Method and system for multi-robot collaborative harvesting of woody oil plants based on swarm intelligence

Abstract

The present application relates to the field of intelligent agricultural robots, especially to a woody oil plant multi-robot cooperative harvesting method and system based on swarm intelligence, the present application proposes to build a distributed sensing network and a semantic enhanced digital twin model, realizing efficient harvesting of woody oil plants, the system adopts a three-layer decision architecture to formulate a harvesting strategy, and improves work efficiency through adaptive task allocation and cooperative harvesting, multi-robots share experience through federated learning, realize group collaborative learning, and optimize harvesting performance, the system reliability is improved by 85%, there is no single point failure risk, and the multi-modal sensing system improves the identification accuracy by 50%, bringing an innovative harvesting solution to the field of intelligent agricultural robots.

Description

Technical Field

[0001] This invention relates to the field of intelligent agricultural robots, and in particular to a multi-robot collaborative harvesting method and system for woody oil crops based on swarm intelligence, applicable to intelligent and efficient harvesting operations of woody oil crops such as camellia seeds, olives, and tea seeds. Background Technology

[0002] With the improvement of agricultural modernization and the rise in rural labor costs in my country, intelligent agricultural robot technology has been gradually applied and developed. Woody oil crops are characterized by their wide distribution, high value, and high cost of manual harvesting, making intelligent harvesting an important research direction in modern agriculture.

[0003] Currently, there is relevant research on multi-robot collaborative operation systems. For example, Chinese patent CN 111240319 B discloses an outdoor multi-robot collaborative operation system and method. This system adopts a master-slave structure, consisting of a master robot and several slave robots. Each robot is equipped with navigation and positioning, motion control, sensors, communication, and master control modules. The master robot of this system constructs an environmental map using high-precision LiDAR, the slave robots receive map information, and the central control backend is responsible for task allocation and path planning.

[0004] However, the aforementioned existing technologies have the following shortcomings:

[0005] 1. Centralized control architecture results in low system reliability; failure of the main robot or the central control backend can paralyze the entire system.

[0006] 2. Excessive reliance on communication can easily lead to communication bottlenecks in complex orchard environments;

[0007] 3. Limited sensing capabilities; relying solely on lidar makes it difficult to effectively identify the characteristics and maturity of oilseed crops.

[0008] 4. The algorithm lacks adaptability; fixed algorithm combinations are difficult to adapt to the complex and ever-changing environment and task requirements of orchards.

[0009] 5. The collaboration mechanism is simple, mainly limited to task allocation and obstacle avoidance, and lacks in-depth collaboration for harvesting operations;

[0010] 6. The plan lacks a time dimension, only considering spatial path planning, and does not take into account the optimization of harvesting sequence;

[0011] 7. No consideration was given to crop characteristics; optimization was not performed for the specific growth characteristics of woody oil crops. Summary of the Invention

[0012] The purpose of this invention is to provide a multi-robot collaborative harvesting method and system for woody oil crops based on swarm intelligence. Through distributed sensing networks, a three-layer decision architecture, adaptive task allocation and execution, swarm collaborative learning, and dynamic equilibrium optimization, it achieves efficient and intelligent harvesting of woody oil crops, overcoming the aforementioned defects in the prior art.

[0013] This invention proposes a multi-robot collaborative harvesting method for woody oilseeds based on swarm intelligence, comprising:

[0014] A distributed sensing network is constructed in the oilseed orchard to acquire fruit distribution data, maturity data, and tree structure data. The distributed sensing network consists of multimodal sensors equipped on multiple robots, with each robot collecting local sensing data and sharing it.

[0015] Based on the data acquired by the distributed sensing network, a semantically enhanced digital twin model containing fruit attributes is established;

[0016] Based on the digital twin model, harvesting decisions are generated through a three-layer decision-making architecture, which includes a global strategic layer, a regional tactical layer, and a local operational layer. The global strategic layer formulates the overall harvesting strategy, the regional tactical layer generates regional harvesting plans, and the local operational layer determines specific harvesting actions.

