Interference and trajectory learning method based on integrated neural network

Abstract

The invention relates to the field of machine learning, in particular to an interference and trajectory learning method based on an integrated neural network, which comprises the following steps: (1) preprocessing multi-source input data in a collaborative scene, camera depth images carried by the platforms, real-time motion state data and collaborative task expected motion target data are collected in the simulator; performing noise reduction and feature extraction preprocessing on the visual data, and performing standardization processing on a motion state and expected target data; and (2) integrated neural network construction and training optimization: constructing an end-to-end integrated neural network, determining that the network input is the preprocessed multi-source data, and the output is an expected moving trajectory of each platform and an external time-varying interference prediction result. According to the method, a mapping link and a multi-module connection step in a traditional method are omitted through an end-to-end architecture, so that the calculation time is greatly saved, and collaborative optimization of'trajectory planning-interference prediction 'is realized at the same time.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of machine learning, in particular to an interference and trajectory learning method based on an integrated neural network. BACKGROUND

[0002] Today, mobile platforms such as unmanned aerial vehicles and robotic dogs are increasingly widely used in complex scenarios such as inspection exploration, material transportation, and emergency rescue, which puts higher requirements on the accuracy of platform motion trajectories, the adaptability to external interference, and the efficiency of control processes. However, in traditional solutions, "trajectory planning" and "interference prediction" are usually independent modules: trajectory planning relies on pre-set models or step-by-step mapping, and does not fully consider real-time interference; interference prediction is based solely on motion data to deduce, which is separate from trajectory generation, resulting in the inability to optimize both simultaneously. This separation mode not only increases the computational delay of multi-module data interaction, but also makes it difficult for the generated trajectory to adapt to a dynamic interference environment, limiting the anti-interference performance.

[0003] In existing technologies, some methods attempt to combine visual information and motion state, but either focus on a single task or use a multi-stage processing flow, failing to achieve end-to-end output from "multi-source input to trajectory and interference results". At the same time, most solutions are designed for a single platform, lacking generality for different types of platforms such as unmanned aerial vehicles and robotic dogs. Therefore, how to build an integrated framework to simultaneously output expected trajectories and accurate interference prediction results that meet safety requirements through an end-to-end mode, achieving deep collaborative optimization of both, has become a core challenge to improve the performance of multiple platforms in complex interference environments.

[0004] Although the "Medical Rescue Unmanned Aerial Vehicle and Robotic Dog Cooperative Linkage Method" with patent publication number CN120351922A also targets scenarios where unmanned aerial vehicles and robotic dogs work together, it integrates multi-modal data through time alignment and spatial alignment techniques, constructs a global semantic map using parallel SLAM calculation and GTSAM library graph optimization, and then plans a dynamic path for both platforms based on the map. The core focus is on "multi-modal mapping - path planning based on the map", and it does not use data-driven learning methods such as neural networks.

[0005] The "Robotic Dog Intelligent Gait Planning Method Based on Deep Learning" with patent publication number CN120373348A uses an optimized deep learning system to correct and learn gait data, achieving planning and flexibility improvement of robotic dog walking posture, but this method only focuses on gait optimization of robotic dogs, cannot obtain expected trajectories, and cannot actively learn the rules of external time-varying interference, lacking adaptability to motion stability in complex interference environments.

[0006] The patent with the publication number CN120773068A discloses a fusion control method and system for mechanical arm flaw detection operation during machine dog climbing, which combines multi-round breeding evolution optimization to generate a mechanical arm arm-out control strategy, and obtains a disturbance compensation strategy through disturbance triggering prediction and optimization compensation. However, this method only focuses on the specific operation scene of "climbing-flaw detection" of a single platform of the machine dog, and cannot actively learn the time-varying law and characteristics of the disturbance. SUMMARY

[0007] To solve the problem that the trajectory planning and disturbance prediction are separated, the multi-module processing flow is redundant, the calculation time is high, and the end-to-end collaborative optimization is difficult to realize in the existing method, the present application provides a disturbance and trajectory learning method based on an integrated neural network, which directly outputs the expected trajectory and disturbance prediction result by constructing an end-to-end integrated framework with multi-source perception data (visual image, motion state) in synchronization, without the need for an independent mapping link, and realizes the collaborative back propagation optimization of the two by means of a joint loss function, thereby significantly improving the response speed and anti-disturbance performance of the multi-platform in a complex environment.

