A tracked vehicle path tracking method based on liquid time constant neural network

Abstract

This invention discloses a path tracking method for tracked vehicles based on a liquid time constant neural network, belonging to the field of intelligent control technology for tracked vehicles. The method includes: establishing an expert demonstration database based on MPC (Multi-Process Control) and collecting path tracking error and drive motor torque data; performing robust scaling normalization and long sequence segmentation preprocessing on the data; building an LTCNN model with an NCP (Non-Conceptual Processing) architecture in the PyTorch framework and training it using the AdamW optimizer through MPC behavior cloning; inputting the real-time path tracking error into the trained model and outputting the drive motor torque to control the vehicle to track the desired path. This invention ensures path tracking accuracy through MPC behavior cloning, achieving a lateral error RMSE of 0.089m under small curvature conditions and 0.13m under large curvature conditions; the forward inference of LTCNN improves real-time performance, saving 34.8% of runtime compared to MPC; data preprocessing and the NCP architecture enhance the model's robustness and interpretability, resulting in smoother output torque, making it suitable for path tracking of tracked vehicles in complex off-road environments.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for tracked vehicles, specifically to a path tracking method for tracked vehicles based on a Liquid Time-Constant Neural Network (LTCNN), which aims to improve the path tracking accuracy and real-time performance of tracked vehicles in complex off-road environments. Background Technology

[0002] Tracked vehicles are widely used in military, rescue, and agricultural fields due to their excellent terrain adaptability and maneuverability. With the development of intelligent technologies, the autonomous navigation and path-tracking capabilities of tracked vehicles have become a research hotspot.

[0003] However, in the random and ever-changing off-road environment, tracked vehicles face challenges such as time-varying ground parameters and random disturbances, making it difficult for traditional control methods to meet the requirements of high precision and real-time performance. Therefore, developing an efficient and robust path-following control method is particularly important. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a path tracking method for tracked vehicles based on a liquid time constant neural network. This method solves the technical problems of low path tracking accuracy, poor real-time performance, and insufficient robustness of traditional control methods in complex off-road environments, and achieves efficient and accurate path tracking control for tracked vehicles.

[0005] To solve the above problems, the technical solution adopted by the present invention is as follows: A path tracking method for tracked vehicles based on a liquid time constant neural network includes the following steps: Step 1: Establish an expert demonstration database based on model predictive control algorithms: Obtain the position coordinates of the desired path, calculate the desired heading angle based on the geometric relationship between the coordinate points, and obtain the path tracking error by combining the vehicle's driving status. Based on the kinematic model of the tracked vehicle with instantaneous steering center, an MPC controller is designed. The control quantity is determined to be the winding speed of the left and right tracks. The corresponding vehicle speed and yaw rate are tracked by the PID algorithm to obtain the torque of the left and right drive motors. An MPC controller and a tracked vehicle model were built in the Simulink environment. Paths for various continuous steering conditions were designed. Expert demonstration data were collected through path tracking simulation. The input of the expert demonstration data was the path tracking error, and the output was the torque of the left and right drive motors. Step Two: Data Collection Based on the Expert Demonstration Database The expert demonstration data is divided into training set, validation set and test set in a ratio of 3:1:1; A robust scaling method is used to standardize the input and output data. First, a robust scaler is fitted to the input and output data of the training set, and the median and interquartile range of the data are calculated. Then, the fitted scaler is used to perform the same scaling transformation on the training set, validation set and test set. The long sequences of working conditions in each set are divided into multiple short sequences with some overlap to form the final training set, validation set, and test set. Step 3: Build and train the liquid time constant neural network control model: An LTCNN path tracking control model was built in the PyTorch framework. The model was wired using the NCP neural circuit architecture, which includes sensory neurons, interneurons, command neurons, and motor neurons. The model was trained using the AdamW optimizer with MPC behavior cloning. The training loss function was the mean square error between the model's predicted value and the expert demonstration value of the drive motor torque. Step 4: Tracked vehicle path tracking control: Input the real-time tracked vehicle path tracking error into the trained LTCNN path tracking control model. The model outputs the torque of the left and right drive motors, and controls the tracked vehicle to track the desired path based on this torque.

