Intelligent agricultural machine end-to-end momentum perception trajectory planning method, intelligent agricultural machine, medium

Abstract

The application provides an intelligent agricultural machine end-to-end momentum perception trajectory planning method, comprising the following steps: acquiring farmland operation related data, including farmland operation scene perception data, current farmland operation pose data and high-precision farmland map data; performing adaptive momentum feature extraction on the farmland operation related data, fusing dynamic historical space-time features, and generating an enriched planning query; inputting the enriched planning query as a condition parameter into a diffusion model for trajectory generation, and performing global search optimization through Monte Carlo tree search during the trajectory generation process to generate a multi-modal candidate trajectory. The method greatly reduces the collision rate in a complex dynamic scene and improves the trajectory prediction consistency.

Description

Technical Field

[0001] This invention relates to the field of intelligent agricultural machinery, and more particularly to an end-to-end momentum sensing trajectory planning method for intelligent agricultural machinery, intelligent agricultural machinery, and storage medium. Background Technology

[0002] With the development of smart agriculture, agricultural production is rapidly evolving from mechanization to intelligence and automation, and "unmanned farms" have become a development trend. In autonomous agricultural equipment, path planning and trajectory prediction are core capabilities that directly determine the stability and safety of agricultural machinery in complex farmland environments.

[0003] Traditional methods for automated agricultural machinery often rely on rule-based or simple regression models. However, farmland environments are characterized by significant unstructured features (such as irregular furrows, crop shading, and rugged terrain) and high dynamic noise (such as swaying vegetation and dust), making it difficult to directly apply traditional methods. While cutting-edge end-to-end automated driving can directly map from sensors to trajectories, existing frameworks such as momentum-aware driving still face challenges in agricultural applications, including pattern collapse, short-sightedness in local guidance, and static temporal coupling. This invention aims to improve upon these limitations. Summary of the Invention

[0004] This invention aims to solve the problems of trajectory prediction jitter, poor consistency of long-term planning, sensitivity to farmland noise, and difficulty in multimodal trajectory selection in existing end-to-end automatic driving of intelligent agricultural machinery. It proposes an end-to-end momentum perception trajectory planning method for intelligent agricultural machinery, an intelligent agricultural machinery, and a storage medium.

[0005] This invention provides an end-to-end momentum sensing trajectory planning method for intelligent agricultural machinery, comprising:

[0006] Acquire data related to farmland operations, including farmland operation scenario perception data, current operating posture data of agricultural machinery, and high-precision farmland map data;

[0007] Adaptive momentum feature extraction is performed on the farmland operation-related data, and dynamic historical spatiotemporal features are fused to generate an enriched planning query.

[0008] The enrichment planning query is input as a conditional parameter into the diffusion model for trajectory generation. During the trajectory generation process, a global search optimization is performed through Monte Carlo tree search to generate multimodal candidate trajectories.

[0009] In one embodiment of the present invention, the step of performing adaptive momentum feature extraction on the farmland operation-related data, fusing dynamic historical spatiotemporal features, and generating an enriched planning query includes:

[0010] The dynamics of the current farmland scene are quantified by calculating core indicators, including: the motion entropy and minimum collision time of the current scene; where the minimum collision time is the minimum estimated collision time between the current trajectory of the agricultural machinery and all obstacles within the field of view; motion entropy is used to measure the uncertainty of the motion of objects within the field of view; where, if N dynamic obstacles are detected within the field of view, the velocity distribution probability of the dynamic obstacles is... Then the entropy of motion The calculation formula is: ;

[0011] Based on the minimum collision time and motion entropy, the optimal number of historical review frames at the current moment is calculated using an adaptive control function, and the calculation formula is as follows:

[0012]

[0013] in, , For preset weighting coefficients, These are the preset minimum and maximum number of historical review frames, respectively. Minimum collision time, For the entropy of motion, For historical review frames; when TTC is less than the preset collision time threshold or When the number of historical review frames exceeds a preset motion entropy threshold, reduce the number of frames. To focus on the latest perceived information; when TTC is greater than or equal to the preset collision time threshold and When the number of historical review frames is less than or equal to the preset motion entropy threshold, increase the number of historical review frames. To utilize long-term contextual information; The function is used to restrict the elements in an array to between a specified minimum and maximum value.