[0008] To achieve the above-mentioned purposes, the technical solutions adopted by the present application are as follows:

[0009] A disturbance and trajectory learning method based on an integrated neural network, comprising the following steps:

[0010] (1) Multi-source input data preprocessing in a collaborative scene: for the cooperative task of the unmanned aerial vehicle and the machine dog, collect the camera depth image, real-time motion state data and cooperative task expected motion target data carried by each platform in the simulator; perform noise reduction and feature extraction preprocessing on the visual data, and standardize the motion state and expected target data;

[0011] (2) Integrated neural network construction and training optimization: an end-to-end integrated neural network is constructed, the network input is the preprocessed multi-source data, and the output is the expected running trajectory of each platform and the external time-varying disturbance prediction result; in the trajectory prediction link, the network directly outputs the position point in the image coordinate system, the minimum energy trajectory from the current position of the platform to the target position is calculated based on the position point, and the corresponding trajectory score is generated; in the disturbance prediction link, the deviation between the disturbance prediction value output by the network and the actual disturbance value is taken as the disturbance loss term; the network simultaneously optimizes the trajectory score and the disturbance estimation error, and constructs a "trajectory deviation-disturbance prediction deviation" joint loss function; the preprocessed multi-source data is input into the network, and the joint loss function is taken as the optimization target to iteratively adjust the network weight parameters; through continuous training, the expected trajectory output by the network meets the safety requirement, the disturbance prediction accuracy is improved, and the trajectory planning process is deeply integrated with the disturbance information.

[0012] Further, the step (1) is specifically as follows:

[0013] (1.1) First, build a three-dimensional task scene map with a size of in the simulator, which is used to simulate the physical environment of the unmanned aerial vehicle and the robot dog working together; for each data collection sample in the collaborative task, randomly configure the initial motion state parameters of the platform;

[0014] (1.2) Generate a polynomial expected trajectory with the "start-end" position, velocity, and acceleration constraints as boundary conditions;

[0015] (1.3) In the time dimension of the expected trajectory, generate a third-order continuous disturbance force acting on the base of the unmanned aerial vehicle / robot dog, which satisfies the second-order derivative continuity characteristic to simulate the slowly changing and smooth external disturbance in the actual environment (such as air flow disturbance, ground friction change disturbance), and the disturbance force amplitude range is set according to the carrying capacity of the platform to ensure that the disturbance scene conforms to the actual application scene;

[0016] (1.4) Take the trajectory data of 10 continuous dynamics iteration steps as a sample, collect the complete actual motion state data in 10 steps: and the actual disturbance force at the tenth time step; then build a random forest obstacle in the simulation map, and establish an Euclidean signed distance field (ESDF) map combined with the geometric size parameters of the platform, and use a collision detection algorithm to determine whether the sample trajectory collides with the random forest obstacle: if there is a collision, the sample is removed, and only the valid samples without collision are retained; then add the depth image shot by the camera in each sample;

[0017] Further, the step (1.1) is as follows:

[0018] Randomly generate the position vector of the starting point of the platform , the starting point velocity vector and the starting point acceleration vector ; randomly set the straight line distance from the starting point to the end point, and randomly generate the trajectory direction vector (the direction vector has a length of 1 to ensure randomness), calculate the end point position based on the starting point position , distance and direction vector , set the end point velocity vector and end point acceleration vector to ensure smooth start and end states of the trajectory.

