Robot dynamic path planning method and system based on deep learning

Abstract

The invention provides a robot dynamic path planning method and system based on deep learning, and the method comprises the steps: building a motion response model which can reflect the dynamic characteristics of a robot, and carrying out the global sensitivity analysis of key motion parameters including the track smoothness, the obstacle avoidance safety distance and the motion speed; determining a feasible interval of each key motion parameter according to an analysis result; a path point sequence and a system control instruction are taken as decision variables, path length and motion time minimization is taken as a target, numerical optimization solution is carried out under the constraint of a motion response model, and an optimal path point sequence and optimal control instruction combination and a corresponding optimal motion track are obtained through forward simulation; and determining an optimal trajectory tracking control parameter of each motion stage according to the feasible interval of each key motion parameter, the optimal path point sequence and the optimal control instruction. By adopting the scheme of the invention, accurate track tracking control of the multi-target optimization driven robot can be realized, so that the operation efficiency under a complex working condition is improved.

Description

Technical Field

[0001] This application relates to the field of automatic control technology, and more specifically, to a method and system for robot dynamic path planning based on deep learning. Background Technology

[0002] Robot automatic control technology has evolved from early programmed fixed-point operations to today's intelligent systems with certain environmental perception and trajectory planning capabilities. In traditional control architectures, robot dynamics models are often constructed using mechanism modeling methods based on Lagrange equations or Newton-Euler methods, and trajectory tracking is achieved by combining proportional-integral-derivative control, fuzzy control and other methods.

[0003] In existing technologies, traditional mechanism-based modeling control methods struggle to accurately describe the true dynamic characteristics of robots under the coupling effects of dynamic loads, environmental disturbances, and system delays. Due to the complexity of robot mechanical structures, their dynamic characteristics are influenced by multiple factors such as joint flexibility, friction variations, and load fluctuations. Simplified models built solely based on physical laws cannot fully reflect the system's dynamic response in actual operation. This model mismatch is particularly pronounced in unstructured environments with random obstacles, leading to significant deviations in the trajectory planned based on the model during actual tracking. While some research attempts to improve model accuracy through online parameter identification, the real-time performance and accuracy of parameter identification are difficult to guarantee in scenarios with multi-factor coupling and rapidly changing operating conditions. This results in a decline in robot dynamic performance, failing to simultaneously meet the multi-objective optimization requirements of optimal path length, shortest motion time, and smooth trajectory. Therefore, achieving precise trajectory tracking control of robots driven by multi-objective optimization to improve operational efficiency under complex conditions has become a challenging problem for the industry. Summary of the Invention

[0004] This application provides a robot dynamic path planning method and system based on deep learning, which can realize precise tracking control of robot trajectory driven by multi-objective optimization, thereby improving the operating efficiency under complex working conditions.

[0005] Firstly, this application provides a deep learning-based method for dynamic path planning of robots, comprising: A motion model of the robot in a dynamic environment is established. The motion model is constructed by a deep neural network to learn the evolution law of the robot's motion state under dynamic load, environmental disturbance and system response delay. The deep neural network is trained and its parameters are calibrated using actual collected path tracking data to obtain a motion response model that can truly reflect the dynamic characteristics of the robot. Based on the motion response model, a global sensitivity analysis is performed on key motion parameters including trajectory smoothness, obstacle avoidance safety distance, and motion speed. Based on the analysis results, the feasible range of each key motion parameter is determined. Using the path point sequence and system control commands as decision variables, and minimizing the path length and motion time as objectives, numerical optimization is performed under the constraints of the motion response model. The optimal path point sequence, optimal control command combination, and their corresponding optimal motion trajectory are obtained through forward simulation. Based on the feasible range of each key motion parameter, the optimal path point sequence, and the optimal control command, the optimal trajectory tracking control parameters for each motion stage are determined, and the robot's motion trajectory is controlled by feedback based on the optimal trajectory tracking control parameters and the optimal motion trajectory.

[0006] Preferably, establishing a motion model of the robot in a dynamic environment specifically includes: A deep neural network structure is constructed, which includes an input layer, multiple hidden layers, and an output layer. The input layer is used to receive the robot's current state variables and environmental perception data, and the output layer is used to output the robot's next state. Define the input variables of the deep neural network structure, including the robot's position, velocity, acceleration, dynamic load parameters, environmental disturbance characteristics, and system delay time; Define the output variables of the deep neural network structure, including the robot's predicted position, predicted velocity, and predicted orientation angle; The deep neural network structure is pre-trained using historical motion data to initialize network weights and bias parameters, thereby obtaining the motion model.

[0007] Preferably, training and calibrating the deep neural network using actual collected path tracking data to obtain a motion response model that can truly reflect the robot's dynamic characteristics specifically includes: Serialized path tracking data with timestamps is collected through sensor modules; The serialized path tracking data is preprocessed to eliminate data noise and standardize data scale; The preprocessed data is input into the deep neural network in chronological order, and forward propagation calculation is performed to obtain the loss function value. Based on the loss function value, the weight parameters of the deep neural network are adjusted through backpropagation until the model converges, thus obtaining the motion response model.

[0008] Preferably, the path tracking data is collected synchronously by multi-source sensors deployed on the robot body.

[0009] Preferably, the global sensitivity analysis determines the feasible range of parameters by evaluating the disturbance response of trajectory smoothness, obstacle avoidance safety distance, and motion speed.

[0010] Preferably, using the path point sequence and system control commands as decision variables, and minimizing the path length and motion time as the objective, numerical optimization is performed under the constraints of the motion response model. The optimal path point sequence, optimal control command combination, and their corresponding optimal motion trajectory are obtained through forward simulation, specifically including: The continuous path planning problem is discretized into a decision problem involving multiple path points, and the path point coordinates and system control commands are used together as decision variables. A multi-objective optimization function is constructed with the objectives of minimizing path length and motion time. The motion response model is used as a dynamic constraint, and the feasible intervals of each key motion parameter are used as boundary constraints to form the constraint conditions of the optimization problem. The decision variables are solved using a constrained multi-objective optimization algorithm, and the Pareto optimal solution set is output. According to the preset decision rules, a set of solutions is selected from the Pareto optimal solution set as the optimal path point sequence and the optimal control command combination; The optimal path point sequence and the optimal control command are combined and input into the motion response model for forward simulation to generate the corresponding optimal motion trajectory.

[0011] Preferably, the optimal trajectory tracking control parameters for each motion stage are determined based on the feasible range of each key motion parameter, the optimal path point sequence, and the optimal control command. Specifically, these parameters include: The optimal path point sequence is divided into multiple motion stages in chronological order: acceleration, constant speed, deceleration, and turning. For each motion phase, extract the curvature features and velocity change features of the trajectory within that phase; Based on the curvature characteristics and velocity change characteristics, the initial parameters of the trajectory tracking controller are obtained by querying a preset parameter mapping table. The initial parameters are validated and fine-tuned by considering the feasible ranges of each key motion parameter to generate the optimal trajectory tracking control parameters for each motion stage.