[0014] Based on the optimal number of historical review frames, select the corresponding historical query set. Introducing temporal encoding into the cross-attention mechanism The enrichment program query is obtained by weighting historical queries based on time relevance. The formula for calculating attention is:

[0015]

[0016] in, These are the query matrix, key matrix, and value matrix for the attention mechanism, respectively. Given the dimension of the key matrix, the output of the attention calculation is the enrichment programming query. .

[0017] In one embodiment of the present invention, during trajectory generation in the diffusion model...

[0018] During the training phase, a positive diffusion process is performed, gradually adding Gaussian noise to the real expert driving trajectory until the trajectory becomes pure Gaussian noise, simulating the gradual loss of trajectory information.

[0019] During the inference phase, an inverse denoising process is performed to recover the expert driving trajectory from the pure Gaussian noise trajectory using a neural network. Implementation, in which These are conditional parameters; the iterative formula for inverse denoising is:

[0020]

[0021] in, Let be the noise standard deviation of the t-th step of reverse denoising. It is random Gaussian noise. This represents the current diffusion step number. For the noisy trajectory at step t, For the noisy trajectory at step t-1, Let be the diffusion coefficient at step t. It is the cumulative product of the diffusion coefficients in the first t steps. The parameter is The deep learning backbone network.

[0022] In one embodiment of the present invention, a global search is performed using Monte Carlo tree search during trajectory generation.

[0023] The optimization process includes:

[0024] Construct a search space where the state of a node in the search tree corresponds to a noisy node at any step in the diffusion process.

[0025] The trajectory and action space are all possible sampling paths from the current noisy state to the next state after denoising.

[0026] During the tree search loop, a Monte Carlo tree search loop is executed at preset intervals of diffusion steps, including:

[0027] The UCB confidence upper bound algorithm is used to balance the process of selecting known high-scoring paths with trying unknown paths, and to select the most promising child nodes.

[0028] At the leaf nodes, the diffusion model is used to predict multiple possible next denoising states to complete node expansion.

[0029] Starting from the expanded nodes, a simplified diffusion strategy is used to quickly extrapolate until a complete expert driving trajectory is generated;

[0030] Calculate the reward score for the simulated expert driving trajectory, and then backpropagate this reward score to update the value of all nodes on the path.

[0031] In one embodiment of the present invention, the reward score is composed of a weighted average of three parts: collision probability constraint, comfort constraint, and topology constraint, and the calculation formula is as follows:

[0032]

[0033] in, , , These are preset weighting coefficients; This represents a collision probability constraint.

[0034] Based on the overlap integral of the surrounding agent prediction distribution, the collision probability is directly minimized; This indicates a comfort constraint, penalizing excessive acceleration. and curvature ,Right now ; It represents topological constraints and uses high-precision farmland map data to calculate the deviation of the trajectory from the center line of the furrow or the boundary of the field.

[0035] In one embodiment of the present invention, based on the DiffuSpec framework, a draft model with fewer parameters than the main diffusion model is used to quickly generate draft trajectory waypoints, and then the main diffusion model is used to perform parallel verification and fine-tuning of the draft trajectory.

[0036] Semantic perturbations of farmland scenes are added to the model training data to enable the model to learn robust features and adapt to harsh field operation environments; wherein, the semantic perturbations include at least one of the following: camera partially obscured by mud, sudden changes in illumination, and missing radar point clouds.

[0037] In one embodiment of the present invention, the farmland operation scene perception data includes at least one of farmland environmental point cloud, field operation image, and farmland obstacle detection data, and the current operation posture data of the agricultural machinery includes the position, heading, and speed information of the agricultural machinery.