[0019] Furthermore, step (1.2) is as follows:

[0020] For uniaxial motion, a fifth-order polynomial is used to describe the relationship between position and time. , , The axes are solved independently, with Taking the axis as an example, the expression is:

[0021] ;

[0022] in, for Location at any given moment For polynomial coefficients, The total duration of the trajectory is defined as follows: ,in Given the maximum velocity; substitute the boundary conditions into the position and derivative formulas to construct the coefficient matrix equation:

[0023] ;

[0024] Solving for the coefficient vector, for a three-dimensional trajectory, , , Each axis is independently modeled using the aforementioned 5th-order polynomial model. The polynomial coefficients of each axis are solved using the same method, and the coefficient vector matrix is ​​calculated. To obtain the three-dimensional trajectory ;

[0025] By analyzing the desired trajectory Find the first derivative to obtain the desired speed The desired acceleration is obtained by taking the second derivative. This forms a complete sequence of desired motion states.

[0026] Furthermore, step (1.3) is as follows:

[0027] The desired motion state sequence and third-order continuous disturbance force Input the dynamics model of the drone / robot dog and set the dynamics iteration time step. The actual motion state of the drone / robot dog base under disturbance is calculated in real time using dynamic equations. ,in The attitude of the drone / robot dog base.

[0028] Furthermore, step (1.4) is as follows:

[0029] Actual velocity data in the valid sample Actual acceleration data Standardization process: , ;

[0030] in These are the minimum and maximum values ​​of the velocity data among all valid samples. These are the minimum and maximum values ​​of acceleration data among all valid samples, respectively.

[0031] Furthermore, step (2) is as follows:

[0032] (2.1) Construct an integrated neural network, adopting a "multimodal feature fusion - dual-task parallel output" structure;

[0033] (2.2) The definition of network learning loss value includes two items: trajectory loss and interference loss.

[0034] Furthermore, step (2.1) is as follows:

[0035] 1) Visual feature processing: processing depth images ResNet network is used for feature extraction, and high-level visual features such as the distribution and relative position of environmental obstacles are extracted through convolutional layers and residual connections.

[0036] 2) Motion state processing: The standardized velocity, acceleration, and attitude data are processed using 1D convolutional layers to capture the temporal variation patterns of the motion state;

[0037] 3) Feature Fusion and Output: Visual features and motion state features are fused through a cross-modal attention mechanism to obtain the following outputs: one is an estimate of external time-varying disturbances. Secondly, the velocity acceleration at the endpoint of the desired trajectory. .

[0038] Furthermore, step (2.2) is as follows:

[0039] Trajectory loss term: based on network output Based on the current platform location As the starting state, the corresponding trajectory is generated using the 5th-order polynomial defined in step (1.1);

[0040] Calculate in three directions respectively, with Taking direction as an example, for time Current position for: , ;

[0041] The defined safe distance function for each location is:

[0042] ;

[0043] in It is the closest distance to the obstacle calculated using an ESDF map. Given a distance threshold, The coefficients are given; therefore, the loss term for the entire curve is... Defined as:

[0044] ; ;

[0045] Interference loss item Interference estimate of network output Compared with actual interference value of Loss function:

[0046] ;

[0047] Joint learning: building a system that includes and The joint loss function (which balances the effects of the two losses through weight coefficients) feeds the preprocessed input data into the network for iterative training. The network parameters are optimized through backpropagation, which enables the network to improve the accuracy of trajectory endpoint prediction and interference estimation at the same time, realizing end-to-end collaborative learning of "motion state-environment perception-trajectory generation and interference prediction".

[0048] In summary, the present invention has at least one of the following beneficial technical effects:

[0049] This invention discloses an interference and trajectory learning method based on an integrated neural network, comprising multi-source input data preprocessing in a collaborative scenario, construction and training optimization of an integrated neural network. For collaborative tasks involving UAVs and robotic dogs, depth images, real-time motion states (velocity, acceleration, attitude), and desired moving target data are collected in a simulation environment. After feature extraction and standardization, an effective dataset is obtained. An end-to-end integrated neural network is constructed, using the preprocessed multi-source data as input, and simultaneously outputting the desired trajectory endpoint and external time-varying interference estimates. A joint loss function is constructed using trajectory score deviation and interference prediction error. Network parameters are simultaneously optimized through backpropagation, enabling the network to accurately generate safe and feasible trajectories and predict interference in real time.