[0012] Preferably, the trajectory tracking controller is an adaptive control module that dynamically adjusts the control gain based on trajectory curvature and velocity characteristics. Preferably, feedback control of the robot's motion trajectory based on the optimal trajectory tracking control parameters and the optimal motion trajectory specifically includes: During the robot's movement, its global positioning data and inertial measurement unit data are acquired in real time; The real-time acquired global positioning data and inertial measurement unit data are compared with the current expected state of the optimal motion trajectory to generate a state error vector; The state error vector is input to the trajectory tracking controller for the corresponding motion stage; The trajectory tracking controller calculates a compensation control command for eliminating state errors based on the optimal trajectory tracking control parameters. The compensation control command is superimposed with the optimal control command to generate a drive command, which is then sent to the robot's actuator.

[0013] Secondly, this application provides a deep learning-based robot dynamic path planning system, comprising: The model building module is used to build a motion model of the robot in a dynamic environment. The motion model is constructed by a deep neural network to learn the evolution law of the robot's motion state under dynamic load, environmental disturbance and system response delay conditions. The deep neural network is trained and its parameters are calibrated using actual collected path tracking data to obtain a motion response model that can truly reflect the robot's dynamic characteristics. The processing module is used to perform global sensitivity analysis on key motion parameters, including trajectory smoothness, obstacle avoidance safety distance and motion speed, based on the motion response model, and determine the feasible range of each key motion parameter according to the analysis results. The processing module is also used to perform numerical optimization under the constraints of the motion response model, with path point sequence and system control commands as decision variables and minimizing path length and motion time as objectives, and to obtain the optimal path point sequence and optimal control command combination and their corresponding optimal motion trajectory through forward simulation. The execution module is used to determine the optimal trajectory tracking control parameters for each motion stage based on the feasible range of each key motion parameter, the optimal path point sequence, and the optimal control command, and to perform feedback control on the robot's motion trajectory based on the optimal trajectory tracking control parameters and the optimal motion trajectory.

[0014] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: In this embodiment, a motion model of the robot in a dynamic environment is first established. This motion model is constructed using a deep neural network to learn the evolution of the robot's motion state under dynamic load, environmental disturbance, and system response delay conditions. The deep neural network is trained and its parameters are calibrated using actual collected path tracking data to obtain a motion response model that truly reflects the robot's dynamic characteristics. Based on the motion response model, a global sensitivity analysis is performed on key motion parameters, including trajectory smoothness, obstacle avoidance safety distance, and motion speed. The feasible range of each key motion parameter is determined based on the analysis results. Using the path point sequence and system control commands as decision variables, and minimizing the path length and motion time as the objective, numerical optimization is performed under the constraints of the motion response model. The optimal path point sequence, optimal control command combination, and their corresponding optimal motion trajectory are obtained through forward simulation. Based on the feasible range of each key motion parameter and the optimal path point sequence and optimal control command, the optimal trajectory tracking control parameters for each motion stage are determined. Feedback control of the robot's motion trajectory is then performed based on the optimal trajectory tracking control parameters and the optimal motion trajectory.

[0015] Therefore, this application determines the optimal trajectory tracking control parameters for each motion stage by using the feasible ranges of each key motion parameter, the optimal path point sequence, and the optimal control command. Based on these optimal trajectory tracking control parameters and the optimal motion trajectory, feedback control is performed on the robot's motion trajectory. First, by establishing a motion response model for the robot, the temporal motion state of the robot can be accurately predicted under dynamic disturbances and load changes, ensuring that path planning is based on real dynamic responses, reducing trajectory deviations caused by model mismatch, and improving the stability and execution efficiency of the entire planning process from the source. Second, by determining the feasible ranges of each key motion parameter based on the global sensitivity analysis results, a verifiable set of constraints can be established between performance indicators, converging the optimization search space and improving the robustness and executability of the planning solution. Then, using the path point sequence and system control command as decision variables, and minimizing the path length and motion time as the objective, the optimal trajectory tracking control parameters for each motion stage are determined. Numerical optimization is performed under the constraints of the model. Forward simulation yields the optimal path point sequence, optimal control command combination, and corresponding optimal motion trajectory. Multi-objective optimization generates a solution set that satisfies dynamic feasibility and lies on the Pareto front, achieving a trade-off between path efficiency and time cost. Forward simulation verifies the executability and safety margin of the selected solution in a real dynamic environment. Finally, by determining the feasible intervals of key motion parameters and the optimal path point sequence and optimal control commands, the optimal trajectory tracking control parameters for each motion stage are determined. This enables the construction of a feedforward and feedback coordinated control mechanism, allowing the robot to maintain stable and high-precision trajectory following performance during acceleration, constant speed, deceleration, and turning phases. This reduces additional energy consumption and adjustment time in dynamic environments, thereby improving task execution efficiency. In summary, the proposed solution achieves precise trajectory tracking control of a robot driven by multi-objective optimization, thereby improving operational efficiency under complex working conditions. Attached Figure Description

[0016] Figure 1 This is an exemplary flowchart of a deep learning-based robot dynamic path planning method according to some embodiments of this application; Figure 2 This is a flowchart illustrating the feasible range of each key motion parameter according to some embodiments of this application; Figure 3 This is a schematic flowchart illustrating the generation of optimal trajectory tracking control parameters according to some embodiments of this application; Figure 4 This is a schematic diagram of the structure of a deep learning-based robot dynamic path planning system according to some embodiments of this application; Figure 5 This is a schematic diagram of the structure of a computer device that implements a deep learning-based robot dynamic path planning method according to some embodiments of this application. Detailed Implementation

[0017] To better understand the technical solution of this application, the technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] refer to Figure 1 The figure is an exemplary flowchart of a deep learning-based robot dynamic path planning method according to some embodiments of this application. The deep learning-based robot dynamic path planning method mainly includes the following steps: In step 101, a motion model of the robot in a dynamic environment is established. The motion model is constructed by a deep neural network to learn the evolution law of the robot's motion state under dynamic load, environmental disturbance and system response delay conditions. The deep neural network is trained and its parameters are calibrated using actual collected path tracking data to obtain a motion response model that can truly reflect the robot's dynamic characteristics.

[0019] In some embodiments, establishing a motion model of the robot in a dynamic environment can be achieved by the following steps: A deep neural network structure is constructed, which includes an input layer, multiple hidden layers, and an output layer. The input layer is used to receive the robot's current state variables and environmental perception data, and the output layer is used to output the robot's next state. Define the input variables of the deep neural network structure, including the robot's position, velocity, acceleration, dynamic load parameters, environmental disturbance characteristics, and system delay time; Define the output variables of the deep neural network structure, including the robot's predicted position, predicted velocity, and predicted orientation angle; The deep neural network structure is pre-trained using historical motion data to initialize network weights and bias parameters, thereby obtaining the motion model.