[0038] In another aspect, the present invention provides an intelligent agricultural machine, comprising: an environmental sensing component, a positioning module, an on-board computing unit, and a storage unit, wherein the on-board computing unit is communicatively connected to the environmental sensing component, the positioning module, and the storage unit respectively;

[0039] The environmental perception component collects perception data of farmland operation scenarios; the positioning module collects the current operating posture data of agricultural machinery; and the storage unit stores high-precision map data of farmland.

[0040] The on-board computing unit includes:

[0041] The acquisition module is used to acquire data related to farmland operations, including farmland operation scene perception data, current operation posture data of agricultural machinery, and high-precision farmland map data.

[0042] The adaptive momentum planning interaction module is used to perform adaptive momentum feature extraction on the farmland operation-related data, integrate dynamic historical spatiotemporal features, and generate enriched planning queries.

[0043] The diffusion trajectory synthesis module is used to input the enrichment planning query as a condition parameter into the diffusion model to generate the trajectory. During the trajectory generation process, a global search optimization is performed through Monte Carlo tree search to generate multimodal candidate trajectories.

[0044] In one embodiment of the present invention, the intelligent agricultural machinery further includes: a drive-by-wire execution unit, which is communicatively connected to the vehicle-mounted computing unit, for receiving the candidate trajectory and controlling the agricultural machinery to complete field operations according to the candidate trajectory.

[0045] In another aspect, the present invention provides a readable storage medium storing a program or instructions, characterized in that the program or instructions, when executed by a processor, implement the above-described intelligent agricultural machinery end-to-end momentum sensing trajectory planning method.

[0046] As can be seen from the above solutions, the advantages of the present invention are:

[0047] This invention provides an end-to-end momentum-sensing trajectory planning method for intelligent agricultural machinery. It adaptively extracts momentum features from farmland operation-related data, integrates dynamic historical spatiotemporal features, and generates an enriched planning query. This enriched planning query is then input as a conditional parameter into a diffusion model for trajectory generation. During trajectory generation, a Monte Carlo Tree Search (MCTS) is used for global search optimization to generate multimodal candidate trajectories. This method combines a diffusion model with MCTS search and employs an adaptive momentum interaction mechanism, significantly improving the stability and consistency of intelligent agricultural machinery trajectory prediction. It also substantially reduces the collision rate in complex dynamic scenarios, effectively solving the challenge of agricultural machinery operation planning in complex environments. Attached Figure Description

[0048] Figure 1 This is a flowchart illustrating an end-to-end momentum sensing trajectory planning method for intelligent agricultural machinery according to an embodiment of the present invention.

[0049] Figure 2 A detailed flowchart of step S2 is shown;

[0050] Figure 3 A schematic diagram of the frame structure of an intelligent agricultural machine provided in an embodiment of the present invention is shown.

[0051] The attached figures are labeled as follows:

[0052] 31: Environmental sensing components;

[0053] 32: Positioning module;

[0054] 33: Onboard computing unit;

[0055] 34: Storage unit;

[0056] 35: Drive-by-wire actuator;

[0057] 331: Acquisition module;

[0058] 332: Adaptive Momentum Planning Interactive Module;

[0059] 333: Diffusion trajectory synthesis module. Detailed Implementation

[0060] It should be noted that, in this invention, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0061] In the absence of further restrictions, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0062] refer to Figure 1 As shown, Figure 1 The diagram shows a flowchart of an end-to-end momentum sensing trajectory planning method for intelligent agricultural machinery provided in an embodiment of the present invention.

[0063] A method for end-to-end momentum sensing trajectory planning of intelligent agricultural machinery includes the following steps:

[0064] Step S1: Obtain relevant data for farmland operations, including farmland operation scene perception data, current operation posture data of agricultural machinery, and high-precision farmland map data.

[0065] The farmland operation scene perception data includes at least one of farmland environmental point clouds, field operation images, and farmland obstacle detection data. The current operating posture data of the agricultural machinery includes the position, heading, and speed information of the agricultural machinery. The high-precision farmland map data includes information such as plot boundaries and furrow lines.