[0006] Furthermore, the formula for calculating robust scaling in step two is as follows: ,in The median of the characteristic column. The interquartile range is equal to the 75th percentile. Subtract the 25th percentile .

[0007] Furthermore, the partially overlapping short sequences mentioned in step two allow adjacent short sequences to share some time steps, simulating the rolling optimization window of MPC and enhancing the model's ability to capture the continuous dynamics of the sequence.

[0008] Furthermore, in step three, the LTCNN path tracking model uses NCP wiring, with the number of neurons set to 16 (excluding sensory neurons). Among them, the number of intermediate neurons is 9, the number of command neurons is 5, and the number of motor neurons is 2, which is the output feature dimension. The neurons are functionally layered to realize signal transmission and processing.

[0009] Furthermore, the parameters for model training in step three are set as follows: exponential decay rate of the first moment is 0.9, exponential decay rate of the second moment is 0.999, learning rate of LTCNN is 1e-3, weight decay coefficient is 0.01, training sequence length is 50, and number of iterations is 4000.

[0010] Furthermore, in step three, LTCNN performs forward propagation by fusing the ODE solver. The solution process is as follows: for the trajectory... arrive The system will span consecutive time periods Discretized Equally spaced time points Calculate only each time arrive The local trajectory changes are updated iteratively from arrive The hidden states of neurons are ultimately obtained. The system state output at time t is noteworthy; it's important to note that when the number of discrete points... When the size is large enough, the result of this piecewise linear approximation will approach the true continuous-time analytical solution infinitely. Therefore, LTCNN combines explicit Euler and implicit Euler and uses sufficiently fine time discretization to accurately fit its nonlinear features.

[0011] Furthermore, the fused ODE solver is used to solve the ordinary differential equations of LTCNN: ,in It is a time constant. In dynamic state, For system input, The reference potential, specifically, means that at... Enter the following to hide the status. Regarding time The derivative is used by Parameterized Neural Networks Describe it; Transform it into discrete form: in, For time intervals.

[0012] Furthermore, in step one, the location coordinates of the desired path are obtained. The desired heading angle is calculated based on the geometric relationship before and after the coordinate point. , will the desired path information With the vehicle's driving status Combined with the obtained path tracking error .

[0013] Furthermore, in step three, the MPC controller is designed using a tracked vehicle kinematic model based on the instantaneous steering center, according to... To calculate the optimal control quantity According to kinematic relationships, the longitudinal and yaw motions of the vehicle, i.e., vehicle speed and yaw rate, can be affected by changing the winding speed of the left and right tracks. Therefore, the MPC control variable is determined to be... ,in These represent the left and right track winding speeds, respectively. To ensure consistent control, after the MPC obtains the desired track winding speed, a PID algorithm is used to track the corresponding vehicle speed and yaw rate to obtain the torque of the left and right drive motors. .

[0014] Furthermore, in step three, the functional division of neurons in each layer of the NCP neural circuit architecture is clear: sensory neurons receive path tracking error input, intermediate neurons perform information processing and feature extraction, command neurons generate control decisions, and motor neurons output torque commands to drive the motor.

[0015] The beneficial effects of this invention are as follows: 1. High path tracking accuracy: Through MPC behavior cloning training, the LTCNN model can accurately fit expert demonstration data. Under the condition of continuous steering with small curvature, the RMSE of the lateral error of path tracking is only 0.089m; under the condition of continuous steering with large curvature, the RMSE of the lateral error is 0.13m, and the control accuracy is significantly better than that of traditional control methods.

[0016] 2. Excellent real-time performance: The neural network control based on behavior cloning achieves rapid response through forward inference, resulting in high computational efficiency. Under the same simulation conditions, the average runtime of the LTCNN model is only 124.8s, saving 34.8% of the runtime compared to MPC's 191.5s, effectively solving the computational latency problem of traditional MPC.

[0017] 3. Strong robustness: The data preprocessing adopts a robust scaling method, which weakens the impact of outliers and noise on the data. At the same time, the model's ability to capture dynamic changes in sequences is enhanced by the partial overlapping short sequence partitioning, so that the model can still work stably in complex off-road environments with time-varying ground parameters and random disturbances.

