soil types and moisture contents on the force-deformation relationship. Strength parameters such as shear strength, internal friction angle, and cohesion are obtained from the soil samples and input into the Drucker-Prager yield criterion to calculate the yield stress threshold of the soil. The calculated stress is compared with the yield stress threshold to determine the plastic deformation region of the soil. This determination is used to determine the material property distribution in finite element analysis. Based on the calculated plastic deformation region and the preliminary foot end sinking depth distribution, finite element analysis is performed using ANSYS software to simulate the foot end sinking process in the soil. In this analysis, the data from the six-dimensional force sensor at the foot end is converted into boundary conditions and loads for the finite element model. On this basis, a model predictive control (MPC) algorithm was incorporated to optimize the response strategy, predict future soil changes, and adjust the gait in real time, enhancing adaptability and stability. Furthermore, a pre-established soil deformation database, containing stress-strain curves for different soil types, moisture contents, and compaction levels, was integrated into the ANSYS material model. During the finite element analysis, the foot's sinking depth distribution in the soil was iteratively calculated using the Drucker-Prager yield criterion and the soil deformation database. The Newton-Raphson iteration method was used to solve the nonlinear equations and determine the displacement of each node on the foot-soil interface. Cubic spline interpolation was performed on the calculated node displacements to generate a continuous sinking depth distribution surface. Finally, the foot's sinking depth distribution in the soil was output, including the maximum sinking depth, average sinking depth, and sinking depth contour maps.

[0033] In one embodiment, the foot-mounted six-dimensional force sensor collects data at a frequency of 1000 Hz, acquiring force and torque data in the X, Y, and Z directions. The force data is converted to the ground coordinate system through coordinate transformation to obtain the normal and tangential force distributions. The force-deformation relationship model is fitted using a polynomial. For clay soils, the fitted relationship is F = k1d + k2d^2 + k3*d^3, where F is the force, d is the deformation, and k1, k2, and k3 are fitting coefficients. Preliminary calculations indicate a maximum sinking depth of 15 mm. The soil sample was found to have a shear strength of 50 kPa, an internal friction angle of 30°, and a cohesion of 20 kPa. Substituting these parameters into the Drucker-Prager yield criterion, the yield stress threshold is calculated to be 70 kPa. Comparing the calculated stress with the yield stress threshold, the plastic deformation region is determined to be a circular region with a radius of 30 mm. A foot-mounted soil contact model was established in ANSYS, using hexahedral elements with a 2 mm cell size. The six-dimensional force sensor data was converted into model boundary conditions, with the normal force acting as a vertical pressure and the tangential force acting as a shear force applied to the contact surface. A soil deformation database contains stress-strain curves for various moisture contents (10% to 40%) and compaction levels (85% to 95%). In the finite element analysis, the nonlinear equations were solved using the Newton-Raphson iteration method with a convergence accuracy of 0.1%. A stable solution was obtained after five iterations, yielding displacement data for 500 nodes on the contact surface between the foot and the soil. Cubic spline interpolation was performed on these discrete node displacements to generate a continuous 100x100 grid-like surface for the subsidence depth distribution. The final output showed a maximum subsidence depth of 18.5 mm and an average of 10.2 mm. Nine contour lines were plotted at 2 mm intervals.

[0034] Step 20 : Based on the posture data and the sinking depth distribution obtained by the foot-end inertial measurement unit, the slip amount of the foot end relative to the soil is processed to obtain an estimated value of the motion state of the foot end in the soil.

[0035] Optionally, it includes: performing Kalman filtering on the posture data to obtain smoothed three-axis angle and three-axis angular velocity data; constructing a three-dimensional grid model of the contact between the foot end and the soil based on the sinking depth distribution and combined with the geometric parameters of the foot end; the geometric parameters of the foot end include the length, width, height and center of mass position of the foot end; inputting the three-axis angular velocity data into the three-dimensional grid model to obtain the position, velocity and acceleration changes of the foot end in three-dimensional space; comparing the position of the foot end in three-dimensional space with the initial position of the foot end, and obtaining the slip amount of the foot end relative to the soil based on the comparison result and the preset soil deformation model; based on the slip amount of the foot end relative to the soil, combined with the position, velocity and acceleration changes of the foot end in three-dimensional space for analysis, obtain an estimated value of the motion state of the foot end in the soil.

[0036] Furthermore, attitude data, including three-axis angles and three-axis angular velocities, is acquired from the foot's inertial measurement unit. This attitude data is then processed using a Kalman filter to eliminate measurement noise and obtain smooth angle and angular velocity data. In the Kalman filter parameter setting, the measurement noise covariance is determined based on the sensor specifications, and the process noise covariance is obtained through experimental data tuning. Based on the previously acquired foot's soil sink depth distribution and combined with the foot's geometric parameters (including foot's length, width, height, and center of mass), a three-dimensional mesh model of the foot's contact with the soil is constructed. This mesh model uses tetrahedral elements, with the element size set to an appropriate size to ensure sufficient accuracy. The foot's kinematic equations are established using quaternions to avoid the gimbal lock problem that may arise from Euler angles. The filtered angular velocity data is substituted into the quaternion differential equation, and a quaternion representation of the foot's attitude is obtained through numerical integration. The quaternion is used to calculate the foot's position, velocity, and acceleration changes in three-dimensional space. By comparing the difference between the calculated foot position and the initial position and combining it with a soil deformation model, the slip of the foot relative to the soil is calculated. The soil deformation model considers the elastic-plastic properties of the soil and uses the Mohr-Coulomb yield criterion to describe the stress-strain relationship of the soil. By comprehensively considering position, velocity, and acceleration information, an estimate of the foot's motion state in the soil is obtained, including the three-dimensional position coordinates, linear velocity vector, angular velocity vector, and attitude quaternion. Monte Carlo error propagation analysis is performed to evaluate the accuracy of the motion state estimate and generate confidence intervals for the estimate.