[0066] Step S2: Perform adaptive momentum feature extraction on the farmland operation-related data, integrate dynamic historical spatiotemporal features, and generate an enriched planning query.

[0067] Step S3: Input the enrichment planning query as a condition parameter into the diffusion model to generate the trajectory. During the trajectory generation process, global search optimization is performed through Monte Carlo tree search to generate multimodal candidate trajectories.

[0068] In this embodiment, adaptive momentum feature extraction is performed on farmland operation-related data, and dynamic historical spatiotemporal features are fused to generate an enriched planning query. This enriched planning query is then input as a conditional parameter into a diffusion model for trajectory generation. During trajectory generation, a Monte Carlo Tree Search (MCTS) is used for global search optimization to generate multimodal candidate trajectories. By combining the diffusion model with MCTS search and employing an adaptive momentum interaction mechanism, the stability and consistency of intelligent agricultural machinery trajectory prediction are significantly improved. This method substantially reduces the collision rate in complex dynamic scenarios, improves trajectory prediction consistency, and possesses generalization capabilities for long-tailed scenarios.

[0069] Research shows that existing momentum planning interactors (MPIs) typically employ a fixed history frame window (e.g., a fixed lookback to the past). The system extracts spatiotemporal features using a fixed window (in seconds). However, the operating speed of agricultural machinery varies greatly. During precision sowing, the speed may be as low as 3 km / h, while during field migration, it may reach 30 km / h. In low-speed precision operations, a fixed short window leads to insufficient utilization of historical information and an inability to capture slow-moving trends. In high-speed or highly dynamic scenarios (such as when someone suddenly enters), a fixed long window introduces outdated historical features, diluting the effectiveness of current perception and resulting in suboptimal utilization of perception momentum. The system cannot flexibly adjust its "attention" according to the current operating rhythm. To address this problem, in one embodiment of the present invention, an adaptive momentum interaction is considered. This algorithm uses motion entropy and time to minimum collision (TTC) as two indicators to dynamically adjust the length of the historical window. Adaptive momentum feature extraction is performed on the farmland operation-related data, fusing dynamic historical spatiotemporal features to generate an enriched planning query.

[0070] Specifically, refer to Figure 2 As shown, Figure 2 A schematic diagram of the specific process for step S2 is shown.

[0071] Step S21: Quantify the dynamics of the current farmland scene and calculate core indicators, including: the motion entropy and minimum collision time of the current scene; where the minimum collision time is the minimum expected collision time between the current trajectory of the agricultural machinery and all obstacles within the field of view; motion entropy is used to measure the uncertainty of the motion of objects within the field of view; where, if N dynamic obstacles (such as pedestrians, other agricultural machinery, etc.) are detected within the field of view, the velocity distribution probability of the dynamic obstacles is: Then the entropy of motion The calculation formula is: , Let be the speed of the i-th dynamic obstacle.

[0072] Step S22: Based on the minimum collision time and motion entropy, calculate the optimal number of historical review frames at the current moment using an adaptive control function. The calculation formula is as follows:

[0073]

[0074] in, , For preset weighting coefficients, These are the preset minimum and maximum number of historical review frames, respectively. Minimum collision time, For the entropy of motion, For historical review frames; when TTC is less than the preset collision time threshold or When the number of historical review frames exceeds a preset motion entropy threshold, reduce the number of frames. To focus on the latest perceived information; when TTC is greater than or equal to the preset collision time threshold and When the number of historical review frames is less than or equal to the preset motion entropy threshold, increase the number of historical review frames. It utilizes long-term context information; the Clip function is used to restrict the elements in an array to between a specified minimum and maximum value.

[0075] Step S23: Select the corresponding historical query set based on the optimal number of historical review frames. Introducing temporal encoding into the cross-attention mechanism Historical queries based on time relevance Perform weighted calculations to obtain the enrichment programming query. The formula for calculating attention is:

[0076]

[0077] in, These are the query matrix, key matrix, and value matrix for the attention mechanism, respectively. Given the dimension of the key matrix, the output of the attention calculation is the enrichment programming query. .