[0018] 4. Good output stability: The torque curve of the drive motor output by the trained LTCNN model is more stable than that of MPC. The torque fluctuation is less than 20 N·m under small curvature conditions and less than 50 N·m under large curvature conditions. This alleviates the problem of violent torque fluctuation caused by the forward movement of the desired path point, which is beneficial to protecting the vehicle drive system and improving driving stability.

[0019] 5. High interpretability: The model adopts the NCP neural circuit architecture, with neurons designed in functional layers. The function of neurons in each layer is clear, which facilitates the understanding and optimization of the model's control logic and reduces the model's maintenance cost. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the robust scaling effect, where (a) is the input feature before preprocessing, (b) is the input feature after preprocessing, (c) is the output feature before preprocessing, and (d) is the output feature after preprocessing. Figure 2 The diagram shows a comparison of wiring methods, where (a) is the fully connected method, (b) is the random sparse method, and (c) is the NCP method. Figure 3 The results of LTCNN path tracking control based on MPC behavior cloning (small curvature condition) are shown, where (a) is the path tracking effect, (b) is the path tracking error input, and (c) is the drive motor torque output. Figure 4 The results of LTCNN path tracking control based on MPC behavior cloning (high curvature condition) are shown, where (a) is the path tracking effect, (b) is the path tracking error input, and (c) is the drive motor torque output. Detailed Implementation

[0021] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0022] A path tracking method for tracked vehicles based on a liquid time constant neural network includes the following steps: Step 1: Establishment of an expert demonstration database based on model predictive control algorithms In the tracked vehicle path tracking control problem, the position coordinates of the desired path are first obtained. The desired heading angle is calculated based on the geometric relationship before and after the coordinate point. , will the desired path information With the vehicle's driving status Combined with the obtained path tracking error The MPC controller is designed using a tracked vehicle kinematic model based on the instantaneous steering center. To calculate the optimal control quantity According to kinematic relationships, the longitudinal and yaw motions of the vehicle, i.e., vehicle speed and yaw rate, can be affected by changing the winding speed of the left and right tracks. Therefore, the MPC control variable is determined to be... ,in These represent the left and right track winding speeds, respectively. To ensure consistent control, after the MPC obtains the desired track winding speed, a PID algorithm is used to track the corresponding vehicle speed and yaw rate to obtain the torque of the left and right drive motors. .

[0023] An MPC controller and tracked vehicle model were built in the Simulink environment. Paths for various continuous steering conditions were designed, and expert demonstration data was collected through path tracking simulation. The input of the expert demonstration data included path tracking error. The output sample collects the torque of the left and right drive motors. .

[0024] Step Two: Data Collection Based on the Expert Demonstration Database Considering the potential for extreme values ​​in the initial control input due to overshoot, and the differences in control input magnitude between turning and straight-line driving conditions, expert demonstration data is preprocessed to improve the stability and convergence speed of neural network training. First, all driving conditions are divided into training, validation, and test sets in a 3:1:1 ratio. Then, because the data is non-normally distributed and contains outliers, robust scaling is used to standardize the input and output data. This method is less sensitive to outliers and noise, scaling the data based on the median and interquartile range, unlike traditional methods which have strict requirements on data distribution and rely on the mean, standard deviation, or maximum / minimum values. The formula for robust scaling is as follows: The median of the characteristic column. The interquartile range is equal to the 75th percentile. Subtract the 25th percentile : The robust scaling process is as follows: In order to prevent data leakage, a robust scaler is first fitted to the input and output data of the training set, respectively, and the median and interquartile range are calculated. Then, the fitted scaler is used to perform the same scaling transformation on the training set, validation set and test set. Figure 1 The diagram illustrates the robust scaling results, where (a) represents the input features before preprocessing, (b) represents the input features after preprocessing, (c) represents the output features before preprocessing, and (d) represents the output features after preprocessing. The input features before preprocessing fluctuate due to the forward movement of the expected path tracking point, and the curve range is narrow (-1.0 to 0.3), which may mask the actual trend of change. The scaled input features have a wider range, from -12.5 to 7.5, and the peaks of dx around 20 and 40 seconds are clearly preserved, reflecting the true dynamics. The fluctuation range is amplified year-on-year, but it is not compressed due to the sharp fluctuations in dx. This avoids the distortion of the overall data by outliers while amplifying meaningful fluctuations. The output feature range before preprocessing is very large (-1200 to 1800), with obvious outliers around 0 seconds and 40 seconds; the range of the scaled output feature is relatively small, approximately -0.6 to 0.6, the influence of outliers is weakened, and the overall distribution is more uniform. Among them, the input features before preprocessing... The unit is [m]. The unit is [rad]; since the values ​​after robust scaling are standardized relative values, the units of the preprocessed input and output features are not labeled.