[0037] In one embodiment, the foot-end inertial measurement unit collects attitude data at a frequency of 100 Hz to obtain three-axis angles (pitch angle, roll angle, yaw angle) and three-axis angular velocity. The measurement noise covariance of the Kalman filter is set to a diagonal matrix with a diagonal element value of 0.01 rad 2 The process noise covariance is 0.001rad through experimental data tuning. 2 / s 2. The accuracy of the filtered angle data is improved to ±0.1°, and the accuracy of the angular velocity data reaches ±0.05° / s. The geometric parameters of the foot end are 100mm long, 50mm wide, and 30mm high, with the center of mass located at the geometric center. The constructed three-dimensional mesh model contains 5000 tetrahedral elements with a unit size of 1mm. In the quaternion representation, the attitude quaternion q = [q0,q1,q2,q3], where q0 is the scalar part and q1,q2,q3 are the vector parts. The quaternion differential equation is dq / dt = 0.5qω, where ω is the angular velocity vector. The fourth-order Runge-Kutta method is used for numerical integration with a time step of 0.01s. In the soil deformation model, the Mohr-Coulomb yield criterion parameters are set to internal friction angle φ = 30° and cohesion c = 10kPa. The calculated foot motion state estimates include: three-dimensional spatial position coordinates (x, y, z) = (50 mm, 30 mm, -15 mm), linear velocity vector v = (0.2 m / s, 0.1 m / s, -0.05 m / s), angular velocity vector ω = (1 rad / s, 0.5 rad / s, 0.2 rad / s), and attitude quaternion q = [0.966, 0.058, 0.249, 0.058]. Using the Monte Carlo method for 1000 simulations, the 95% confidence intervals for the position estimate were ±2 mm, the 95% confidence intervals for the velocity estimate were ±0.02 m / s, and the 95% confidence intervals for the attitude estimate were ±0.5°.

[0038] In step 30, the motion state estimation value is input into the deep reinforcement learning agent to obtain the control action parameters output by the deep reinforcement learning agent; the control state parameters include gait parameters, joint torques, and continuous body motions.

[0039] Optionally, the state space of the deep reinforcement learning agent includes the three-dimensional position coordinates of the foot end, the three-axis velocity, the three-axis angular velocity, the three-axis acceleration, and the soil deformation parameters; the action space of the deep reinforcement learning agent includes the gait cycle, step length, step height, foot end trajectory, joint torque, and body posture angle; the reward function of the deep reinforcement learning agent is:

[0040] R=w1*v-w2*(|roll|+|pitch|)-w3*E;

[0041] Among them, v represents velocity, roll and pitch represent attitude angles; E represents energy consumption; w1, w2 and w3 are all weight coefficients.

[0042] Furthermore, based on the estimated foot-end motion state, a state space for a deep reinforcement learning agent is constructed, including the foot-end three-dimensional position coordinates, three-axis velocity, three-axis angular velocity, three-axis acceleration, and soil deformation parameters. Soil deformation parameters are represented by soil stiffness and damping coefficients. Maximum-minimum normalization is used to map each state variable to the interval [-1, 1]. The action space of the deep reinforcement learning agent is designed, including gait period, step length, step height, foot-end trajectory, joint torque, and body posture angle, represented as a continuous action space. The output is mapped to the interval [-1, 1] using the tanh function and then scaled according to the robot's joint limits to achieve action limiting. A reward function is constructed, comprehensively considering the robot's velocity, posture stability, and energy consumption in the soil. The reward function is expressed as: R = w1*v - w2*(|roll| + |pitch|) - w3*E, where v is velocity, roll and pitch are posture angles, E is energy consumption, and w1, w2, and w3 are weight coefficients. A deep deterministic policy gradient algorithm with a dual network structure is used to train the agent, and an evaluation network and a target network are constructed. The evaluation network contains 3 convolutional layers and 2 fully connected layers, and the output layer corresponds to the action space dimension. The target network structure is the same as the evaluation network, and the parameters are synchronized through a soft update mechanism. An experience replay buffer is set to store training samples, and batch data is randomly sampled each time for training. During the training process, the agent interacts with the simulation environment to collect state, action, reward, and next state data. The Adam optimizer is used to update the network parameters, and the learning rate is set to an appropriate value. The target network is updated every set value. After training is completed, the learned strategy is applied to the actual robot control, and the state is input into the evaluation network to obtain the corresponding action output. The learned strategy is verified in the simulation environment, the robot's movement performance under different terrain conditions is recorded, and the average reward value is calculated to evaluate the performance.