[0078] Furthermore, research indicates that existing technologies often employ L1 / L2 norm regression loss functions for model training. This deterministic regression method essentially seeks the "average behavior" within the dataset. When faced with long-tailed distribution scenarios such as farmland edges, mixed obstacles, or extreme weather, the model tends to output a "compromise" and mediocre trajectory instead of choosing from multiple feasible avoidance strategies (such as detouring to the left, detouring to the right, or stopping and waiting). This results in a lack of trajectory diversity, making it unable to cope with low-probability but high-risk complex conditions and reducing operational safety. To address this issue, in one embodiment of this invention, a diffusion model with strong distribution modeling capabilities is used to replace the regression model. Specifically, during the trajectory generation process of the diffusion model, a forward diffusion process is performed during the training phase, guiding the model to real expert driving trajectories. Gradually add Gaussian noise until the trajectory becomes pure Gaussian noise, i.e., a noisy trajectory. The simulation demonstrates the gradual loss of trajectory information. During the inference phase, a reverse denoising process is performed, transforming the trajectory from pure Gaussian noise... The expert's driving trajectory was recovered. By reversing the process, a high-fidelity trajectory that matches the momentum of the current environment is recovered from random noise. Specifically, this is achieved through a neural network. Implementation, in which These are conditional parameters. The iterative formula for inverse denoising is:

[0079]

[0080] in, Let be the noise standard deviation of the t-th step of reverse denoising. It is random Gaussian noise. This represents the current diffusion step number. For the noisy trajectory at step t, For the noisy trajectory at step t-1, Let be the diffusion coefficient at step t. It is the cumulative product of the diffusion coefficients in the first t steps. The parameter is The deep learning backbone network.

[0081] Furthermore, the generation process of ordinary diffusion models is blind, only pursuing probability maximization. In this embodiment, Monte Carlo Tree Search (MCTS) is introduced during the inference phase to perform a global search optimization process. Global guidance addresses the problem in existing end-to-end models that use local guidance mechanisms and lack strategic extrapolation of the long-term future. This leads to small errors generated during single-step inference accumulating rapidly over time, resulting in poor consistency in long-distance planning, trajectory jitter, frequent direction corrections, and even problems such as agricultural machinery deviating from the furrow and damaging crops.

[0082] Specifically, the search space is first constructed, where the state is the nodes of the search tree. This corresponds to an intermediate state in the diffusion process, i.e., the noisy trajectory at any step. The action space represents all possible sampling paths from the current noisy state to the next state. Because the sampling in the diffusion model is random, the same... It can be transformed into multiple types through different sampling. These constitute the action space.

[0083] Then, during the tree search loop, a Monte Carlo tree search loop is executed at preset diffusion steps (usually selecting key diffusion steps, such as every 5 diffusion steps). During the loop, the UCB (upper confidence bound) algorithm is used to balance selecting known high-scoring paths with trying unknown paths, choosing the most promising child nodes. At leaf nodes, the diffusion model is used to predict multiple possible next denoising states, completing node expansion. Starting from the expanded nodes, a simplified diffusion strategy is used to quickly extrapolate until a complete expert driving trajectory is generated. Calculate a reward score for the simulated expert driving trajectory, and then backpropagate this reward score to update the value of all nodes along the path.

[0084] Furthermore, existing end-to-end models typically learn driving behavior implicitly, lacking direct constraints on jerk, probability of collision (PoC), and road topology. While the generated trajectory may be geometrically connected, it can contain abrupt accelerations and decelerations (excessive jerk), which is fatal for tractors carrying heavy implements, easily leading to implement damage or reduced work quality (such as inconsistent seeding depth). To address this, in one embodiment of the present invention, the reward score... It consists of three weighted components: collision probability constraints, comfort constraints, and topology constraints. The calculation formula is as follows:

[0085]

[0086] in, , , These are preset weighting coefficients; This represents a collision probability constraint.