[0025] Then, the long sequences of operating conditions in each set are divided into multiple partially overlapping short sequences to form the final training set, validation set, and test set. This approach increases the amount of data and improves the training effect of the time series prediction model. Compared to completely non-overlapping sequences, the partially overlapping approach allows adjacent short sequences to share some time steps, similar to the rolling optimization window of MPC, which helps the model capture smooth transitions and continuous dynamics between sequences. At the same time, the overlap allows the model to encounter similar state transitions multiple times, which can enhance the robustness of the model in filtering noise.

[0026] Step 3: Building and Training the Liquid Time Constant Neural Network Control Model The LTCNN path tracking control model was built in the PyTorch framework and trained using the AdamW optimizer with MPC behavior cloning. The loss function was the mean square error between the model's predicted values ​​and the expert's demonstrated values ​​of the drive motor torque. The specific training parameters are shown in Table 1.

[0027] Table 1 Training parameters of the path tracking control model

[0028] LTCNN uses a fusion solver for forward propagation, and its solution process is as follows: For the trajectory arrive The system will span consecutive time periods Discretized Equally spaced time points Calculate only each time arrive The local trajectory changes are updated iteratively from arrive The hidden states of neurons are ultimately obtained. The system state output at time t. It is worth noting that when the number of discrete points... When the time complexity is sufficiently large, this piecewise linear approximation will approach the true continuous-time analytical solution infinitely. Therefore, LTCNN fuses explicit and implicit Euler functions, using sufficiently fine-grained time discretization to accurately fit its nonlinear characteristics. The fused ODE solver is then used to solve the ordinary differential equations of LTCNN:

[0029] in It is a time constant. In dynamic state, For system input, The reference potential, specifically, means that at... Enter the following to hide the status. Regarding time The derivative is used by Parameterized Neural Networks Describe it.

[0030] Transform it into discrete form: in, The time interval is denoted by . Based on equation (4.6), the forward propagation of LTCNN can be described by Table 2.

[0031] Table 2 LTCNN Forward Propagation Algorithm

[0032] The LTCNN path tracking control model is built using the NCP (Neural Circuit Policies) architecture. NCP is a compact and sparse neural network architecture inspired by the topology of the nematode nervous system, employing a four-layer network. Neurons are functionally classified into sensory neurons, interneurons, command neurons, and motor neurons. These neural layers, composed of neurons with different functions, sequentially receive environmental input, process information, generate control decisions, and output control commands, resulting in good interpretability. Furthermore, the connections between the same number of neurons are sparser, with fewer synapses, enabling efficient representation learning. Figure 2For comparison of wiring methods, straight arrows pointing from one neuron to another indicate the direction of information transmission. Arrows circling back to a neuron indicate that the neuron's output at the current time step will be fed back as input for the next time step, a core characteristic of recurrent neural networks. Compared to the NCP method, the fully connected and random sparse methods are not internally layered, consisting entirely of intermediate neurons. The fully connected method has the most synapses, and even at the same sparsity level, the random sparse method has more synapses than the NCP method. Given the superiority of the NCP wiring method, the LTCNN path tracking model adopts the NCP wiring method, with a neuron count of 16 (excluding sensory neurons, whose number corresponds to the input feature dimension of 3). This includes 9 intermediate neurons, 5 command neurons, and 2 motor neurons (corresponding to the output feature dimension).