[0050] The advantages of this invention compared to existing technologies are as follows: This invention achieves end-to-end inference of "multi-source input - trajectory and interference dual output" through an integrated neural network, eliminating the independent environment mapping stage and data interaction steps between trajectory planning and interference prediction modules in traditional methods. This significantly reduces computational redundancy and processing time, making it more suitable for the real-time requirements of multi-platform collaborative tasks. Simultaneously, by employing a joint loss function, trajectory safety optimization and interference prediction accuracy improvement are incorporated into the same backpropagation process, allowing them to mutually constrain and collaboratively optimize each other during parameter updates. Trajectory generation directly incorporates interference information to improve anti-interference adaptability, while interference prediction further corrects accuracy through trajectory deviation feedback. Furthermore, the network exhibits strong compatibility with input features; only the local input dimension needs to be adjusted according to the differences in motion states between the UAV and the robot dog, without reconstructing the network structure, enabling rapid cross-platform deployment and effectively expanding the applicability of the method. Attached Figure Description

[0051] Figure 1 This is a flowchart of the interference and trajectory learning method based on an integrated neural network according to the present invention.

[0052] Figure 2 The image shows the interference prediction diagram of the quadcopter drone used for method verification in a specific embodiment.

[0053] Figure 3 The image shows the trajectory tracking diagram of the quadcopter drone used for method verification in a specific embodiment.

[0054] Figure 4 The image shows the interference prediction graph of the robot dog used for method verification in a specific embodiment.

[0055] Figure 5 The image shows the trajectory tracking diagram of the robot dog used for method verification in a specific embodiment. Detailed Implementation

[0056] A method for learning interference and trajectory based on an integrated neural network includes the following steps:

[0057] (1) Preprocessing of multi-source input data in collaborative scenarios: For the collaborative work tasks of UAVs and robot dogs, depth images of cameras carried by each platform, real-time motion state data and expected motion target data of collaborative tasks are collected in the simulator; the visual data is denoised and feature extracted for preprocessing, and the motion state and expected target data are standardized.

[0058] (2) Construction and training optimization of integrated neural network: An end-to-end integrated neural network is constructed, and the network input is defined as preprocessed multi-source data, and the output is the expected running trajectory of each platform and the prediction result of external time-varying interference. In the trajectory prediction stage, the network directly outputs the position point in the image coordinate system, calculates the minimum energy trajectory from the current position of the platform to the target position based on the position point, and generates the corresponding trajectory score. In the interference prediction stage, the deviation between the interference prediction value output by the network and the actual interference value is used as the interference loss term. The network optimizes the trajectory score and the interference estimation error at the same time and constructs a joint loss function of "trajectory deviation - interference prediction deviation". The preprocessed multi-source data is input into the network, and the network weight parameters are iteratively adjusted with the joint loss function as the optimization target. Through continuous training, the expected trajectory output by the network meets the safety requirements, while improving the interference prediction accuracy and ensuring that the interference information is deeply integrated into the trajectory planning process.

[0059] Furthermore, step (1) is as follows:

[0060] (1.1) First, build a simulator with a size of [size missing]. A 3D mission scenario map is used to simulate the physical environment of the drone and robot dog working together; for each data acquisition sample in the collaborative mission, the initial motion state parameters of the platform are randomly configured;

[0061] (1.2) Using the position, velocity, and acceleration constraints of the "start-end point" as boundary conditions, generate the polynomial expected trajectory. ;

[0062] (1.3) Generate a third-order continuous disturbance force acting on the UAV / robot dog base in the time dimension of the desired trajectory. The interference force satisfies the continuous characteristic of the second derivative to simulate the slowly changing and smooth external interference in the actual environment (such as airflow interference and ground friction change interference). The amplitude range of the interference force is set according to the platform's carrying capacity to ensure that the interference scenario conforms to the actual application scenario.