[0017] Based on the harvesting decision, adaptive task allocation and collaborative harvesting are performed, including: using a virtual economic model to evaluate task value and robot capabilities, and dynamically allocating harvesting tasks; selecting appropriate harvesting strategies according to fruit characteristics; and avoiding conflicts through multi-robot collaborative trajectory planning.

[0018] Collect task execution results and realize collaborative learning and experience sharing among groups through a federated learning architecture;

[0019] Based on the learning results, dynamic equilibrium optimization is performed to seek the global optimum among harvest quantity, quality, energy efficiency, and uniformity.

[0020] Preferably, the construction of the oilfield distributed sensing network includes:

[0021] Each robot is configured with a multimodal sensor suite, which includes an RGB camera, a multispectral camera, a lidar, a force sensor, and an environmental sensor.

[0022] The RGB camera and multispectral camera were used to identify fruit characteristics and classify fruit maturity into 5 levels.

[0023] The lidar is used to generate point cloud data to identify tree structures;

[0024] Based on the multimodal sensor data, a fruit distribution density heat map and a harvesting difficulty assessment are generated, with the harvesting difficulty quantified on a scale of 1 to 10.

[0025] Preferably, the establishment of the semantically enhanced digital twin model includes:

[0026] Distributed SLAM technology is used to construct local environment maps, with each robot maintaining a local map covering an area of ​​50m × 50m;

[0027] Map stitching is achieved through ORB feature point matching, requiring a map overlap of more than 30%.

[0028] Upgrade traditional SLAM to semantic SLAM by introducing semantic tags such as trees, fruits, and obstacles;

[0029] Construct a complete digital twin model that includes a tree geometry model, a fruit attribute database, terrain obstacle representation, and historical harvesting data.

[0030] Preferably, the method of generating harvesting decisions through a three-layer decision architecture further includes:

[0031] Implement a distributed consensus protocol based on blockchain principles, using a proposal-verification-voting model, and execute the proposal only when more than 67% of the robots agree to it;

[0032] Introducing a time decay factor λ = e -αt This gives higher weight to recent data;

[0033] Establish a multi-objective optimization function F(x) = w1·output(x) + w2·quality(x) - w3·energy consumption(x) - w4·time(x), where w1, w2, w3, and w4 are dynamically adjusted weight parameters;

[0034] Implement a multi-agent collaborative reasoning system that supports distributed information aggregation and joint reasoning.

[0035] Preferably, the adaptive task allocation and collaborative harvesting includes:

[0036] The harvesting task is broken down into hierarchical levels: park level (L1), region level (L2), tree level (L3), and fruit level (L4).

[0037] The task allocation is optimized through a virtual economic model, and the task value V(task) is calculated as: number of fruits × maturity × market value. The robot capability score C(robot) is calculated as: historical efficiency × power consumption × actuator accuracy.

[0038] Identify five typical fruit states (easily detached, firm, clustered, shaded, and high-positioned) and select a specific harvesting sequence for each type;

[0039] Implement multi-robot collaborative trajectory planning, predict potential path intersections within 10 seconds, and execute priority-based avoidance strategies.

[0040] Preferably, the method of achieving collaborative learning and experience sharing through a federated learning architecture includes:

[0041] The model is trained locally on each robot, aggregating only the model parameters rather than the raw data;

[0042] Apply differential privacy protection mechanisms and add appropriate noise to protect sensitive information;

[0043] Construct a collective experience library that prioritizes the storage of high-value experiences. The experience representation includes scene features, actions performed, result evaluations, and contextual information.

[0044] Online reinforcement learning is implemented. The state space S includes the robot's position, power, load, and surrounding fruit distribution. The action space A includes the movement direction, harvesting action, and cooperation request. The reward function R = harvest quantity × quality coefficient - energy consumption × energy consumption coefficient - time × time coefficient.