[0020] It should be noted that the motion model in this application is used to describe the evolution of the robot's motion state in a dynamic environment. After the network parameters are pre-trained and initialized with historical motion data, the model can provide state prediction support for path planning.

[0021] In practical implementation, the deep neural network structure can be constructed as follows: Considering the nonlinear mapping requirements of robot dynamic motion, a deep neural network structure is designed, comprising an input layer, multiple hidden layers, and an output layer. The number of neurons in the input layer is determined based on the total dimension of the input variables, and each neuron corresponds to a receiving channel for input parameters, used to accurately receive the robot's current state variables and environmental perception data. The number of hidden layers is set to 3-5 layers, and the number of neurons in each layer is optimized using gradient descent based on the dimensions of the input and output layers. Adjacent hidden layers are connected via fully connected layers, and each hidden layer uses the ReLU activation function to enhance the network's ability to approximate nonlinear motion patterns. The ReLU activation function's threshold parameter is set to 0, meaning it outputs the input value when it's greater than 0 and outputs 0 when it's less than or equal to 0. The number of neurons in the output layer matches the dimension of the output variable, and a linear activation function is used to output the robot's next state without distortion. The network architecture, which includes an input layer, multiple hidden layers, and an output layer, and clearly defines the activation function type, connection method, and method for determining the number of neurons, serves as the deep neural network structure for constructing the motion model. Secondly, the input variables of the deep neural network structure can be defined as follows: the robot's position refers to its three-dimensional coordinate parameters in a preset Cartesian coordinate system, i.e., the XYZ axis coordinates. These parameters are determined by the robot's... The robot's speed is acquired by laser positioning sensors; the robot's velocity refers to its three-dimensional linear velocity parameters, i.e., linear velocity along the XYZ axes, acquired by the robot's wheel speed encoders or inertial measurement units; the robot's acceleration refers to its three-dimensional acceleration parameters, i.e., acceleration along the XYZ axes, acquired by the inertial measurement units; dynamic load parameters include load mass and load distribution coefficient, with load mass acquired by pressure sensors and load distribution coefficient calculated based on the ratio of load mass to the robot's own mass; environmental disturbance characteristics include obstacle movement speed, obstacle position change rate, and terrain flatness coefficient, with obstacle movement speed and position change rate acquired by visual sensors and terrain flatness coefficient acquired by tactile sensors. The system delay time is calculated from the terrain height difference data collected by the sensors; the system delay time refers to the time interval from the issuance of the control command to the robot's actuator generating a response, which is obtained by recording the difference between the command issuance time and the response trigger time using a timer; then, the output variables of the deep neural network structure can be defined as follows: the predicted position refers to the three-dimensional coordinate parameters of the robot after one sampling period, which uses the same Cartesian coordinate system as the robot position in the input variables, and is used to characterize the robot's spatial position at future moments; the predicted velocity refers to the three-dimensional linear velocity parameters of the robot after one sampling period, which maintains the same parameter dimension as the robot velocity in the input variables, and is used to characterize the robot's motion speed and direction at future moments;The predicted orientation angle refers to the angle between the robot's forward direction and the positive X-axis of the Cartesian coordinate system after one sampling period, ranging from 0 to 360 degrees, used to characterize the robot's future posture orientation. Finally, the deep neural network structure is pre-trained using historical motion data to initialize network weights and bias parameters. The resulting motion model can be implemented as follows: historical motion data of the robot under different dynamic loads, environmental disturbances, and system delays can be collected. The collection equipment includes laser positioning sensors, wheel speed encoders, inertial measurement units, pressure sensors, vision sensors, tactile sensors, and timers. The collected historical motion data needs to be... The data collection covers all parameters corresponding to the input variables and the actual observed parameters corresponding to the output variables, with a collection time span of no less than 72 hours to ensure that the data covers common robot motion scenarios and working conditions. The collected historical motion data is preprocessed. First, median filtering is used to remove outliers, with outliers defined as data deviating from the mean by three standard deviations. Then, min-max normalization is used to map all data to the interval between 0 and 1. The normalization formula is: processed data = (original data - minimum data value) / (maximum data value - minimum data value). The preprocessed historical motion data is divided into training and validation sets in a 7:3 ratio. The training set is used for network weights. The iteration update of the weight and bias parameters is performed, and the validation set is used to monitor overfitting during network training. The training parameters of the deep neural network are set, with the mean squared error loss function used to calculate the error between the predicted and actual values. The loss function formula is: mean squared error = (sum of squares of predicted values ​​- actual values) divided by the number of samples. The optimization algorithm is an adaptive momentum estimation algorithm. The initial learning rate is set to 0.001, and as the number of training iterations increases, the learning rate decays to 0.9 times the previous iteration every 100 iterations. The total number of training iterations is set to 1000. The convergence condition is that the mean squared error of the validation set is less than 0.0001 or... The maximum number of iterations is reached; the network weight parameters are initialized using the Xavier initialization method to ensure consistent input and output variances for each neuron layer, and all bias parameters are initialized to 0; based on the training set data, the network weights and bias parameters are iteratively updated using the backpropagation algorithm. After each iteration, the mean squared error is calculated using the validation set data. If the mean squared error on the validation set does not decrease for 20 consecutive iterations, training is terminated early; after training, the final network weights and bias parameters are saved, and the pre-trained deep neural network with initialized weights and bias parameters, capable of outputting the predicted state for the next moment based on the current input state variables and environmental perception data, is used as the initial motion model.

[0022] In some embodiments, training and calibrating the deep neural network with actual collected path tracking data to obtain a motion response model that can truly reflect the robot's dynamic characteristics can be achieved in the following manner: Serialized path tracking data with timestamps is collected through sensor modules; The serialized path tracking data is preprocessed to eliminate data noise and standardize data scale; The preprocessed data is input into the deep neural network in chronological order, and forward propagation calculation is performed to obtain the loss function value. Based on the loss function value, the weight parameters of the deep neural network are adjusted through backpropagation until the model converges, thus obtaining the motion response model.

[0023] It should be noted that the path tracking data in this application refers to the time-series correlation data collected by sensors during the robot's path tracking process, which includes its own motion state, environmental influencing factors, and actual motion results; the motion response model is a trained and calibrated deep neural network model used to realistically reflect the robot's dynamic characteristics and provide state prediction support for path planning and trajectory control.