[0033] The effects of this invention are illustrated below using simulation data experiments: Figure 3 The results of LTCNN behavior cloning under continuous steering conditions with small curvature are shown in Figure (a). The tracked vehicle tracks the desired path with small curvature at a speed of 20 km / h. Under the control of the LTCNN path tracking control algorithm based on MPC behavior cloning, the tracked vehicle can travel well along the desired path, with an RMSE of 0.089 m for the lateral error (dy) of path tracking, indicating high control accuracy. Figures (b) and (c) show the path tracking error input of the controller and the torque output of the drive motors during the driving process. The trained LTCNN can well fit the torque outputs of the two drive motors given by MPC under the path tracking error input under this small curvature condition, with RMSEs of 10.32 N·m and 11.06 N·m, respectively. The predicted torque curve deviates little from MPC at the peak, with fluctuations less than 20 N·m. Furthermore, the output is more stable than MPC, mitigating the severe torque fluctuations caused by the forward shift of the desired path point.

[0034] Figure 3 Performance of LTCNN Path Tracking Control Based on MPC Behavior Cloning (Small Curvature Condition) Figure 4To demonstrate the effectiveness of LTCNN behavior cloning under continuous steering conditions with high curvature, Figure (a) shows the path tracking performance. The tracked vehicle, traveling at 20 km / h, tracks the desired path with high curvature. Under the control of the LTCNN path tracking control algorithm based on MPC behavior cloning, the tracked vehicle can generally travel along the desired path, with an RMSE of 0.13 m for the lateral error (dy) of path tracking, indicating high control accuracy. Figures (b) and (c) show that the trained LTCNN can better fit the torque output of the two drive motors given by MPC under the path tracking error input in this high curvature condition, with RMSEs of 20.63 N·m and 30.29 N·m, respectively. The predicted torque curve deviates little from the MPC at its peak, with fluctuations less than 50 N·m, and the output torque is more stable than that of MPC.

[0035] Real-time performance was evaluated by runtime. The average simulation time for the two path tracking scenarios was recorded, and the results are shown in Table 3. It can be seen that LTCNN has better real-time performance, taking 124.8 seconds, saving 34.8% of runtime compared to MPC. This is because MPC needs to solve an optimization problem involving high-dimensional state space and constraints in each control cycle, resulting in high computational complexity. In contrast, neural network control based on behavior cloning achieves fast response through forward inference, resulting in higher computational efficiency. Furthermore, the efficiency of MPC is affected by the prediction time domain; the longer the prediction time domain and the larger the optimization problem, the more significant the computational delay.

[0036] Table 3 Real-time performance evaluation of control algorithm

[0037] This invention ensures path tracking accuracy through MPC behavior cloning, with a lateral error RMSE of 0.089m under small curvature conditions and 0.13m under large curvature conditions; LTCNN forward inference improves real-time performance, saving 34.8% of runtime compared to MPC; data preprocessing and NCP architecture enhance model robustness and interpretability, resulting in smoother output torque, making it suitable for tracked vehicle path tracking in complex off-road environments.

[0038] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A track vehicle path tracking method based on a liquid time constant neural network, characterized by, Includes the following steps: Step 1: Establish an expert demonstration database based on model predictive control algorithms: Obtain the position coordinates of the desired path, calculate the desired heading angle based on the geometric relationship between the coordinate points, and obtain the path tracking error by combining the vehicle's driving status. Based on the kinematic model of the tracked vehicle with instantaneous steering center, an MPC controller is designed. The control quantity is determined to be the winding speed of the left and right tracks. The corresponding vehicle speed and yaw rate are tracked by the PID algorithm to obtain the torque of the left and right drive motors. An MPC controller and a tracked vehicle model were built in the Simulink environment. Paths for various continuous steering conditions were designed. Expert demonstration data were collected through path tracking simulation. The input of the expert demonstration data was the path tracking error, and the output was the torque of the left and right drive motors. Step Two: Data Collection Based on the Expert Demonstration Database The expert demonstration data is divided into training set, validation set and test set in a ratio of 3:1:1; A robust scaling method is used to standardize the input and output data. First, a robust scaler is fitted to the input and output data of the training set, and the median and interquartile range of the data are calculated. Then, the fitted scaler is used to perform the same scaling transformation on the training set, validation set and test set. The long sequences of working conditions in each set are divided into multiple short sequences with some overlap to form the final training set, validation set, and test set. Step 3: Build and train the liquid time constant neural network control model: An LTCNN path tracking control model was built in the PyTorch framework. The model was wired using the NCP neural circuit architecture, which includes sensory neurons, interneurons, command neurons, and motor neurons. The model was trained using the AdamW optimizer with MPC behavior cloning. The training loss function was the mean square error between the model's predicted value and the expert demonstration value of the drive motor torque. Step 4: Tracked vehicle path tracking control: Input the real-time tracked vehicle path tracking error into the trained LTCNN path tracking control model. The model outputs the torque of the left and right drive motors, and controls the tracked vehicle to track the desired path based on this torque.