[0087] Based on the overlap integral of the surrounding agent prediction distribution, the collision probability is directly minimized; It represents topological constraints and uses high-precision farmland map data to calculate the deviation between the trajectory and the center line of the furrow or the boundary of the field, ensuring that the agricultural machinery does not crush the seedlings or drive out of the working boundary. This represents a comfort constraint, penalizing excessive jerk and curvature. ,in For the jerk (Jerk) term, This is the curvature term. Jerk directly reflects the smoothness of agricultural machinery operation. An excessively large value indicates strong jerking during acceleration, deceleration, and steering, which not only accelerates the mechanical wear of the machinery's chassis and operating components but also leads to decreased precision in operations such as sowing, fertilizing, and spraying, resulting in problems like missed sowing and uneven application. Essentially, this penalty is designed to reduce aggressive driving maneuvers, ensuring smooth and stable operation of agricultural machinery and matching the response speed of the equipment. The penalty for high curvature is to prevent the machinery from turning too sharply, thus avoiding tipping over, crushing seedlings, or damaging the ridges.

[0088] In one embodiment, to meet the control frequency requirements of agricultural machinery above 10Hz, the DiffuSpec framework is employed. Based on the DiffuSpec framework, a draft model with fewer parameters than the main diffusion model is used to quickly generate draft trajectory waypoints, and then the main diffusion model is used to perform parallel verification and fine-tuning of the draft trajectory. This method significantly reduces inference latency while maintaining generation quality.

[0089] In one embodiment, semantic perturbations of a farmland scene are added to the model training data. These perturbations include at least one of the following: camera partially obscured by mud, abrupt changes in illumination, or missing radar point clouds. This forces the model to learn robust features that are insensitive to these perturbations, thereby maintaining stability in harsh farmland environments.

[0090] In summary, the end-to-end momentum-sensing trajectory planning method for intelligent agricultural machinery provided by this invention adaptively extracts momentum features from farmland operation-related data, integrates dynamic historical spatiotemporal features, and generates an enriched planning query. This enriched planning query is then input as a conditional parameter into a diffusion model for trajectory generation. Furthermore, during trajectory generation, a Monte Carlo Tree Search (MCTS) is used for global search optimization to generate multimodal candidate trajectories. This method combines a diffusion model with MCTS search and employs an adaptive momentum interaction mechanism, significantly improving the stability and consistency of intelligent agricultural machinery trajectory prediction. It also substantially reduces the collision rate in complex dynamic scenarios, effectively solving the challenge of agricultural machinery operation planning in complex environments.

[0091] Another embodiment of the present invention also provides an intelligent agricultural machine, see reference. Figure 3 As shown, Figure 3 A schematic block diagram of the intelligent agricultural machine is shown. It includes: an environmental perception component 31, a positioning module 32, an on-board computing unit 33, and a storage unit 34, wherein the on-board computing unit 33 is communicatively connected to the environmental perception component 31, the positioning module 32, and the storage unit 34.

[0092] The environmental perception components are common, existing perception configurations, such as a combination of LiDAR, agricultural-specific cameras, and millimeter-wave radar, to collect perception data of farmland operation scenarios. This includes farmland environmental point clouds collected by LiDAR, field operation images collected by cameras, and farmland obstacle detection data collected by millimeter-wave radar. The current operating posture data of the agricultural machinery includes the machine's position, heading, and speed information. The high-precision farmland map data includes information such as plot boundaries and furrow lines.

[0093] The positioning module collects the current working posture data of the agricultural machinery, and the storage unit stores high-precision map data of farmland.

[0094] The onboard computing unit employs an existing, widely deployed autonomous driving computing platform, such as the NVIDIA Orin NX / Orin X computing board, to execute the intelligent agricultural machinery end-to-end momentum perception trajectory planning method described in the above embodiment. The onboard computing unit 33 specifically includes:

[0095] The acquisition module 331 is used to acquire data related to farmland operations, including farmland operation scene perception data, current operation posture data of agricultural machinery, and high-precision farmland map data.

[0096] The adaptive momentum planning interaction module 332 is used to perform adaptive momentum feature extraction on the farmland operation-related data, integrate dynamic historical spatiotemporal features, and generate enriched planning queries.