[0043] In one embodiment, the state space of the deep reinforcement learning agent contains 15 dimensions, where the three-dimensional position coordinate range of the foot end is [-0.5m, 0.5m], the three-axis velocity range is [-2m / s, 2m / s], the three-axis angular velocity range is [-πrad / s, πrad / s], and the three-axis acceleration range is [-10m / s 2 ,10m / s 2], the soil stiffness coefficient ranges from [1000 N / m to 10000 N / m], and the damping coefficient ranges from [10 Ns / m to 100 Ns / m]. After normalization, all state variables are mapped to the interval [-1, 1]. The motion space contains 10 dimensions, with a gait period range of [0.5 s, 2 s], a step length range of [0.1 m, 0.5 m], and a step height range of [0.05 m, 0.2 m]. The foot trajectory is defined by five control points, each with x and z coordinates in the range [-0.1 m, 0.1 m]. The joint torque range is [-100 Nm, 100 Nm], and the body posture angle range is [-0.5 rad, 0.5 rad]. The tanh function maps the network output to [-1, 1] and then scales it according to the respective ranges. In the reward function, w1 = 1, w2 = 0.5, and w3 = 0.01. Velocity v is in m / s, roll and pitch angles are in rad, and energy consumption E is in J. The evaluation network's three convolutional layers have 32, 64, and 128 kernels, respectively. Each fully connected layer has 128 neurons, and the output layer has 10 neurons corresponding to the action space dimensions. The experience replay buffer has a capacity of 100,000, and 256 samples are randomly sampled for training at each time. The agent interacts with the simulated environment for 10,000 episodes, with a maximum of 1,000 steps per episode. The Adam optimizer has β1 = 0.9 and β2 = 0.999. The soft update parameter τ of the target network is 0.001. After training, 100 tests were conducted under five different soil conditions, and the average reward and success rate were recorded. Finally, the learned strategy achieved an average reward value of 850 on wet and soft mud, with a success rate of 92%, while the average reward value on dry and hard mud was 920, with a success rate of 98%. The above content can also be completed using a multi-layer perceptron (MLP) model. The MLP model is used to estimate the motion state of the foot end and output control actions. The input of the MLP model includes 15 dimensions: the three-dimensional position coordinates of the foot end (range [-0.5m, 0.5m]), the three-axis velocity (range [-2m / s, 2m / s]), the three-axis angular velocity (range [-πrad / s, πrad / s]), the three-axis acceleration (range [-10m / s 2 ,10m / s 2]), soil stiffness coefficient (range [1000 N / m, 10,000 N / m]), and damping coefficient (range [10 Ns / m, 100 Ns / m]). All input variables are normalized and mapped to the interval [-1, 1]. The MLP model consists of three hidden layers, each with 128 neurons, and a Rectified Linear Unit (ReLU) activation function. The output layer uses a linear activation function to directly output the robot's motion. The output of the MLP model includes 10 dimensions: gait period (range [0.5 s, 2 s]), step length (range [0.1 m, 0.5 m]), step height (range [0.05 m, 0.2 m]), x and z coordinates of the foot trajectory control points (each control point ranges from [-0.1 m, 0.1 m]), joint torques (range [-100 Nm, 100 Nm]), and body posture angles (range [-0.5 rad, 0.5 rad]). The output is mapped to the interval [-1, 1] using a tanh function and then scaled according to the respective ranges. Training data comes from actual interaction data in a simulated environment. The MLP model uses mean squared error (MSE) as the loss function and is trained using the Adam optimizer with a learning rate of 0.001. After sufficient training, the MLP model can directly output decisions controlling the robot's movements based on the input foot state and soil parameters.

[0044] Step 40 updates the actor network of the deep reinforcement learning agent based on the control state parameters to obtain the control strategy and optimized control action parameters. Based on the obtained contact area and shape between the foot end and the soil, combined with the soil's water content, particle size distribution, and porosity, the hardness and bearing capacity of the soil are processed, and based on the processing results, it is determined whether the foot end will sink or slip in the soil.

[0045] Optionally, the actor network of the deep reinforcement learning agent is updated based on the control state parameters to obtain a control strategy and optimized control action parameters, including:

[0046] The actor network is updated according to the proximal policy optimization algorithm, and gait parameters, joint torques, and body movement continuous actions are generated according to the current state. The policy update step size is adjusted through an adaptive learning rate to keep the KL divergence between the new and old policies within the preset threshold range of 0.01 to 0.05.

[0047] Based on the obtained contact area and shape between the foot and the soil, combined with the soil's moisture content, particle size distribution, and porosity, the soil's hardness and bearing capacity are processed. Based on the processing results, it is determined whether the foot will sink or slip in the soil, including:

[0048] A pressure distribution map is constructed based on the contact area and shape between the foot and the soil obtained by the foot pressure sensor array. The pressure distribution map is binarized to obtain a binary image, and the edges of the binary image are extracted according to the Canny edge detection algorithm to obtain the contact area size and shape characteristic parameters. The shape characteristic parameters include roundness, eccentricity and Hu invariant moments. A soil physical property database is established according to the water content, particle size distribution and porosity of the soil samples. Based on the soil physical property database, the hardness and bearing capacity measurement values ​​under different water content, particle size distribution and porosity conditions are obtained. The contact area, shape characteristic parameters of the foot and the soil hardness and bearing capacity are input into a decision tree classifier to determine whether the foot will sink or slip in the soil, and the sinking risk level and slipping risk level are output.

[0049] Furthermore, a proximal policy optimization algorithm is used to update the actor network, generating gait parameters, joint torques, and continuous body motion based on the current state. The policy update step size is adjusted using an adaptive learning rate to ensure that the KL divergence between the old and new policies remains within a preset threshold range of 0.01 to 0.05. The adaptive learning rate is dynamically adjusted based on the difference between the KL divergence and the target threshold, decreasing the learning rate when the KL divergence is excessive and increasing it when it is too small. An array of foot-end pressure sensors is used to detect the contact area and shape of the foot end with the soil. The pressure distribution map is binarized using the Otsu threshold method, and the Canny edge detection algorithm is used to extract edges and calculate the contact area size and shape characteristic parameters. Shape characteristic parameters include roundness, eccentricity, and Hu invariant moments. A soil physical property database is established based on soil sample moisture content, particle size distribution, and porosity data. A support vector regression algorithm using a radial basis function kernel is used to train a soil hardness and bearing capacity prediction model. The soil physical property database contains hardness and bearing capacity measurements under different conditions of moisture content, particle size distribution (clay, silt, sand), and porosity. The foot contact area, shape characteristic parameters, and predicted soil hardness and bearing capacity are input into a decision tree classifier to determine whether the foot will sink or slip in the soil and output the sinking risk level and slipping risk level. The decision tree classifier uses the CART algorithm and the Gini coefficient as the classification criterion. The sinking risk level and slipping risk level are divided into three levels: low, medium, and high. The actions generated by the proximal policy optimization algorithm are combined with the results of foot contact detection and soil property analysis. The generated actions are evaluated and adjusted through a reinforcement learning reward function to optimize the control strategy to reduce the sinking and slipping risks.