[0063] (1.4) Using trajectory data from 10 consecutive dynamic iteration steps as a sample, collect complete actual motion state data within 10 steps: and the actual interference force at the tenth time step Then, a random forest obstacle is constructed in the simulation map, and an Euclidean symbolic distance field (ESDF) map is established. Combined with the platform's geometric dimensions, a collision detection algorithm is used to determine whether the sample trajectory collides with the random forest obstacle: if a collision occurs, the sample is discarded, and only valid samples without collisions are retained; then, a depth image captured by the camera is added to each sample. ; Obtain complete sample data.

[0064] Furthermore, step (1.1) is as follows:

[0065] Randomly generate the position vector of the platform's starting point. Starting velocity vector and the initial acceleration vector Randomly set the straight-line distance from the starting point to the ending point of the trajectory. , Simultaneously, a trajectory direction vector is randomly generated. (The direction vector has a magnitude of 1 to ensure randomness of direction), based on the starting position. ,distance and direction vector Calculate the endpoint position Set the final velocity vector Final acceleration vector This ensures a stable start and end point for the trajectory.

[0066] Furthermore, step (1.2) is as follows:

[0067] For uniaxial motion, a fifth-order polynomial is used to describe the relationship between position and time. , , The axes are solved independently, with Taking the axis as an example, the expression is:

[0068] ;

[0069] in, for Location at any given moment For polynomial coefficients, The total duration of the trajectory is defined as follows: ,in Given the maximum velocity; substitute the boundary conditions into the position and derivative formulas to construct the coefficient matrix equation:

[0070] ;

[0071] Solving for the coefficient vector, for a three-dimensional trajectory, , , Each axis is independently modeled using the aforementioned 5th-order polynomial model. The polynomial coefficients of each axis are solved using the same method, and the coefficient vector matrix is ​​calculated. To obtain the three-dimensional trajectory ;

[0072] By analyzing the desired trajectory Find the first derivative to obtain the desired speed The desired acceleration is obtained by taking the second derivative. This forms a complete sequence of desired motion states.

[0073] Furthermore, step (1.3) is as follows:

[0074] The desired motion state sequence and third-order continuous disturbance force Input the dynamics model of the drone / robot dog and set the dynamics iteration time step. The actual motion state of the drone / robot dog base under disturbance is calculated in real time using dynamic equations. ,in The attitude of the drone / robot dog base.

[0075] Furthermore, step (1.4) is as follows:

[0076] Actual velocity data in the valid sample Actual acceleration data Standardization process: , ;

[0077] in These are the minimum and maximum values ​​of the velocity data among all valid samples. These are the minimum and maximum values ​​of acceleration data among all valid samples, respectively.

[0078] Furthermore, step (2) is as follows:

[0079] (2.1) Construct an integrated neural network, adopting a "multimodal feature fusion - dual-task parallel output" structure;

[0080] (2.2) The definition of network learning loss value includes two items: trajectory loss and interference loss.

[0081] Furthermore, step (2.1) is as follows:

[0082] 1) Visual feature processing: processing depth images ResNet network is used for feature extraction, and high-level visual features such as the distribution and relative position of environmental obstacles are extracted through convolutional layers and residual connections.

[0083] 2) Motion state processing: The standardized velocity, acceleration, and attitude data are processed using 1D convolutional layers to capture the temporal variation patterns of the motion state;

[0084] 3) Feature Fusion and Output: Visual features and motion state features are fused through a cross-modal attention mechanism to obtain the following outputs: one is an estimate of external time-varying disturbances. Secondly, the velocity acceleration at the endpoint of the desired trajectory. .