[0024] In specific implementation, firstly, the serialized path tracking data with timestamps can be acquired through a sensor module in the following way: A sensor module consisting of a laser positioning sensor, wheel speed encoder, inertial measurement unit, pressure sensor, vision sensor, tactile sensor, and high-precision timer is used to collect real-time data on the robot's three-dimensional position, velocity, acceleration, dynamic load parameters, environmental disturbance characteristics, system delay time, and the corresponding actual next-moment position, velocity, and orientation angle. During acquisition, the timer adds a microsecond-level timestamp to each data set, and the sampling frequency is set to 100Hz. Secondly, the serialized path tracking data is preprocessed to eliminate data noise and unify data scale, which can be achieved as follows: A median filtering method with a window size of 5 can be used to remove outliers deviating from three standard deviations of the median in the window. The original data is replaced with the median in the window to remove noise. Then, a minimum-maximum normalization method is used to map all the denoised data to the 0-1 interval. The input parameters and the actual output parameters undergo the same operation to ensure consistency. Finally, the preprocessed data is input into the deep neural network in chronological order to perform forward propagation calculation and obtain the loss. The function value can be implemented as follows: Preprocessed input parameters are input into the constructed deep neural network in time-stamp order. After weighted calculation and activation transformation at each layer, the predicted position, velocity, and orientation angle are output. The mean squared error loss function is used to calculate the deviation between the predicted value and the actual output parameters, obtaining a numerical index of the quantized deviation. The quantized deviation index obtained from forward propagation is used as the loss function value. Finally, based on the loss function value, the weight parameters of the deep neural network are adjusted through backpropagation until the model converges. The motion response model can be implemented as follows: Based on... The loss function value is calculated using the chain rule of backpropagation to determine the gradient of each layer's weights. The weight parameters are then adjusted using an adaptive momentum estimation algorithm. The initial learning rate is set to 0.001, decaying to 0.9 times the previous rate every 100 rounds, with a maximum iteration count of 2000 rounds. The convergence condition is that the loss function value is less than 0.0001 for 30 consecutive rounds. If the validation set loss increases for 25 consecutive rounds, the model is stopped early. After the model meets the convergence condition, the final weight parameters and network configuration are saved. The deep neural network that has been trained and calibrated, meets the convergence condition, and can truly reflect the robot's dynamic characteristics is used as the motion response model.

[0025] In step 102, based on the motion response model, a global sensitivity analysis is performed on key motion parameters including trajectory smoothness, obstacle avoidance safety distance, and motion speed, and the feasible range of each key motion parameter is determined according to the analysis results.

[0026] In some embodiments, based on the motion response model, performing a global sensitivity analysis on key motion parameters, including trajectory smoothness, obstacle avoidance safety distance, and motion speed, can be achieved using the following steps: In the motion response model, a parameter sampling space is set for trajectory smoothness, obstacle avoidance safety distance, and motion speed; Within the parameter sampling space, multiple sets of parameter samples are generated using the Latin hypercube sampling method; Each set of parameter samples is input into the motion response model to calculate the path length and motion time. Based on the path length and motion time corresponding to all parameter samples, the sensitivity index of each key motion parameter is calculated using regression analysis.

[0027] It should be noted that the parameter sampling space in this application is used to define the value boundaries of each key motion parameter; the sensitivity index is a quantitative indicator used to quantify the degree of influence of each key motion parameter on path length and motion time, providing a basis for determining the feasible range of parameters.

[0028] In specific implementation, firstly, in the motion response model, the parameter sampling space for trajectory smoothness, obstacle avoidance safety distance, and motion speed can be implemented in the following way: trajectory smoothness is measured by the rate of change of trajectory curvature, and the value range is set to 0.1 to 1.0 per meter based on the maximum turning angle and motion stability requirements of the robot's steering mechanism; the obstacle avoidance safety distance is set to a value range of 0.3 to 2.0 meters based on the robot's own dimensions (length and width) and the motion uncertainty of environmental obstacles; the motion speed is set to a value range of 0.1 to 1.5 meters per second, referencing the robot's rated travel speed, acceleration and deceleration capabilities, and load-bearing limits. Then, the parameters with clearly defined values ​​and boundaries are further specified. The set of key motion parameter numerical intervals serves as the parameter sampling space for trajectory smoothness, obstacle avoidance safety distance, and motion speed. Secondly, each set of parameter samples is input into the motion response model. The calculation of path length and motion time can be achieved as follows: Based on the set parameter sampling space, the number of sampling groups is determined to be 100 to 200, ensuring coverage of all areas of the parameter space and avoiding sample redundancy. Then, the Latin hypercube sampling method is used to uniformly divide the sampling space of each key motion parameter into non-overlapping sub-intervals according to the number of sampling groups. A sample point is randomly selected within each sub-interval, and the sample points for each parameter are not repeated in their dimensions, ensuring the uniformity and representativeness of the samples. This is then processed through... The Latin hypercube sampling method generates multiple sets of key motion parameter combinations covering the value range of each parameter within the parameter sampling space as parameter samples. Then, each set of parameter samples is input into the motion response model. The path length and motion time can be calculated as follows: each generated set of parameter samples is sequentially input into the trained and calibrated motion response model. The model simulates the robot's motion process in a dynamic environment based on the input trajectory smoothness, obstacle avoidance safety distance, and motion speed parameters. The path length is obtained by calculating the sum of the straight-line distances between adjacent path points in the optimal path point sequence output by the model; the motion time is calculated by dividing the path length by the robot's average speed under that parameter sample. The average motion speed is calculated. Then, based on the path length and motion time corresponding to all parameter samples, the sensitivity index of each key motion parameter is calculated using regression analysis. This can be achieved in the following way: using trajectory smoothness, obstacle avoidance safety distance, and motion speed as independent variables, and the corresponding path length and motion time as dependent variables, a regression model is constructed using multiple linear regression analysis. By fitting the regression equation, the regression coefficients of each independent variable are calculated. Then, the regression coefficients are standardized to eliminate the influence of the difference in the dimensions of each parameter, and the quantitative value of the influence of each key motion parameter on the path length and motion time is obtained. The quantitative value of the influence is used as the sensitivity index of the corresponding key motion parameter.

[0029] In some embodiments, reference Figure 2As shown in the figure, this is a flowchart illustrating the process of determining the feasible ranges of each key motion parameter in some embodiments of this application. In this embodiment, the determination of the feasible ranges of each key motion parameter based on the analysis results can be achieved using the following steps: In step 1021, based on the sensitivity index of each key motion parameter, key motion parameters that have an impact on path length and motion time exceeding a preset threshold are selected. In step 1022, the monotonicity relationship between the selected key motion parameters and path length and motion time is analyzed. In step 1023, based on the monotonicity relationship and the preset path length and upper limit of motion time, the range of values ​​for key motion parameters is derived in reverse. In step 1024, the intersection of the value range and the physical limits of the robot is taken to determine the feasible range of the selected key motion parameters.