[0097] The diffusion trajectory synthesis module 333 is used to input the enrichment planning query as a condition parameter into the diffusion model for trajectory generation. During the trajectory generation process, a global search optimization is performed through Monte Carlo tree search to generate multimodal candidate trajectories. The finally generated multimodal candidate trajectories will be stored in the storage unit.

[0098] In one embodiment, the intelligent agricultural machinery further includes: a drive-by-wire execution unit 35, which is communicatively connected to the on-board computing unit 33, for receiving the candidate trajectory and controlling the agricultural machinery to complete field operations according to the candidate trajectory.

[0099] This device implementation can be combined with the above-described implementations. The relevant technical details mentioned in the above-described implementations remain valid in this device implementation; therefore, to avoid repetition, they will not be repeated here.

[0100] In addition, this embodiment of the invention also provides a readable storage medium storing a program or instructions that, when executed by a processor, implement the steps of the above-described intelligent agricultural machinery end-to-end momentum sensing trajectory planning method and achieve the same technical effect.

[0101] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described intelligent agricultural machinery end-to-end momentum sensing trajectory planning method and achieves the same technical effect.

[0102] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.

Claims

1. A method for end-to-end momentum sensing trajectory planning of intelligent agricultural machinery, characterized in that, include: Acquire data related to farmland operations, including farmland operation scenario perception data, current operating posture data of agricultural machinery, and high-precision farmland map data; Adaptive momentum feature extraction is performed on the farmland operation-related data, and dynamic historical spatiotemporal features are fused to generate an enriched planning query. The enrichment planning query is input as a conditional parameter into the diffusion model for trajectory generation. During the trajectory generation process, a global search optimization is performed through Monte Carlo tree search to generate multimodal candidate trajectories.

2. The method according to claim 1, characterized in that, The steps of adaptive momentum feature extraction, fusion of dynamic historical spatiotemporal features, and generation of enrichment programming query for the farmland operation-related data include: The dynamics of the current farmland scene are quantified, and core indicators are calculated, including: the motion entropy of the current scene and the minimum collision time; where the minimum collision time is the minimum expected collision time between the current trajectory of the agricultural machinery and all obstacles in the field of vision. Motion entropy is used to measure the uncertainty of the motion of objects within the field of view; where, if N dynamic obstacles are detected within the field of view, the velocity distribution probability of the dynamic obstacles is... Then the entropy of motion The calculation formula is: ; Based on the minimum collision time and motion entropy, the optimal number of historical review frames at the current moment is calculated using an adaptive control function, and the calculation formula is as follows: in, , For preset weighting coefficients, These are the preset minimum and maximum number of historical review frames, respectively. Minimum collision time, For the entropy of motion, For historical review frames; when TTC is less than the preset collision time threshold or When the number of historical review frames exceeds a preset motion entropy threshold, reduce the number of frames. To focus on the latest perceived information; when TTC is greater than or equal to the preset collision time threshold and When the number of historical review frames is less than or equal to the preset motion entropy threshold, increase the number of historical review frames. To utilize long-term contextual information; Based on the optimal number of historical review frames, select the corresponding historical query set. Introducing temporal encoding into the cross-attention mechanism The enrichment program query is obtained by weighting historical queries based on time relevance. The formula for calculating attention is: in, These are the query matrix, key matrix, and value matrix for the attention mechanism, respectively. Given the dimension of the key matrix, the output of the attention calculation is the enrichment programming query. .

3. The method according to claim 2, characterized in that, During trajectory generation using a diffusion model, During the training phase, a positive diffusion process is performed, gradually adding Gaussian noise to the real expert driving trajectory until the trajectory becomes pure Gaussian noise, simulating the gradual loss of trajectory information. During the inference phase, an inverse denoising process is performed to recover the expert driving trajectory from the pure Gaussian noise trajectory using a neural network. Implementation, in which These are conditional parameters; the iterative formula for inverse denoising is: in, Let be the noise standard deviation of the t-th step of reverse denoising. It is random Gaussian noise. This represents the current diffusion step number. For the noisy trajectory at step t, For the noisy trajectory at step t-1, Let be the diffusion coefficient at step t. It is the cumulative product of the diffusion coefficients in the first t steps. The parameter is The deep learning backbone network.