[0085] Furthermore, step (2.2) is as follows:

[0086] Trajectory loss term: based on network output Based on the current platform location As the starting state, the corresponding trajectory is generated using the 5th-order polynomial defined in step (1.1);

[0087] Calculate in three directions respectively, with Taking direction as an example, for time Current position for:

[0088] , ;

[0089] The defined safe distance function for each location is:

[0090] ;

[0091] in It is the closest distance to the obstacle calculated using an ESDF map. Given a distance threshold, The coefficients are given; therefore, the loss term for the entire curve is... Defined as:

[0092] ; ;

[0093] Interference loss item Interference estimate of network output Compared with actual interference value of Loss function:

[0094] ;

[0095] Joint learning: building a system that includes and The joint loss function (which balances the effects of the two losses through weight coefficients) feeds the preprocessed input data into the network for iterative training. The network parameters are optimized through backpropagation, which enables the network to improve the accuracy of trajectory endpoint prediction and interference estimation at the same time, realizing end-to-end collaborative learning of "motion state-environment perception-trajectory generation and interference prediction".

[0096] The following specific embodiments illustrate the implementation principle of the present invention:

[0097] The experiment focused on verifying the interference estimation and trajectory tracking performance of the quadcopter drone and the robot dog.

[0098] Experimental interference settings: The quadcopter drone is subjected to three-axis external interference forces with an amplitude range of 0.5~4.5N, including two types of interference: step type and sinusoidal wave type; the robot dog is subjected to planar (X, Y direction) external interference forces with an amplitude range of 2.0~10.0N.

[0099] Trajectory planning constraints: The motion map size used in the experiment was 40m×40m×5m (quadcopter, 3D space) and 40m×40m (robot dog, planar space); the quadcopter UAV's expected trajectory had a maximum speed range of 0.5~2.0m / s, a maximum acceleration range of 0.3~1.0m / s², a starting position range of (0~1m, 0~1m, 0.5~1m), and an ending position range of (35~40m, 35~40m, 2~3m); the robot dog's expected trajectory had a maximum speed range of 0.3~1.2m / s and an expected yaw rate of 0.1~0.7rad / s.

[0100] Test results are as follows Figures 2 to 5 As shown.

[0101] Figure 2 The results show a comparison between the actual and estimated values ​​of the three-axis external disturbance force. The estimated curves in the X, Y, and Z directions basically coincide with the actual curves in the entire time domain, and the peak value and phase changes can be tracked synchronously without significant divergence or long-term lag. The results demonstrate that the proposed method can achieve stable and low-bias estimation of the three-axis disturbance force under dynamic disturbance conditions.

[0102] like Figure 3 As shown, the desired trajectory exhibits a continuous and smooth trend across all coordinate axes: no abrupt changes or sharp peaks occur throughout the time domain, and the transitions between trajectory segments are natural. The proposed integrated neural network can improve the stability and low bias characteristics of three-axis disturbance estimation during the learning process (corresponding to...). Figure 2 On the one hand, it reduces the error index, and on the other hand, it reduces the oscillation risk of trajectory generation, so that the actual trajectory can maintain a high degree of consistency with the expected trajectory during dynamic phases such as acceleration and turning. Figure 3 (As shown in the close fit). This achieves the following: under the condition of external disturbance, the system's trajectory tracking accuracy and motion smoothness are improved through the collaborative mechanism of "disturbance estimation compensation + integrated trajectory generation".

[0103] Figure 4The comparison results of the "actual value - estimated value" of external disturbance force in the plane (X, Y) are presented for the robot dog. It can be seen that the estimated curves in both the X and Y directions can synchronously track the amplitude change and phase evolution of the actual disturbance in the entire time domain, and the overall trend is consistent with the local peak and valley positions, without obvious divergence or persistent lag. This shows that the proposed method can achieve stable and low-bias plane disturbance estimation during the robot dog's dynamic movement.

[0104] Figure 5 The position tracking results of the robot dog are presented. The expected trajectory exhibits a continuous and smooth trend in both the X and Y axes, with natural transitions between trajectory segments. The actual position curve and the expected position curve maintain a high degree of consistency throughout the entire time domain, closely matching each other even during dynamic changes such as trajectory acceleration and turning, without significant oscillations or overshoot. Figure 2 It can be seen that, under a unified integrated neural network framework, the tracking accuracy and motion smoothness of the robot dog can be effectively improved by estimating external disturbances online and using them for compensation / trajectory generation.