[0030] It should be noted that the monotonicity relation in this application is used to reflect the variation law between the selected key motion parameters and the path length and motion time; the feasible interval is the effective numerical range used to constrain the values ​​of key motion parameters in robot path planning.

[0031] In specific implementation, firstly, based on the sensitivity index of each key motion parameter, the key motion parameters whose impact on path length and motion time exceeds a preset threshold can be screened out in the following way: First, a preset threshold for the sensitivity index is set. This threshold is determined based on the judgment standard of the degree of parameter influence in engineering applications, and is set to 0.6. When the sensitivity index is greater than 0.6, it is determined that the parameter has a significant impact on path length or motion time. The sensitivity index of each key motion parameter is compared with the preset threshold, and all parameters with a sensitivity index exceeding 0.6 are screened out. The parameters that have a significant impact on path length and motion time, as determined by the threshold comparison, are selected as the key motion parameters. Secondly, targeting... The monotonicity analysis of the selected key motion parameters with path length and motion time can be achieved as follows: For each selected key motion parameter, extract all corresponding parameter samples, path length, and motion time data. By plotting parameter-path length curves and parameter-motion time curves, observe the changing trends of the two performance indicators as the parameter values ​​change. If the path length continuously increases as the parameter value increases, it is determined to be a positively correlated monotonic relationship; if the path length continuously decreases as the parameter value increases, it is determined to be a negatively correlated monotonic relationship; if the changing trend has no fixed pattern, it is determined to be a non-monotonic relationship. Then, based on the monotonicity relationship and the preset upper limits of path length and motion time, reverse the process... The derivation of the value range of key motion parameters can be achieved as follows: First, based on the actual application requirements of robot path planning, set upper limits for path length and motion time. For parameters with a positively correlated monotonic relationship, according to the rule that "increasing the parameter leads to an increase in performance index," deduce the maximum value of the parameter that ensures the performance index does not exceed the preset upper limit. For parameters with a negatively correlated monotonic relationship, according to the rule that "increasing the parameter leads to a decrease in performance index," deduce the minimum value of the parameter that ensures the performance index does not exceed the preset upper limit. Combine the physical meaning of the parameters to supplement lower limits (such as velocity not less than 0) or upper limit constraints, forming a preliminary numerical range. This preliminary numerical range is then used as the value range of the key motion parameters. Finally, the intersection of the value range and the robot's physical limits is used to determine the feasible range of the selected key motion parameters. This can be achieved in the following way: First, the robot's physical limits are defined. These limits are determined by the robot's hardware performance parameters and structural design constraints, including the maximum trajectory smoothness allowed by the steering mechanism, the minimum obstacle avoidance safety distance determined by the mechanical structure, the maximum motion speed supported by the power system, and the minimum stable speed. The value range of the key motion parameters is then intersected with the corresponding robot physical limits. Values ​​exceeding the physical limits are eliminated, and the value range that simultaneously meets the performance requirements and physical limitations is retained. This value range is then used as the feasible range of the selected key motion parameters.

[0032] In step 103, the path point sequence and system control commands are used as decision variables, and the path length and motion time are minimized as objectives. Numerical optimization is performed under the constraints of the motion response model. The optimal path point sequence and optimal control command combination and their corresponding optimal motion trajectory are obtained through forward simulation.

[0033] In some embodiments, using the path point sequence and system control commands as decision variables, and minimizing the path length and motion time as the objective, numerical optimization is performed under the constraints of the motion response model. The optimal path point sequence, optimal control command combination, and their corresponding optimal motion trajectory can be obtained through forward simulation using the following steps: The continuous path planning problem is discretized into a decision problem involving multiple path points, and the path point coordinates and system control commands are used together as decision variables. A multi-objective optimization function is constructed with the objectives of minimizing path length and motion time. The motion response model is used as a dynamic constraint, and the feasible intervals of each key motion parameter are used as boundary constraints to form the constraint conditions of the optimization problem. The decision variables are solved using a constrained multi-objective optimization algorithm, and the Pareto optimal solution set is output. According to the preset decision rules, a set of solutions is selected from the Pareto optimal solution set as the optimal path point sequence and the optimal control command combination; The optimal path point sequence and the optimal control command are combined and input into the motion response model for forward simulation to generate the corresponding optimal motion trajectory.

[0034] It should be noted that the decision variables in this application are a set of variables used to describe path point coordinates and system control commands; the multi-objective optimization function is a mathematical function used to simultaneously minimize path length and motion time, wherein the path length is calculated by the sum of the Euclidean distances between adjacent points in the path point sequence, and the motion time is calculated by the ratio of the path length to the robot's average motion speed; the constraints are a set of dynamic constraints and boundary constraints used to limit the solution space of the optimization problem; the Pareto optimal solution set is used to represent the solution set that cannot be further improved among multiple objectives; the optimal path point sequence and optimal control command combination are used to represent the best path point sequence and control command combination selected according to the decision rules in the Pareto optimal solution set, wherein the path point sequence is the robot's position sequence, and the control command is the robot's speed and control parameter sequence; the optimal motion trajectory is used to represent the robot motion path generated by forward simulation, which ensures feasibility and optimality in a dynamic environment and meets the objectives of minimizing path length and motion time.

[0035] In practical implementation, firstly, the continuous path planning problem is discretized into a decision problem involving multiple path points, and the path point coordinates and system control commands are used as decision variables. This can be achieved as follows: First, the robot's continuous motion path is divided into multiple discrete path points, each defined by three-dimensional coordinates, determined by the division of the robot's workspace. System control commands include velocity and acceleration commands, which are set based on the robot's actuator capabilities. The path point coordinates and system control commands together constitute the decision variables, used to adjust the path and control strategy during optimization. Secondly, with the shortest path length and minimum motion time as optimization objectives, a multi-objective optimization function is constructed. This can be achieved as follows: First, the path length is defined as one of the objective functions, obtained by summing the straight-line distances between all adjacent path points in the path point sequence. Motion time is defined as another objective function, calculated by dividing the path length by the robot's average motion speed on that path. These two objective functions are combined into a multi-objective optimization function, expressed using a weighted sum method or vector optimization. A multi-objective optimization function is constructed by taking the shortest path length and the minimum movement time as optimization objectives. It should be further explained that the specific form of the multi-objective optimization function is a mathematical function that simultaneously minimizes the total path length and the total movement time. This function merges two independent objectives into a solvable scalar function or a set of target vectors to be optimized using either a weighted sum method or a target vector method. The total path length is obtained by calculating the sum of the Euclidean distances between all adjacent path points in the path point sequence, used to quantify the spatial distance traveled by the robot. The total movement time is obtained by dividing the total path length by the robot's average speed along that path, used to quantify the time consumed by the robot's movement. In actual construction, if the weighted sum method is used, weight coefficients are assigned to each objective, and the weighted objectives are summed to form a scalar function. The weight coefficients are adjusted according to task requirements; for example, when efficiency is emphasized, movement time is given higher weight. If the target vector method is used, the independence of the two objectives is maintained, and they are expressed as a single target vector that needs to be optimized simultaneously. This vector is then directly processed by a multi-objective optimization algorithm, and the above mathematical expression is ultimately used as the multi-objective optimization function for numerical optimization.Then, the motion response model is used as a dynamic constraint, and the feasible intervals of each key motion parameter are used as boundary constraints. The constraints of the optimization problem can be implemented in the following way: the motion response model, which is a robot motion prediction model built based on a deep neural network, is used to verify the dynamic feasibility of the path point sequence and control commands. The feasible intervals of each key motion parameter are used as boundary constraints. These feasible intervals are the ranges of parameter values ​​determined through global sensitivity analysis, such as the ranges for trajectory smoothness, obstacle avoidance safety distance, and motion speed. These constraints together constitute the constraints of the optimization problem, limiting the values ​​of the decision variables.