[0050] In one embodiment, the proximal policy optimization algorithm uses 32 parallel environments to collect samples when updating the actor network. Each environment is updated after running 2048 steps. The initial learning rate is set to 3e-4, and the KL divergence target value is 0.02. When the KL divergence exceeds 0.025, the learning rate is reduced by 10%; when it is lower than 0.015, it is increased by 10%. The foot pressure sensor array consists of 8x8 sensor units with a sampling frequency of 100Hz. The Otsu threshold method is used to calculate the pressure distribution. Figure 2 The threshold is usually between 20 and 50N. The low threshold of Canny edge detection is set to 50 and the high threshold is 150. The formula for roundness calculation is 4πA / P 2 , A is the area, P is the perimeter; eccentricity e=√(1-b 2 / a 2 ), with a and b being the lengths of the major and minor axes, respectively. A support vector regression model was trained using 5,000 soil samples, with a cross-validation error of less than 5%. In the soil physical property database, moisture content was categorized into 5% intervals, particle size was classified into clay (<0.002 mm), silt (0.002–0.05 mm), and sand (0.05–2 mm), and porosity was categorized into 5% intervals. A decision tree classifier was trained using 10,000 labeled samples and achieved an accuracy of 92% using 10-fold cross-validation. Subsidence risk was categorized into low (<5 mm), medium (5–20 mm), and high (>20 mm); and slip risk was categorized into low (<10 mm), medium (10–30 mm), and high (>30 mm). The reinforcement learning reward function R = w1*v - w2*(sinking risk + slipping risk) - w3*E, where v is speed, E is energy consumption, w1 = 1, w2 = 0.5, and w3 = 0.01. After 1000 iterations of optimization, the average reward increased from an initial -50 to +150, and the sinking and slipping risks were reduced.

[0051] Step 50: If the foot sinks or slips in the soil, the sliding detection is started, and the gait parameters are updated according to the preset sliding control strategy, sliding detection priority and multi-thread coordination strategy to obtain the emergency adjustment amount.

[0052] Optionally, the foot's sinking depth and slipping distance in the soil are calculated based on data from the foot pressure sensor and inertial measurement unit. If the sinking depth exceeds 10 mm or the slipping distance exceeds 20 mm, the slip detection thread is triggered, setting the slip status flag to true. Mutexes and semaphores are used to coordinate multithreaded access to gait parameters. When the slip detection thread modifies gait parameters, a read-write lock is used to protect shared data, allowing multiple read threads to access the data simultaneously, but requiring exclusive access for write operations. The read-write lock is acquired and released according to the "reader-first" principle; the write thread waits for all read threads to complete before acquiring the lock. Condition variables are used to notify other threads of the updated gait parameters, and inter-thread synchronization is achieved using the pthread_cond_wait and pthread_cond_signal functions. After the slip detection thread is started, a fuzzy control method is used to calculate emergency adjustments to the gait parameters based on pre-defined slip control rules. The severity of the sinking depth and slipping distance are classified as "small," "medium," and "large," corresponding to adjustments to stride length, cadence, and foot trajectory. The robot's center of mass position and posture angle are taken into account simultaneously to ensure that the adjusted gait can quickly restore balance. The earliest deadline first (EDF) scheduling algorithm is used to assign high priority to the slip detection thread, and a priority inheritance protocol is used to prevent priority inversion. When a low-priority task holds a shared resource, its priority is temporarily raised to the same level as the highest-priority task requesting the resource, ensuring that the slip detection thread can respond and execute in a timely manner. The calculated emergency adjustment of the gait parameters is applied to the robot control, and an interpolation algorithm is used to smoothly transition to the new gait parameters to avoid instability caused by sudden changes. An adaptive filter is used to fuse the original control algorithm output and the emergency adjustment value to achieve real-time adjustment of the robot's gait.

[0053] In one embodiment, the foot pressure sensor samples at a frequency of 1000 Hz, and the inertial measurement unit collects data at a frequency of 200 Hz. The two sensor data are fused using a Kalman filter algorithm to calculate the foot sink depth and slip distance. When the sink depth exceeds 10 mm or the slip distance exceeds 20 mm, a slip detection thread with a priority of 90 is triggered. The read-write lock is implemented using the POSIX pthread library, with the read lock wait time set to 10 ms and the write lock wait time to 50 ms. The condition variable uses the pthread_cond_t type, with a timeout of 100 ms. The fuzzy control method uses five input variables (sink depth, slip distance, center of mass position x, y, and attitude angle) and seven output variables (step length adjustment ratio, step frequency adjustment ratio, foot trajectory adjustment in the x, y, and z directions, and body pitch and roll angle adjustment). The fuzzy rule base contains 125 rules. The time slice length of the EDF scheduling algorithm is set to 5 ms, and the relative deadline of the slip detection thread is set to 20 ms. In the priority inheritance protocol, low-priority tasks can be promoted to a maximum priority of 95. Contingency adjustments to gait parameters are implemented using a cubic spline interpolation algorithm, enabling a smooth transition to new gait parameters within 100 milliseconds. The adaptive filter uses a least mean square error (LMS) algorithm with a step size parameter μ set to 0.01 and a filter order of 32. After 1,000 simulations, this method reduced the robot's sinking depth on soft mud by an average of 62.5% and its slip distance by an average of 71.3%, while maintaining stable gait transitions.