[0105] Experimental results from both a quadcopter platform (three-axis perturbation estimation and 3D position tracking) and a robot dog platform (planar perturbation estimation and planar position tracking) demonstrate that the proposed method achieves consistent results of "trackable perturbation, low estimation bias, and close and smooth trajectory" on platforms with different dynamic structures, motion constraints, and perturbation dimensions (3D / 2D). This indicates that the method does not rely on the specific structure of a particular platform or a single operating condition assumption, possessing strong cross-platform applicability and universality, and can be extended to various types of mobile robots for perturbation estimation and trajectory tracking tasks under external perturbation conditions.

[0106] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made according to the structure, shape, and principle of the present invention should be covered within the scope of protection of the present invention. Contents not described in detail in this specification are prior art known to those skilled in the art.

Claims

1. A method for learning interference and trajectory based on an integrated neural network, characterized in that, Includes the following steps: (1) Preprocessing of multi-source input data in collaborative scenarios: For the collaborative work tasks of UAVs and robot dogs, depth images of cameras carried by each platform, real-time motion state data and expected motion target data of collaborative tasks are collected in the simulator; the visual data is denoised and feature extracted for preprocessing, and the motion state and expected target data are standardized. (2) Construction and training optimization of integrated neural network: An end-to-end integrated neural network is constructed, and the network input is defined as preprocessed multi-source data, and the output is the expected running trajectory of each platform and the prediction result of external time-varying interference. In the trajectory prediction stage, the network directly outputs the position point in the image coordinate system, calculates the minimum energy trajectory from the current position of the platform to the target position based on the position point, and generates the corresponding trajectory score. In the interference prediction stage, the deviation between the interference prediction value output by the network and the actual interference value is used as the interference loss term. The network optimizes the trajectory score and the interference estimation error at the same time and constructs a joint loss function of "trajectory deviation - interference prediction deviation". The preprocessed multi-source data is input into the network, and the network weight parameters are iteratively adjusted with the joint loss function as the optimization target. Through continuous training, the expected trajectory output by the network meets the safety requirements, while improving the interference prediction accuracy and ensuring that the interference information is deeply integrated into the trajectory planning process.

2. The interference and trajectory learning method based on an integrated neural network according to claim 1, characterized in that, The specific steps (1) are as follows: (1.1) First, build a simulator with a size of [size missing]. A 3D mission scenario map is used to simulate the physical environment of the drone and robot dog working together; for each data acquisition sample in the collaborative task, the initial motion state parameters of the platform are randomly configured; (1.2) Using the position, velocity, and acceleration constraints of the "start-end point" as boundary conditions, generate the polynomial expected trajectory. ; (1.3) Generate a third-order continuous disturbance force acting on the UAV / robot dog base in the time dimension of the desired trajectory. The interference force satisfies the continuous characteristic of the second derivative to simulate the slowly varying and smooth external interference in the actual environment. The amplitude range of the interference force is set according to the platform's carrying capacity to ensure that the interference scenario conforms to the actual application scenario. (1.4) Using trajectory data from 10 consecutive dynamic iteration steps as a sample, collect complete actual motion state data within 10 steps: and the actual disturbance force at the tenth time step Then, a random forest obstacle is constructed in the simulation map, and an Euclidean symbolic distance field (ESDF) map is established. Combined with the platform's geometric dimensions, a collision detection algorithm is used to determine whether the sample trajectory collides with the random forest obstacle: if a collision occurs, the sample is discarded, and only valid samples without collisions are retained; then, a depth image captured by the camera is added to each sample. ; Obtain complete sample data.