2. The method for tracking tracked vehicles based on a liquid time constant neural network according to claim 1, characterized in that, The formula for calculating robust scaling in step two is as follows: ,in The median of the characteristic column. The interquartile range is equal to the 75th percentile. Subtract the 25th percentile .

3. The method for tracking tracked vehicles based on a liquid time constant neural network according to claim 1, characterized in that, The partially overlapping short sequences mentioned in step two allow adjacent short sequences to share some time steps, simulating the rolling optimization window of MPC and enhancing the model's ability to capture the continuous dynamics of the sequence.

4. The method for tracking tracked vehicles based on a liquid time constant neural network according to claim 1, characterized in that, In step three, the LTCNN path tracking model uses NCP wiring, with a neuron count of 16 (excluding sensory neurons). The number of intermediate neurons is 9, command neurons are 5, and motor neurons are 2, which is the output feature dimension. The neurons are functionally layered to achieve signal transmission and processing.

5. The method for tracking tracked vehicles based on a liquid time constant neural network according to claim 1, characterized in that, In step three, the parameters for model training are set as follows: exponential decay rate of the first moment is 0.9, exponential decay rate of the second moment is 0.999, learning rate of LTCNN is 1e-3, weight decay coefficient is 0.01, training sequence length is 50, and number of iterations is 4000.

6. The method for tracking tracked vehicles based on a liquid time constant neural network according to claim 1, characterized in that, In step three, LTCNN performs forward propagation by fusing the ODE solver. The solution process is as follows: for the trajectory... arrive The system will span consecutive time periods Discretized Equally spaced time points Calculate only each time arrive The local trajectory changes are updated iteratively from arrive The hidden states of neurons are ultimately obtained. The system state output at time t is noteworthy; it's important to note that when the number of discrete points... When the size is large enough, the result of this piecewise linear approximation will approach the true continuous-time analytical solution infinitely. Therefore, LTCNN combines explicit Euler and implicit Euler and uses sufficiently fine time discretization to accurately fit its nonlinear features.

7. The method for tracking tracked vehicles based on a liquid time constant neural network according to claim 6, characterized in that, Use the fused ODE solver to solve the ordinary differential equations of LTCNN: ,in It is a time constant. For dynamic state, For system input, The reference potential, specifically, means that at... Enter the following to hide the status. Regarding time The derivative is used by Parameterized Neural Networks Describe it; Transform it into discrete form: ， in, For time intervals.

8. The method for tracking tracked vehicles based on a liquid time constant neural network according to claim 1, characterized in that, In step one, obtain the location coordinates of the desired path. The desired heading angle is calculated based on the geometric relationship before and after the coordinate point. , will the desired path information With the vehicle's driving status Combined with the obtained path tracking error .

9. The method for tracking tracked vehicles based on a liquid time constant neural network according to claim 1, characterized in that, In step three, the MPC controller is designed using a tracked vehicle kinematic model based on the instantaneous steering center, according to... To calculate the optimal control quantity According to kinematic relationships, the longitudinal and yaw motions of the vehicle, i.e., vehicle speed and yaw rate, can be affected by changing the winding speed of the left and right tracks. Therefore, the MPC control variable is determined to be... ,in These represent the left and right track winding speeds, respectively. To ensure consistent control, after the MPC obtains the desired track winding speed, a PID algorithm is used to track the corresponding vehicle speed and yaw rate to obtain the torque of the left and right drive motors. .

10. The method for tracking tracked vehicles based on a liquid time constant neural network according to claim 1, characterized in that, In step three, the functional division of neurons in each layer of the NCP neural circuit architecture is clear: sensory neurons receive path tracking error input, intermediate neurons perform information processing and feature extraction, command neurons generate control decisions, and motor neurons output torque commands to drive the motor.

Images