[0054] The method further includes: step 60, after the robot executes the optimized gait parameters and joint torque in the mud environment, obtaining the real-time state based on the foot-end force sensor and the inertial measurement unit, and updating the critic network using the temporal difference algorithm according to the real-time reward and the next moment state to estimate the long-term expected reward of the current state-action pair.

[0055] Optionally, the robot executes optimized gait parameters and joint torques in a muddy environment. Data is collected through foot-end force sensors and an inertial measurement unit. A Kalman filter algorithm is used to fuse this multi-source sensor information to determine the robot's current state, including foot-end position, velocity, acceleration, and attitude angle. The state transition matrix of the Kalman filter algorithm is set to the identity matrix. An immediate reward is calculated based on the current state and a predefined reward function: R = w1*v - w2*E - w3*(|roll| + |pitch|) + w4*(1-s / s_max), where v is velocity, E is energy consumption, roll and pitch are attitude angles, s is sink depth, s_max is the maximum allowable sink depth, w1 = 1, w2 = 0.01, w3 = 0.5, and w4 = 2. State prediction is performed using a long short-term memory network (LSTM) consisting of two LSTM layers, each with 128 neurons, and a fully connected output layer. The TD(λ) algorithm is used to update the critic network. The critic network consists of three hidden layers, each with 256, 128, and 64 neurons, respectively, and uses the Reluctant Unified Unit (ReLU) activation function. The current state, action, immediate reward, and predicted next state are used as inputs. The temporal difference error (TD) is calculated and backpropagated to update the network parameters, estimating the long-term expected reward of the current state-action pair. Based on the value estimates output by the critic network, the actor network is updated using the proximal policy optimization (PPO) algorithm with a clipping parameter ε set to 0.2. By maximizing the long-term expected reward objective, the policy gradient is calculated and the network parameters are updated to generate a new policy distribution, which guides the robot's action selection in a muddy environment. The updated policy is applied to the robot's control over the action space consisting of joint angles and torques in the range [-π / 2, π / 2] and [-100 Nm, 100 Nm], respectively. Control commands are sent to the robot's actuators at a frequency of 50 Hz.

[0056] In one embodiment, in a muddy environment, the robot executes optimized gait parameters, such as a step length of 0.3m, a step frequency of 1.2Hz, and a foot-end track height of 0.15m. The joint torque is limited to ±80Nm. The foot-end force sensor acquires three-axis force data at a sampling rate of 1000Hz, and the inertial measurement unit acquires acceleration and angular velocity data at a sampling rate of 200Hz. The Kalman filter algorithm uses a 4x4 state transfer matrix and a 2x4 observation matrix. The fused instantaneous state includes foot-end position accuracy of ±1mm, velocity accuracy of ±0.01m / s, and acceleration accuracy of ±0.1m / s 2, with an attitude angle accuracy of ±0.1°. In the reward function, the robot's average velocity v in the mud is 0.5 m / s, the energy consumption E is 100 J / s, the roll and pitch angles are 3° and 2°, respectively, the sink depth s is 15 mm, and the maximum allowable sink depth s_max is set to 30 mm. The long short-term memory network was trained for 100 epochs using a batch size of 32 and a learning rate of 0.001, and the state prediction error dropped to less than 5%. The TD(λ) algorithm used a discount factor of 0.99, and the critic network was trained for 50,000 steps at a learning rate of 0.0003, converging to an average TD error of less than 0.1. The PPO algorithm used a learning rate of 0.0001 and an entropy coefficient of 0.01. After 10,000 epochs of training, the average reward in the mud environment increased from an initial -50 to 250. The updated policy maps joint angles and torques to the range [-1, 1] and restores them to the actual control variables through linear transformation. After 1,000 steps of testing on wet mud, the robot achieved an average speed of 0.45 m / s, a 30% increase in energy efficiency, a 40% reduction in sinking depth, and maintained stable movement.

[0057] Step 70 , based on the real-time state feedback of the robot in the mud and the changes in the environment, continuously optimize the actor-critic network parameters.

[0058] Optionally, a Kalman filter is used to estimate the robot's state in the soil in real time, including position, velocity, attitude angle, and sinking depth, based on data collected by the foot-end force sensor and inertial measurement unit. The state transition matrix of the Kalman filter is set to the identity matrix, and the soil hardness k and bearing capacity index n are estimated based on the relationship between the foot-end sinking velocity v and the applied force F: F = k*v^n, where k and n are obtained by least squares fitting. Based on the estimated soil properties, the weight coefficients of each term in the reward function are dynamically adjusted. The reward function expression is R = w1*v-w2*E-w3*(|roll|+|pitch|)+w4*(1-s / s_max)+w5*(k / k_ref), where v is velocity, E is energy consumption, roll and pitch are attitude angles, s is the sinking depth, s_max is the maximum allowable sinking depth, k is the estimated soil hardness, and k_ref is the reference hardness. The asynchronous dominant actor-critic (A3C) algorithm is used to continuously optimize network parameters. Experience is collected in different environment instances using 16 parallel worker threads, with each thread allocated 2 CPU cores. A centralized global network aggregates gradient updates from each thread every 100ms. The network structure and learning rate are adjusted using the model-independent meta-learning (MAML) method. The appropriate number of network layers, number of neurons, and learning rate are adaptively selected for soil environments with different hardness and bearing capacity. The adaptability of the algorithm in different soil environments is evaluated using performance indicators such as average speed, energy efficiency, stability index, and adaptation time. A given number of steps are tested in each of five typical soil environments (dry hard, wet soft, sandy, clay, and mixed), and the mean and variance of the performance indicators are recorded.