3. The interference and trajectory learning method based on an integrated neural network according to claim 2, characterized in that, The specific steps (1.1) are as follows: Randomly generate the position vector of the platform's starting point. Starting velocity vector and the initial acceleration vector Randomly set the straight-line distance from the starting point to the ending point of the trajectory. , Simultaneously, a trajectory direction vector is randomly generated. Based on the starting position ,distance and direction vector Calculate the endpoint position Set the endpoint velocity vector Final acceleration vector This ensures a stable start and end point for the trajectory.

4. The interference and trajectory learning method based on an integrated neural network according to claim 3, characterized in that, The specific steps (1.2) are as follows: For uniaxial motion, a fifth-order polynomial is used to describe the relationship between position and time. , , The axes are solved independently, with Taking the axis as an example, the expression is: ； in, for Location at any given moment For polynomial coefficients, The total duration of the trajectory is defined as follows: ,in Given the maximum velocity; substitute the boundary conditions into the position and derivative formulas to construct the coefficient matrix equation: ； Solving for the coefficient vector, for a three-dimensional trajectory, , , Each axis is independently modeled using the aforementioned 5th-order polynomial model. The polynomial coefficients of each axis are solved using the same method, and the coefficient vector matrix is ​​calculated. To obtain the three-dimensional trajectory ; By analyzing the desired trajectory Find the first derivative to obtain the desired speed The desired acceleration is obtained by taking the second derivative. This forms a complete sequence of desired motion states.

5. The interference and trajectory learning method based on an integrated neural network according to claim 4, characterized in that, The specific steps (1.3) are as follows: Desired motion state sequence and third-order continuous disturbance force Input the dynamics model of the drone / robot dog and set the dynamics iteration time step. The actual motion state of the drone / robot dog base under disturbance is calculated in real time using dynamic equations. ,in The attitude of the drone / robot dog base.

6. The interference and trajectory learning method based on an integrated neural network according to claim 5, characterized in that, The specific steps (1.4) are as follows: Actual velocity data in the valid sample Actual acceleration data Standardization process: , ; in These are the minimum and maximum values ​​of the velocity data among all valid samples. These are the minimum and maximum values ​​of acceleration data among all valid samples, respectively.

7. The interference and trajectory learning method based on an integrated neural network according to claim 6, characterized in that, Step (2) is as follows: (2.1) Construct an integrated neural network using a "multimodal feature fusion - dual-task parallel output" structure; (2.2) The definition of network learning loss value includes two items: trajectory loss and interference loss.

8. The interference and trajectory learning method based on an integrated neural network according to claim 7, characterized in that, The specific steps (2.1) are as follows: 1) Visual feature processing: processing depth images ResNet network is used for feature extraction, and high-level visual features such as the distribution and relative positions of environmental obstacles are extracted through convolutional layers and residual connections; 2) Motion state processing: The standardized velocity, acceleration, and attitude data are processed using 1D convolutional layers to capture the temporal variation patterns of the motion state; 3) Feature Fusion and Output: Visual features and motion state features are fused through a cross-modal attention mechanism to obtain the following outputs: one is an estimate of external time-varying disturbances. Secondly, the velocity acceleration at the endpoint of the desired trajectory. .

9. The interference and trajectory learning method based on an integrated neural network according to claim 8, characterized in that, The specific steps (2.2) are as follows: Trajectory loss term: based on network output Based on the current platform location As the starting state, the corresponding trajectory is generated using the 5th order polynomial defined in step (1.1); Calculate in three directions respectively, with Taking direction as an example, for time Current position for: ; The defined safe distance function for each location is: ; in It is the closest distance to the obstacle calculated using an ESDF map. Given a distance threshold, The coefficients are given; therefore, the loss term for the entire curve is... Defined as: ； ； Interference loss item Interference estimate of network output Compared with actual interference value of Loss function: ； Joint learning: building a system that includes and The joint loss function is used to feed the preprocessed input data into the network for iterative training. The network parameters are optimized through backpropagation, which enables the network to improve the accuracy of trajectory endpoint prediction and interference estimation at the same time, realizing end-to-end collaborative learning of "motion state-environment perception-trajectory generation and interference prediction".

Images