[0036] In specific implementation, a constrained multi-objective optimization algorithm is used to solve the decision variables, and the output Pareto optimal solution set can be achieved in the following way: A constrained multi-objective optimization algorithm, such as a non-dominated sorting genetic algorithm or a multi-objective particle swarm optimization algorithm, is used to iteratively search the decision variables. Under the premise of satisfying the constraints, the algorithm evaluates the objective function value of each solution and selects non-dominated solutions. These non-dominated solutions constitute the Pareto optimal solution set, representing the set of optimal trade-off solutions for path length and motion time objectives. The solution set obtained by the constrained multi-objective optimization algorithm is used as the Pareto optimal solution set. A set of solutions is selected from the Pareto optimal solution set according to preset decision rules as the optimal path point sequence and optimal control command combination. This can be achieved in the following way: The decision preference is clarified based on the robot's specific task requirements, such as prioritizing work efficiency or prioritizing energy saving. In the Pareto optimal solution set, if the ideal point method is used, the objective function value (i.e., path length and motion time) of each solution is calculated to the ideal point formed by the theoretical minimum of the two objectives. The Euclidean distance between points is calculated, and the solution with the smallest distance is selected. If a weighted sum method is used, appropriate weight coefficients are assigned to the two objectives of path length and motion time, and the weighted objective function value of each solution is calculated, and the solution with the smallest value is selected. If the task has strict time requirements, a priority method can be used, that is, the solution with the shortest path length among all solutions that meet the upper limit of motion time is selected first. Finally, the unique solution selected by any of the above determination rules is used as the optimal path point sequence and the optimal control command combination. The optimal path point sequence and the optimal control command combination are input into the motion response model for forward simulation to generate the corresponding optimal motion trajectory. This can be achieved in the following way: First, the optimal path point sequence and the optimal control command combination are input into the motion response model. The motion response model is a prediction model based on a deep neural network, used to simulate the evolution of the robot's motion state in a dynamic environment. The changes of the robot's position, velocity, and orientation angle over time are calculated through forward simulation to generate a continuous motion trajectory. This trajectory is used as the optimal motion trajectory for actual robot control.

[0037] In step 104, the optimal trajectory tracking control parameters for each motion stage are determined based on the feasible range of each key motion parameter, the optimal path point sequence, and the optimal control command. Then, the robot's motion trajectory is controlled by feedback based on the optimal trajectory tracking control parameters and the optimal motion trajectory.

[0038] In some embodiments, reference Figure 3 As shown in the figure, this is a flowchart illustrating the generation of optimal trajectory tracking control parameters in some embodiments of this application. In this embodiment, the determination of optimal trajectory tracking control parameters for each motion stage based on the feasible range of each key motion parameter, the optimal path point sequence, and the optimal control command can be achieved through the following steps: In step 1041, the optimal path point sequence is divided into multiple motion stages in chronological order: acceleration, constant speed, deceleration, and turning. In step 1042, for each motion phase, the curvature features and velocity change features of the trajectory within that phase are extracted; In step 1043, based on the curvature characteristics and velocity change characteristics, a preset parameter mapping table is queried to obtain the initial parameters of the trajectory tracking controller; In step 1044, the initial parameters are verified and fine-tuned based on the feasible range of each key motion parameter to generate the optimal trajectory tracking control parameters for each motion stage.

[0039] It should be noted that the trajectory tracking controller in this application is an adaptive control module that dynamically adjusts the control gain based on trajectory curvature and velocity characteristics; the optimal trajectory tracking control parameters are the optimal control parameters used to implement precise trajectory tracking feedback control of the robot in the corresponding motion phase.

[0040] In specific implementation, firstly, dividing the optimal path point sequence into multiple motion stages—acceleration, constant speed, deceleration, and turning—in chronological order can be achieved as follows: The changing trend of the robot's velocity profile can be analyzed based on the timestamps of the optimal path point sequence and the velocity parameters in the associated optimal control commands; continuous path point intervals with continuously increasing speeds are divided into acceleration motion stages, continuous path point intervals with stable speeds are divided into constant speed motion stages, and continuous path point intervals with continuously decreasing speeds are divided into deceleration motion stages; simultaneously, based on the change in the direction angle of the connecting lines of adjacent path points, continuous path point intervals with a direction angle change rate exceeding a preset threshold are separately identified as turning motion stages; secondly, for each motion... In each stage, extracting the curvature and velocity change characteristics of the trajectory within that stage can be achieved as follows: For each defined motion stage, firstly, obtain the coordinates of all path points included in that motion stage; based on these path point coordinates, calculate the curvature of the trajectory using numerical differentiation or curve fitting, and use the average and maximum curvature values ​​within that motion stage as curvature characteristics; simultaneously, extract the velocity command sequence from the optimal control command corresponding to that motion stage, and use the average, maximum, and rate of change of velocity within that motion stage as velocity change characteristics; then, based on the curvature and velocity change characteristics, query a preset parameter mapping table to obtain the initial parameters of the trajectory tracking controller, which can be achieved as follows: The pre-defined parameter mapping table is a database established through simulation or experimentation. This database uses different curvature and velocity change characteristics as input indices and verified trajectory tracking controller parameters that achieve good control performance as output data. The curvature and velocity change characteristics of the motion phase are used as query conditions to perform a matching search in the parameter mapping table, finding a set of controller parameters corresponding to the closest feature combination. This set of controller parameters is then used as the initial parameters for the trajectory tracking controller in the current motion phase. Finally, the initial parameters are validated and fine-tuned based on the feasible ranges of each key motion parameter to generate the optimal trajectory tracking control parameters for each motion phase. The following approach is used: Obtain the feasible ranges of each key motion parameter determined through global sensitivity analysis, especially the feasible ranges of trajectory smoothness and motion speed, which are directly related to controller performance. Substitute the initial parameters of the trajectory tracking controller obtained from the query into the motion response model for rapid verification. Check whether the robot's predicted trajectory controlled by these parameters meets the feasible range constraints of all key motion parameters. If there are cases that exceed the feasible range, iteratively fine-tune the initial parameters until the predicted trajectory fully meets all feasible range constraints, while ensuring that the control performance degradation is within an acceptable range. Finally, use the controller parameters determined after compliance verification and fine-tuning as the optimal trajectory tracking control parameters for each motion stage.