[0059] In one embodiment, in actual application, the robot obtains triaxial force data at a sampling rate of 1000Hz through the foot-end force sensor, and the inertial measurement unit obtains acceleration and angular velocity data at a sampling rate of 200Hz. The Kalman filter updates the state estimate every 10ms, with a position estimation accuracy of ±2mm, a velocity estimation accuracy of ±0.05m / s, and an attitude angle estimation accuracy of ±0.5°. In a wet and soft mud environment, the estimated soil hardness k is 5000N / m 2 , the load capacity index n is 1.3. In the reward function, the robot's average speed v is 0.4m / s, the energy consumption E is 80J / s, the roll and pitch angles are 2° and 1.5° respectively, the sinking depth s is 25mm, the maximum allowable sinking depth s_max is set to 50mm, and the reference hardness k_ref is 10000N / m 2. The A3C algorithm runs on 16 parallel threads, and each thread updates the global network every 100 steps. After 1000 iterations, the average reward increased from the initial -30 to 180. The MAML algorithm performed meta-learning in 5 different mud environments, training each environment for 20 cycles, and finally selected a 4-layer neural network structure with 128 neurons in each layer and a learning rate of 0.0003. In the adaptability test, the average speed of the robot on dry and hard mud was 0.6m / s, and the energy efficiency was improved by 25%; the average speed on wet and soft mud was 0.35m / s, and the sinking depth was reduced by 35%; the stability index on sandy mud was improved by 40%. The adaptation time of the algorithm in the new environment was shortened from the initial 500 steps to 100 steps, showing good generalization ability.

[0060] The challenge that traditional robots often face in soil is foot sinking and slipping, which leads to unstable movement and low energy efficiency. By acquiring foot force information and posture data from an inertial measurement unit through a six-dimensional force sensor and combining it with a deep reinforcement learning agent to analyze the physical properties of the soil, this method accurately calculates the plastic deformation and motion state of the foot, optimizing gait parameters and joint torque. The method adjusts strategies in real time to cope with changes in soil hardness and carrying capacity, effectively improving the robot's stability and energy efficiency in complex soil environments, extending its service life, and promoting the further development of robotics technology in practical applications.

[0061] Furthermore, the quadruped robot based on sliding detection provided by the present invention is described below. The quadruped robot based on sliding detection described below and the quadruped robot based on sliding detection described above can refer to each other.

[0062] Optional, see Figure 2 , Figure 2 : is a schematic structural diagram of a quadruped robot based on sliding detection provided by the present invention, and the quadruped robot based on sliding detection includes:

[0063] The sinking depth processing module 210 is used to obtain the normal force and tangential force distribution of the foot end on the soil based on the six-dimensional force sensor at the foot end, and analyze and process the plastic deformation area of ​​the soil based on the shear strength, internal friction angle, and cohesion of the soil to obtain the sinking depth distribution of the foot end in the soil;

[0064] The motion state estimation module 220 is used to process the slippage of the foot end relative to the soil based on the posture data and the sinking depth distribution obtained by the foot end inertial measurement unit to obtain an estimated value of the motion state of the foot end in the soil;

[0065] Processing module 230, configured to input the motion state estimation value into the deep reinforcement learning agent to obtain control action parameters output by the deep reinforcement learning agent; the control state parameters include gait parameters, joint torques, and continuous body motion;

[0066] Determination module 240 is used to update the actor network of the deep reinforcement learning agent based on the control state parameters to obtain the control strategy and optimized control action parameters. It also processes the hardness and bearing capacity of the soil based on the obtained contact area and shape of the foot end and the soil, combined with the soil's water content, particle size distribution, and porosity, and determines whether the foot end will sink or slip in the soil based on the processed results.

[0067] The emergency adjustment module 250 is used to start sliding detection if the foot sinks or slides in the soil, and update the gait parameters according to the preset sliding control strategy, sliding detection priority and multi-thread coordination strategy to obtain the emergency adjustment amount.

[0068] The challenge that traditional robots often face in soil is foot sinking and slipping, which leads to unstable movement and low energy efficiency. By acquiring foot force information and posture data from an inertial measurement unit through a six-dimensional force sensor and combining it with a deep reinforcement learning agent to analyze the physical properties of the soil, this method accurately calculates the plastic deformation and motion state of the foot, optimizing gait parameters and joint torque. The method adjusts strategies in real time to cope with changes in soil hardness and carrying capacity, effectively improving the robot's stability and energy efficiency in complex soil environments, extending its service life, and promoting the further development of robotics technology in practical applications.

[0069] See also Figure 3 , Figure 3 This is a diagram of an embodiment of an electronic device provided by an embodiment of the present invention. Figure 3 As shown, an embodiment of the present invention provides an electronic device 300, including a memory 310, a processor 320, and a computer program 311 stored in the memory 310 and executable on the processor 320. When the processor 320 executes the computer program 311, the following steps are implemented:

[0070] The six-dimensional force sensor at the foot end is used to obtain the normal and tangential force distributions of the foot end on the soil. The plastic deformation area of ​​the soil is analyzed and processed based on the shear strength, internal friction angle, and cohesion of the soil to obtain the distribution of the foot end's sinking depth in the soil.

[0071] Based on the posture data and sinking depth distribution obtained by the foot end inertial measurement unit, the slip amount of the foot end relative to the soil is processed to obtain the estimated value of the motion state of the foot end in the soil;

[0072] The motion state estimation value is input into the deep reinforcement learning agent to obtain the control action parameters output by the deep reinforcement learning agent; the control state parameters include gait parameters, joint torques and continuous body motion actions;

[0073] The actor network of the deep reinforcement learning agent is updated based on the control state parameters to obtain the control strategy and optimized control action parameters. The hardness and load-bearing capacity of the soil are processed based on the obtained contact area and shape of the foot and the soil, combined with the soil's moisture content, particle size distribution, and porosity. Based on the processed results, it is determined whether the foot will sink or slip in the soil.