4. The method according to claim 3, characterized in that, During trajectory generation, Monte Carlo methods are used. The global search optimization process for Carlo tree search includes: Construct a search space where the state of a node in the search tree corresponds to a noisy node at any step in the diffusion process. The trajectory and action space are all possible sampling paths from the current noisy state to the next state after denoising. During the tree search loop, a Monte Carlo tree search loop is executed at preset intervals of diffusion steps, including: The UCB confidence upper bound algorithm is used to balance the process of selecting known high-scoring paths with trying unknown paths, and to select the most promising child nodes. At the leaf nodes, the diffusion model is used to predict multiple possible next denoising states to complete node expansion. Starting from the expanded nodes, a simplified diffusion strategy is used to quickly extrapolate until a complete expert driving trajectory is generated; Calculate the reward score for the simulated expert driving trajectory, and then backpropagate this reward score to update the value of all nodes on the path.

5. The method according to claim 4, characterized in that, The reward score is composed of a weighted average of collision probability constraints, comfort constraints, and topology constraints, and the calculation formula is as follows: in, , , These are preset weighting coefficients; This represents a collision probability constraint. Based on the overlap integral of the surrounding agent prediction distribution, the collision probability is directly minimized; This indicates a comfort constraint, penalizing excessive acceleration. and curvature ,Right now ; It represents topological constraints and uses high-precision farmland map data to calculate the deviation of the trajectory from the center line of the furrow or the boundary of the field.

6. The method according to claim 2, characterized in that, Based on the DiffuSpec framework, a draft model with fewer parameters than the main diffusion model is used to quickly generate draft trajectory waypoints, and then the main diffusion model is used to perform parallel verification and fine-tuning of the draft trajectory. Semantic perturbations of farmland scenes are added to the model training data to enable the model to learn robust features and adapt to harsh field operation environments; wherein, the semantic perturbations include at least one of the following: camera partially obscured by mud, sudden changes in illumination, and missing radar point clouds.

7. The method according to claim 1, characterized in that, The farmland operation scene perception data includes at least one of farmland environmental point cloud, field operation image, and farmland obstacle detection data, and the current operation posture data of the agricultural machinery includes the position, heading, and speed information of the agricultural machinery.

8. An intelligent agricultural machine, characterized in that, include: It is equipped with an environmental perception component, a positioning module, an on-board computing unit, and a storage unit, wherein the on-board computing unit is communicatively connected to the environmental perception component, the positioning module, and the storage unit respectively; The environmental perception component collects perception data of farmland operation scenarios; the positioning module collects the current operating posture data of agricultural machinery; and the storage unit stores high-precision map data of farmland. The on-board computing unit includes: The acquisition module is used to acquire data related to farmland operations, including farmland operation scene perception data, current operation posture data of agricultural machinery, and high-precision farmland map data. The adaptive momentum planning interaction module is used to perform adaptive momentum feature extraction on the farmland operation-related data, integrate dynamic historical spatiotemporal features, and generate enriched planning queries. The diffusion trajectory synthesis module is used to input the enrichment planning query as a condition parameter into the diffusion model to generate the trajectory. During the trajectory generation process, a global search optimization is performed through Monte Carlo tree search to generate multimodal candidate trajectories.

9. The intelligent agricultural machinery according to claim 8, characterized in that, It also includes: a drive-by-wire execution unit, which is communicatively connected to the on-board computing unit, for receiving the candidate trajectory and controlling the agricultural machinery to complete field operations according to the candidate trajectory.

10. A readable storage medium storing a program or instructions, characterized in that, When the program or instruction is executed by the processor, it implements the intelligent agricultural machinery end-to-end momentum sensing trajectory planning method according to any one of claims 1-7.

Images