[0041] In some embodiments, feedback control of the robot's motion trajectory based on the optimal trajectory tracking control parameters and the optimal motion trajectory can be achieved using the following steps: During the robot's movement, its global positioning data and inertial measurement unit data are acquired in real time; The real-time acquired global positioning data and inertial measurement unit data are compared with the current expected state of the optimal motion trajectory to generate a state error vector; The state error vector is input to the trajectory tracking controller for the corresponding motion stage; The trajectory tracking controller calculates a compensation control command for eliminating state errors based on the optimal trajectory tracking control parameters. The compensation control command is superimposed with the optimal control command to generate a drive command, which is then sent to the robot's actuator.

[0042] It should be noted that the state error vector in this application is a vector used to quantify the deviation between the actual state and the desired state of the robot in multiple dimensions; the trajectory tracking controller refers to the algorithm module used to calculate the control compensation amount based on the state error; the compensation control instruction refers to the incremental control signal used to correct the deviation of the robot's motion trajectory; and the drive instruction refers to the control signal used to directly drive the motion of the robot's actuator.

[0043] In specific implementation, firstly, during the robot's movement, real-time acquisition of its global positioning data and inertial measurement unit (IMU) data can be achieved as follows: The robot's three-dimensional coordinates in the geodetic coordinate system or a preset workspace coordinate system are acquired in real-time using a GPS receiver or indoor positioning base station mounted on the robot body, and these three-dimensional coordinates are used as global positioning data. The robot's three-axis angular velocity and three-axis acceleration are acquired in real-time using the IMU mounted on the robot body, and these raw measurements are used as IMU data. Secondly, comparing the real-time acquired global positioning data and IMU data with the current desired state of the optimal motion trajectory to generate a state error vector can be achieved as follows: Based on the current timestamp, the corresponding desired robot position, desired velocity, and desired orientation angle are retrieved from the optimal motion trajectory, together constituting the current desired state. The real-time acquired global positioning data is then compared with the desired position to calculate the position error. The robot's actual velocity calculated using the IMU data is compared with the desired velocity to calculate the velocity error. The robot's actual orientation angle calculated using the IMU data is compared with the desired orientation angle to calculate the orientation angle error. Finally, the calculated... The calculated position error, velocity error, and orientation angle error are sequentially combined into a multi-dimensional vector, which is then used as the state error vector. The trajectory tracking controller calculates a compensation control command to eliminate the state error based on the optimal trajectory tracking control parameters. This can be achieved in the following way: The trajectory tracking controller uses its internally set optimal trajectory tracking control parameters as control algorithm coefficients. After receiving the state error vector, the controller processes it according to its control algorithm, and then uses the incremental control signal output to reduce or even eliminate the current state error as the compensation control command. Finally, the compensation control command is superimposed with the optimal control command to generate a drive command, which is then sent to the robot's actuators. This can be achieved in the following way: Based on the current timestamp, the corresponding preset control command is retrieved from the optimal control command sequence. This preset control command is a feedforward command derived from the path planning stage. The compensation control command calculated by the trajectory tracking controller is then algebraically superimposed with the preset optimal control command. The resulting comprehensive command is the final drive command, which is then sent to the robot's underlying actuators such as motors, servos, or hydraulic valves via the robot's control bus.

[0044] On the other hand, in some embodiments, this application provides a robot dynamic path planning system based on deep learning, with reference to... Figure 4The figure is a schematic diagram of the structure of a deep learning-based robot dynamic path planning system 400 according to some embodiments of this application. The deep learning-based robot dynamic path planning system 400 includes: a model building module 401, a processing module 402, and an execution module 403, which are described below: The model building module 401 in this application is mainly used to build a motion model of the robot in a dynamic environment. The motion model is constructed by a deep neural network to learn the evolution law of the robot's motion state under dynamic load, environmental disturbance and system response delay conditions. The deep neural network is trained and its parameters are calibrated using actual collected path tracking data to obtain a motion response model that can truly reflect the dynamic characteristics of the robot. Processing module 402, in this application, is used to perform global sensitivity analysis on key motion parameters including trajectory smoothness, obstacle avoidance safety distance and motion speed based on the motion response model, and determine the feasible range of each key motion parameter according to the analysis results; In this application, the processing module 402 is also used to perform numerical optimization under the constraints of the motion response model, with the path point sequence and system control commands as decision variables and the path length and motion time as the objective, and to obtain the optimal path point sequence and optimal control command combination and their corresponding optimal motion trajectory through forward simulation. The execution module 403 in this application is mainly used to determine the optimal trajectory tracking control parameters for each motion stage based on the feasible range of each key motion parameter, the optimal path point sequence, and the optimal control command, and to perform feedback control on the robot's motion trajectory based on the optimal trajectory tracking control parameters and the optimal motion trajectory.

[0045] In addition, this application also provides a computer device, the computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described deep learning-based robot dynamic path planning method.

[0046] In some embodiments, reference Figure 5 The figure is a schematic diagram of the structure of a computer device implementing a deep learning-based robot dynamic path planning method according to some embodiments of this application. The deep learning-based robot dynamic path planning method in the above embodiments can... Figure 5 The computer device shown is used to implement this, and the computer device 500 includes at least one processor 501, a communication bus 502, a memory 503, and at least one communication interface 504.

[0047] Processor 501 can be a general-purpose central processing unit (CPU) or an application-specific integrated circuit (ASIC).

[0048] The communication bus 502 can be used to transmit information between the aforementioned components.

[0049] Memory 503 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CDROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital versatile optical discs, Blu-ray discs, etc.), magnetic disks or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. Memory 503 may exist independently and be connected to processor 501 via communication bus 502. Memory 503 may also be integrated with processor 501.

[0050] The memory 503 stores program code for executing the scheme of this application, and its execution is controlled by the processor 501. The processor 501 executes the program code stored in the memory 503. The program code may include one or more software modules. In the above embodiments, the deep learning-based robot dynamic path planning method can be implemented by the processor 501 and one or more software modules in the program code in the memory 503.