[0074] If the foot sinks or slips in the soil, the sliding detection is activated, and the gait parameters are updated according to the preset sliding control strategy, sliding detection priority and multi-threaded coordination strategy to obtain the emergency adjustment amount.

[0075] See also Figure 4 , Figure 4 Detailed description of an embodiment of a computer-readable storage medium provided by an embodiment of the present invention. Figure 4 As shown, this embodiment provides a computer-readable storage medium 400 on which a computer program 311 is stored. When the computer program 311 is executed by a processor, the following steps are implemented:

[0076] The six-dimensional force sensor at the foot end is used to obtain the normal and tangential force distributions of the foot end on the soil. The plastic deformation area of ​​the soil is analyzed and processed based on the shear strength, internal friction angle, and cohesion of the soil to obtain the distribution of the foot end's sinking depth in the soil.

[0077] Based on the posture data and sinking depth distribution obtained by the foot end inertial measurement unit, the slip amount of the foot end relative to the soil is processed to obtain the estimated value of the motion state of the foot end in the soil;

[0078] The motion state estimation value is input into the deep reinforcement learning agent to obtain the control action parameters output by the deep reinforcement learning agent; the control state parameters include gait parameters, joint torques and continuous body motion actions;

[0079] The actor network of the deep reinforcement learning agent is updated based on the control state parameters to obtain the control strategy and optimized control action parameters. The hardness and load-bearing capacity of the soil are processed based on the obtained contact area and shape of the foot and the soil, combined with the soil's moisture content, particle size distribution, and porosity. Based on the processed results, it is determined whether the foot will sink or slip in the soil.

[0080] If the foot sinks or slips in the soil, the sliding detection is activated, and the gait parameters are updated according to the preset sliding control strategy, sliding detection priority and multi-threaded coordination strategy to obtain the emergency adjustment amount.

[0081] On the other hand, the present invention further provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the quadruped robot based on sliding detection provided by the above methods, which include:

[0082] The six-dimensional force sensor at the foot end is used to obtain the normal and tangential force distributions of the foot end on the soil. The plastic deformation area of ​​the soil is analyzed and processed based on the shear strength, internal friction angle, and cohesion of the soil to obtain the distribution of the foot end's sinking depth in the soil.

[0083] Based on the posture data and sinking depth distribution obtained by the foot end inertial measurement unit, the slip amount of the foot end relative to the soil is processed to obtain the estimated value of the motion state of the foot end in the soil;

[0084] The motion state estimation value is input into the deep reinforcement learning agent to obtain the control action parameters output by the deep reinforcement learning agent; the control state parameters include gait parameters, joint torques and continuous body motion actions;

[0085] The actor network of the deep reinforcement learning agent is updated based on the control state parameters to obtain the control strategy and optimized control action parameters. The hardness and load-bearing capacity of the soil are processed based on the obtained contact area and shape of the foot and the soil, combined with the soil's moisture content, particle size distribution, and porosity. Based on the processed results, it is determined whether the foot will sink or slip in the soil.

[0086] If the foot sinks or slips in the soil, the sliding detection is activated, and the gait parameters are updated according to the preset sliding control strategy, sliding detection priority and multi-threaded coordination strategy to obtain the emergency adjustment amount.

[0087] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units. That is, they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0088] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus the necessary general hardware platform, or of course, by hardware. Based on this understanding, the essence of the above technical solution or the part that contributes to the existing technology can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each embodiment or certain parts of the embodiment.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A quadruped robot anti-fall method based on sliding detection, characterized in that: include: The six-dimensional force sensor at the foot end is used to obtain the normal and tangential force distributions of the foot end on the soil. The plastic deformation area of ​​the soil is analyzed and processed based on the shear strength, internal friction angle, and cohesion of the soil to obtain the distribution of the foot end's sinking depth in the soil. Based on the posture data and sinking depth distribution obtained by the foot end inertial measurement unit, the slip amount of the foot end relative to the soil is processed to obtain the estimated value of the motion state of the foot end in the soil; The motion state estimation value is input into the deep reinforcement learning agent to obtain the control action parameters output by the deep reinforcement learning agent; the control state parameters include gait parameters, joint torques and continuous body motion actions; The actor network of the deep reinforcement learning agent is updated based on the control state parameters to obtain the control strategy and optimized control action parameters. The hardness and load-bearing capacity of the soil are processed based on the obtained contact area and shape of the foot and the soil, combined with the soil's moisture content, particle size distribution, and porosity. Based on the processed results, it is determined whether the foot will sink or slip in the soil. If the foot sinks or slips in the soil, the sliding detection is activated, and the gait parameters are updated according to the preset sliding control strategy, sliding detection priority and multi-threaded coordination strategy to obtain the emergency adjustment amount.

2. The anti-fall method of a quadruped robot based on sliding detection according to claim 1, characterized in that: The normal and tangential force distributions of the foot acting on the soil are obtained using the six-dimensional force sensor at the foot end. The plastic deformation area of ​​the soil is analyzed and processed based on the shear strength, internal friction angle, and cohesion of the soil. The depth distribution of the foot end sinking in the soil is obtained, including: The force data of the foot acting on the soil is obtained based on the six-dimensional force sensor at the foot end, and the force data is analyzed and extracted to obtain the normal force distribution and tangential force distribution; Based on the pre-established force-deformation relationship model, the preliminary foot end sinking depth distribution is determined; The obtained shear strength, internal friction angle and cohesion of the soil are input into the Drucker-Prager yield criterion to obtain the yield stress threshold of the soil; The stress of the combination of normal force distribution and tangential force distribution is compared with the yield stress threshold to obtain the plastic deformation area of ​​the soil; The preliminary foot end sinking depth distribution and plastic deformation area were analyzed, and the sinking process of the foot end in the soil was simulated to obtain the foot end sinking depth distribution results in the soil.