[0045] Preferably, the dynamic balancing optimization includes:

[0046] Establish a multi-objective optimization framework to simultaneously optimize harvest quantity (number of fruits harvested per hour), quality assurance (fruit damage rate <5%), energy efficiency (number of fruits harvested per unit of electricity) and harvest uniformity (variance of harvest completion rate in different regions).

[0047] Implement adaptive resource allocation, assigning short-range tasks to robots with low power and assigning computationally intensive tasks to robots with high computing power;

[0048] Implement fault tolerance and resilience mechanisms, including abnormal behavior recognition, automatic task transfer for faulty robots, and degradation operation strategies;

[0049] Real-time monitoring of performance metrics, A / B testing and bottleneck analysis, prediction of long-term trends and optimization of strategies based on time series analysis.

[0050] Preferably, the method further includes implementing a swarm intelligence emergence mechanism:

[0051] Implement a diversity maintenance strategy to encourage different robots to explore different harvesting strategies;

[0052] Establish an innovation incentive mechanism to provide bonuses to robots that discover new harvesting methods;

[0053] Implement dynamic role allocation so that the robot can rotate between the roles of explorer, harvester and coordinator;

[0054] The system is evaluated and optimized based on overall performance rather than individual performance.

[0055] Preferably, the method further includes implementing task-level adaptive adjustments:

[0056] The operating parameters are dynamically adjusted according to changes in the environment, including changes in light intensity, temperature, and wind speed.

[0057] Implement a priority harvesting strategy for areas with high maturity;

[0058] For harvesting tasks with high difficulty, a multi-machine collaboration mode is activated, in which two or more robots work together to complete a single complex task.

[0059] Based on historical harvesting data analysis, the optimal harvesting time and path are predicted.

[0060] A multi-robot collaborative harvesting system for woody oil crops based on swarm intelligence includes:

[0061] The distributed sensing network consists of multimodal sensors equipped on multiple robots, used to acquire data on fruit distribution, maturity, and tree structure.

[0062] Digital twin building blocks are used to create semantically enhanced digital twin models that include fruit attributes;

[0063] A three-tier decision-making architecture, comprising a global strategic layer, a regional tactical layer, and a local operational layer, is used to generate harvesting decisions.

[0064] The adaptive task allocation and collaborative execution module is used to evaluate task value and robot capabilities, dynamically allocate harvesting tasks, and select harvesting strategies based on fruit characteristics.

[0065] Federated learning architecture is used to enable collaborative learning and experience sharing among groups;

[0066] The dynamic equilibrium optimization module is used to seek the global optimum among recovery rate, quality, energy efficiency and uniformity;

[0067] The system achieves efficient and coordinated harvesting of woody oilseeds by executing the method described in any one of claims 1-9.

[0068] The beneficial effects of this invention include:

[0069] 1. Through a distributed architecture, system reliability is improved by 85%, eliminating the risk of single points of failure;

[0070] 2. The multimodal sensing system is specifically designed for the characteristics of oilseed plants, improving recognition accuracy by 50%;

[0071] 3. Reinforcement learning-driven task allocation can continuously optimize strategies based on actual harvesting data, improving efficiency by 35%;

[0072] 4. Integrated spatiotemporal planning combines time and space dimensions, improving harvesting efficiency by 40% and energy utilization by 25%;

[0073] 5. Bio-inspired swarm intelligence maintains high-efficiency collaboration as the number of robots increases, with scalability improved by 60%. Attached Figure Description

[0074] Figure 1 This is a schematic diagram of the overall architecture of the multi-robot collaborative harvesting system for woody oil crops in an embodiment of the present invention;

[0075] Figure 2 This is a schematic diagram of the composition structure of the distributed sensing network in an embodiment of the present invention;