[0051] Communication interface 504 uses any transceiver-like device to communicate with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0052] In a specific implementation, as one example, a computer device may include multiple processors, each of which may be a single-core (single CPU) processor or a multi-core (multi CPU) processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0053] The aforementioned computer device can be a general-purpose computer device or a special-purpose computer device. In specific implementations, the computer device can be a desktop computer, a portable computer, a network server, a handheld digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. This application does not limit the type of computer device.

[0054] In addition, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described deep learning-based robot dynamic path planning method.

[0055] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0056] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A deep learning-based dynamic path planning method for robots, characterized in that, include: A motion model of the robot in a dynamic environment is established. The motion model is constructed by a deep neural network to learn the evolution law of the robot's motion state under dynamic load, environmental disturbance and system response delay. The deep neural network is trained and its parameters are calibrated using actual collected path tracking data to obtain a motion response model that can truly reflect the dynamic characteristics of the robot. Based on the motion response model, a global sensitivity analysis is performed on key motion parameters including trajectory smoothness, obstacle avoidance safety distance, and motion speed. Based on the analysis results, the feasible range of each key motion parameter is determined. Using the path point sequence and system control commands as decision variables, and minimizing the path length and motion time as objectives, numerical optimization is performed under the constraints of the motion response model. The optimal path point sequence, optimal control command combination, and their corresponding optimal motion trajectory are obtained through forward simulation. Based on the feasible range of each key motion parameter, the optimal path point sequence, and the optimal control command, the optimal trajectory tracking control parameters for each motion stage are determined, and the robot's motion trajectory is controlled by feedback based on the optimal trajectory tracking control parameters and the optimal motion trajectory.

2. The method as described in claim 1, characterized in that, Establishing a motion model for a robot in a dynamic environment specifically includes: A deep neural network structure is constructed, which includes an input layer, multiple hidden layers, and an output layer. The input layer is used to receive the robot's current state variables and environmental perception data, and the output layer is used to output the robot's next state. Define the input variables of the deep neural network structure, including the robot's position, velocity, acceleration, dynamic load parameters, environmental disturbance characteristics, and system delay time; Define the output variables of the deep neural network structure, including the robot's predicted position, predicted velocity, and predicted orientation angle; The deep neural network structure is pre-trained using historical motion data to initialize network weights and bias parameters, thereby obtaining the motion model.

3. The method as described in claim 1, characterized in that, The deep neural network is trained and its parameters calibrated using actual path tracking data to obtain a motion response model that truly reflects the robot's dynamic characteristics. Specifically, this includes: Serialized path tracking data with timestamps is collected through sensor modules; The serialized path tracking data is preprocessed to eliminate data noise and standardize data scale; The preprocessed data is input into the deep neural network in chronological order, and forward propagation calculation is performed to obtain the loss function value. Based on the loss function value, the weight parameters of the deep neural network are adjusted through backpropagation until the model converges, thus obtaining the motion response model.

4. The method as described in claim 1, characterized in that, The path tracking data is collected synchronously by multiple source sensors deployed on the robot body.

5. The method as described in claim 1, characterized in that, The global sensitivity analysis determines the feasible range of parameters by evaluating the disturbance response of trajectory smoothness, obstacle avoidance safety distance, and motion speed.

6. The method as described in claim 1, characterized in that, Using the path point sequence and system control commands as decision variables, and minimizing path length and motion time as the objective, numerical optimization is performed under the constraints of the motion response model. The optimal path point sequence, optimal control command combination, and their corresponding optimal motion trajectory are obtained through forward simulation, specifically including: The continuous path planning problem is discretized into a decision problem involving multiple path points, and the path point coordinates and system control commands are used together as decision variables. A multi-objective optimization function is constructed with the objectives of minimizing path length and motion time. The motion response model is used as a dynamic constraint, and the feasible intervals of each key motion parameter are used as boundary constraints to form the constraint conditions of the optimization problem. The decision variables are solved using a constrained multi-objective optimization algorithm, and the Pareto optimal solution set is output. According to the preset decision rules, a set of solutions is selected from the Pareto optimal solution set as the optimal path point sequence and the optimal control command combination; The optimal path point sequence and the optimal control command are combined and input into the motion response model for forward simulation to generate the corresponding optimal motion trajectory.

7. The method as described in claim 1, characterized in that, Based on the feasible ranges of each key motion parameter, the optimal path point sequence, and the optimal control command, the optimal trajectory tracking control parameters for each motion stage are determined, specifically including: The optimal path point sequence is divided into multiple motion stages in chronological order: acceleration, constant speed, deceleration, and turning. For each motion phase, extract the curvature features and velocity change features of the trajectory within that phase; Based on the curvature characteristics and velocity change characteristics, the initial parameters of the trajectory tracking controller are obtained by querying a preset parameter mapping table. The initial parameters are validated and fine-tuned by considering the feasible ranges of each key motion parameter to generate the optimal trajectory tracking control parameters for each motion stage.

8. The method as described in claim 1, characterized in that, The trajectory tracking controller is an adaptive control module that dynamically adjusts the control gain based on trajectory curvature and velocity characteristics.

9. The method as described in claim 1, characterized in that, Feedback control of the robot's motion trajectory based on the optimal trajectory tracking control parameters and optimal motion trajectory specifically includes: During the robot's movement, its global positioning data and inertial measurement unit data are acquired in real time; The real-time acquired global positioning data and inertial measurement unit data are compared with the current expected state of the optimal motion trajectory to generate a state error vector; The state error vector is input to the trajectory tracking controller for the corresponding motion stage; The trajectory tracking controller calculates a compensation control command for eliminating state errors based on the optimal trajectory tracking control parameters. The compensation control command is superimposed with the optimal control command to generate a drive command, which is then sent to the robot's actuator.

10. A robot dynamic path planning system based on deep learning, characterized in that, include: The model building module is used to build a motion model of the robot in a dynamic environment. The motion model is constructed by a deep neural network to learn the evolution law of the robot's motion state under dynamic load, environmental disturbance and system response delay conditions. The deep neural network is trained and its parameters are calibrated using actual collected path tracking data to obtain a motion response model that can truly reflect the robot's dynamic characteristics. The processing module is used to perform global sensitivity analysis on key motion parameters, including trajectory smoothness, obstacle avoidance safety distance and motion speed, based on the motion response model, and determine the feasible range of each key motion parameter according to the analysis results. The processing module is also used to perform numerical optimization under the constraints of the motion response model, with path point sequence and system control commands as decision variables and minimizing path length and motion time as objectives, and to obtain the optimal path point sequence and optimal control command combination and their corresponding optimal motion trajectory through forward simulation. The execution module is used to determine the optimal trajectory tracking control parameters for each motion stage based on the feasible range of each key motion parameter, the optimal path point sequence, and the optimal control command, and to perform feedback control on the robot's motion trajectory based on the optimal trajectory tracking control parameters and the optimal motion trajectory.

Images