3. The anti-fall method of a quadruped robot based on sliding detection according to claim 1, characterized in that: Based on the posture data and sinking depth distribution obtained by the foot end inertial measurement unit, the slip amount of the foot end relative to the soil is processed to obtain the estimated value of the motion state of the foot end in the soil, including: Perform Kalman filtering on the attitude data to obtain smoothed three-axis angle and three-axis angular velocity data; Based on the sink depth distribution and the geometric parameters of the foot end, a three-dimensional mesh model of the foot end contact with the soil is constructed; the geometric parameters of the foot end include the foot end length, width, height and center of mass position; The three-axis angular velocity data is input into the three-dimensional grid model to obtain the position, velocity and acceleration changes of the foot end in three-dimensional space; The position of the foot end in three-dimensional space is compared with the initial position of the foot end, and the slip amount of the foot end relative to the soil is obtained based on the comparison result and a preset soil deformation model; Based on the slip of the foot end relative to the soil and combined with the changes in the position, velocity and acceleration of the foot end in three-dimensional space, an estimated value of the motion state of the foot end in the soil is obtained.

4. The anti-fall method for a quadruped robot based on sliding detection according to claim 1, characterized in that: The state space of the deep reinforcement learning agent includes the three-dimensional position coordinates of the foot end, three-axis velocity, three-axis angular velocity, three-axis acceleration, and soil deformation parameters; the action space of the deep reinforcement learning agent includes gait cycle, step length, step height, foot end trajectory, joint torque, and body posture angle; the reward function of the deep reinforcement learning agent is: R=w1*v-w2*(|roll|+|pitch|)-w3*E; Where v represents velocity, roll and pitch represent attitude angles; E represents energy consumption; w1, w2 and w3 are all weight coefficients.

5. The anti-fall method of a quadruped robot based on sliding detection according to claim 1, characterized in that: Based on the control state parameters, the actor network of the deep reinforcement learning agent is updated to obtain the control strategy and optimized control action parameters, including: The actor network is updated according to the proximal policy optimization algorithm, and gait parameters, joint torques, and body movement continuous actions are generated according to the current state. The policy update step size is adjusted through an adaptive learning rate to keep the KL divergence between the new and old policies within the preset threshold range of 0.01 to 0.

05.

6. The anti-fall method for a quadruped robot based on sliding detection according to claim 1, characterized in that: Based on the obtained contact area and shape between the foot and the soil, combined with the soil's moisture content, particle size distribution, and porosity, the soil's hardness and bearing capacity are processed. Based on the processing results, it is determined whether the foot will sink or slip in the soil, including: Constructing a pressure distribution map based on the contact area and shape between the foot and the soil obtained by the foot pressure sensor array; The pressure distribution map is binarized to obtain a binary image, and the edge of the binary image is extracted using the Canny edge detection algorithm to obtain the contact area size and shape characteristic parameters; the shape characteristic parameters include roundness, eccentricity and Hu invariant moment; A soil physical properties database is established based on the moisture content, particle size distribution, and porosity of soil samples. Based on this database, hardness and bearing capacity measurements are obtained under different moisture content, particle size distribution, and porosity conditions. The contact area, shape characteristic parameters, soil hardness, and bearing capacity of the foot end are input into the decision tree classifier to determine whether the foot end will sink or slip in the soil, and the sinking risk level and slipping risk level are output.

7. The anti-fall method for a quadruped robot based on sliding detection according to claim 1, characterized in that: Also includes: After the robot executes the optimized gait parameters and joint torques in the mud environment, it obtains the real-time state based on the foot force sensor and inertial measurement unit. Based on the real-time reward and the next-moment state, the critic network is updated using the temporal difference algorithm to estimate the long-term expected reward of the current state-action pair. The actor-critic network parameters are continuously optimized based on the real-time state feedback of the robot in the mud and environmental changes.

8. A quadruped robot based on sliding detection, characterized in that: An anti-fall method for a quadruped robot based on sliding detection as claimed in any one of claims 1 to 7; Quadruped robots based on sliding detection include: The sinking depth processing module is used to obtain the normal force and tangential force distribution of the foot end on the soil based on the six-dimensional force sensor at the foot end. It also analyzes and processes the plastic deformation area of ​​the soil based on the shear strength, internal friction angle, and cohesion of the soil to obtain the sinking depth distribution of the foot end in the soil. A motion state estimation module is used to process the slippage of the foot end relative to the soil based on the posture data and sinking depth distribution obtained by the foot end inertial measurement unit to obtain an estimated value of the motion state of the foot end in the soil; A processing module is used to input the motion state estimation value into the deep reinforcement learning intelligent agent to obtain the control action parameters output by the deep reinforcement learning intelligent agent; the control state parameters include gait parameters, joint torques and continuous body movement actions; The judgment module is used to update the actor network of the deep reinforcement learning agent based on the control state parameters to obtain the control strategy and optimized control action parameters. Based on the obtained contact area and shape of the foot end and the soil, combined with the soil's water content, particle size distribution, and porosity, the module processes the hardness and bearing capacity of the soil, and judges whether the foot end will sink or slip in the soil based on the processing results. The emergency adjustment module is used to start sliding detection if the foot sinks or slips in the soil, and to update the gait parameters according to the preset sliding control strategy, sliding detection priority and multi-threaded coordination strategy to obtain the emergency adjustment amount.

9. An electronic device comprising: Memory for storing computer software programs; A processor for reading and executing a computer software program, wherein when the processor executes the computer software program, it implements the anti-fall method of a quadruped robot based on sliding detection as claimed in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer software program stored therein, characterized in that: When the computer software program is executed by a processor, the anti-fall method of a quadruped robot based on sliding detection as claimed in any one of claims 1 to 7 is implemented.

Images