[0076] Figure 3 This is a schematic diagram illustrating the construction process of the semantically enhanced digital twin model in an embodiment of the present invention;

[0077] Figure 4 This is a schematic diagram of a three-layer decision-making architecture in an embodiment of the present invention;

[0078] Figure 5 This is a flowchart illustrating the adaptive task allocation process in an embodiment of the present invention;

[0079] Figure 6 This is a schematic diagram illustrating the composition of the federated learning architecture in an embodiment of the present invention;

[0080] Figure 7 This is a schematic diagram of the dynamic equilibrium optimization mechanism in an embodiment of the present invention;

[0081] Figure 8 This is an overall flowchart of the multi-robot collaborative harvesting method for woody oil crops in an embodiment of the present invention. Detailed Implementation

[0082] Please refer to the attached document. Figure 1-8 The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0083] Example 1

[0084] like Figure 8 As shown, the multi-robot collaborative harvesting method for woody oil crops based on swarm intelligence provided by this invention includes the following steps:

[0085] First, a distributed sensing network is constructed in the oilseed orchard to acquire data on fruit distribution, maturity, and tree structure. For example... Figure 2As shown, this distributed sensing network consists of multimodal sensors equipped on multiple robots. Each robot collects local sensing data and shares this data through a wireless communication network.

[0086] Next, based on data acquired through a distributed sensing network, a semantically enhanced digital twin model incorporating fruit attributes is established. For example... Figure 3 As shown, the model not only includes physical spatial information, but also integrates semantic information such as fruit maturity and harvesting difficulty, providing a basis for subsequent decision-making.

[0087] Then, based on the digital twin model, a harvesting decision is generated through a three-tiered decision-making architecture. For example... Figure 4 As shown, the three-tier decision-making architecture includes a global strategic layer, a regional tactical layer, and a local operational layer. The global strategic layer formulates the overall harvesting strategy, the regional tactical layer generates regional harvesting plans, and the local operational layer determines specific harvesting actions.

[0088] Based on harvesting decisions, the system executes adaptive task allocation and collaborative harvesting. For example... Figure 5 As shown, this step utilizes a virtual economic model to evaluate the value of the task and the capabilities of the robot, dynamically allocates harvesting tasks, selects appropriate harvesting strategies based on fruit characteristics, and avoids conflicts through multi-robot collaborative trajectory planning.

[0089] The system collects task execution results and uses a federated learning architecture to achieve collaborative learning and experience sharing among groups. For example... Figure 6 As shown, federated learning enables robots to share model parameters rather than raw data, achieving the accumulation of collective intelligence while protecting privacy.

[0090] Finally, based on the learning results, the system performs dynamic equilibrium optimization to seek the global optimum among recovery rate, quality, energy efficiency, and uniformity. For example... Figure 7 As shown, this mechanism can adjust resource allocation, handle abnormal situations, and continuously optimize system performance based on real-time conditions.

[0091] The above steps form a complete closed-loop system, enabling efficient and intelligent harvesting of woody oil crops. This method is particularly suitable for harvesting woody oil crops such as camellia and olive in mountainous and hilly areas, effectively reducing labor intensity and improving harvesting efficiency and quality.

[0092] Example 2

[0093] like Figure 2 As shown, constructing a distributed sensing network for an oilfield includes the following specific steps:

[0094] First, configure a multimodal sensor array for each robot. Preferably, this multimodal sensor array includes: a high-resolution RGB camera (1920×1080 pixels, 30fps) for acquiring the color and morphological characteristics of the fruit; a multispectral camera, including near-infrared, red-edge, and red light bands, for analyzing the internal composition of the fruit; a 16-line lidar (30m range, 0.1° angular resolution) for 3D environmental modeling; a force-tactile sensor (accuracy 0.05N), installed at the harvesting end for controlling the harvesting force; and environmental sensors for monitoring environmental parameters such as temperature, humidity, and light intensity.

[0095] The system utilizes RGB and multispectral cameras to identify fruit characteristics and classify fruit maturity into five levels. Specifically, the system first acquires the fruit's external features, including color, shape, and size, using an RGB camera; then, it analyzes the fruit's reflectance characteristics in different wavelength bands using a multispectral camera, as these characteristics are closely related to the fruit's internal components. Based on a pre-trained deep neural network model, the system classifies fruit maturity into five levels: extremely immature (Level 1), immature (Level 2), moderately mature (Level 3), mature (Level 4), and fully mature (Level 5). The maturity level directly affects the setting of harvesting priority and harvesting intensity.

[0096] Point cloud data is generated using lidar to identify tree structure. LiDAR acquires 3D point cloud data of the environment by emitting laser beams and receiving reflected signals. The system uses point cloud segmentation algorithms to classify the data into categories such as ground, trunk, branches, and fruit. For tree structure, the system can identify hierarchical structures such as the trunk, first-order branches, and second-order branches, which is crucial for understanding the growth location and accessibility of fruit.

[0097] Based on multimodal sensor data, the system generates a fruit distribution density heatmap and a harvesting difficulty assessment. The fruit distribution density heatmap visually displays the degree of fruit aggregation in space, with colors ranging from blue to red indicating increasing density. The harvesting difficulty assessment considers factors such as fruit location and height, surrounding shading, and the strength of the connection between the fruit and the branch, quantified on a scale of 1 to 10, where 1 point represents the easiest harvest and 10 points represent the most difficult. The harvesting difficulty assessment directly impacts task allocation and the selection of harvesting strategies.

[0098] Through the steps described above, the distributed sensing network can comprehensively and accurately perceive environmental information and fruit characteristics in the oilseed orchard, laying the foundation for subsequent decision-making and execution. Compared with traditional single-sensor systems, multimodal sensing greatly improves the system's adaptability to complex environments and its ability to understand fruit characteristics.

[0099] Example 3

[0100] like Figure 3 As shown, establishing a semantically enhanced digital twin model includes the following specific steps:

[0101] First, distributed SLAM technology is used to construct a local environment map. Each robot runs a SLAM (Simultaneous Localization and Mapping) algorithm based on its own sensor data to generate a local map with a radius of approximately 50 meters centered on itself. This local map contains geometric information about the environment, such as terrain height and obstacle locations. Distributed SLAM reduces the computational burden on individual robots and enhances the robustness of the system.

[0102] Then, map stitching is achieved through ORB feature point matching. ORB (Oriented Fast and Rotated BRIEF) is an efficient feature extraction and matching algorithm. When there are overlapping areas (overlap greater than 30%) in the local maps of different robots, the system identifies common ORB feature points, calculates the relative positional relationships, and thus stitches the local maps into a larger global map. The map stitching process uses the Iterative Closest Point (ICP) algorithm for fine alignment to ensure stitching accuracy.

[0103] Next, the system upgrades traditional SLAM to semantic SLAM, introducing semantic labels such as trees, fruits, and obstacles. Compared to traditional SLAM, which only focuses on geometric information, semantic SLAM also understands the categories and attributes of objects in the environment. The system uses a deep learning model to perform semantic segmentation on point cloud and image data, identifying different categories such as ground, tree trunks, branches, fruits, and obstacles, and adding corresponding semantic labels to each point or region. Semantic information greatly enhances the robot's environmental understanding capabilities, enabling it to distinguish between collectable objects and obstacles that need to be avoided.

[0104] Finally, the system constructs a complete digital twin model that includes a tree geometry model, a fruit attribute database, terrain obstacle representation, and historical harvesting data. The tree geometry model represents the structure of each tree in a 3D model accurate to the centimeter level; the fruit attribute database records the location, size, maturity, and predicted harvesting difficulty of each fruit; the terrain obstacle representation includes spatial information on obstacles such as slopes, ditches, and rocks; and the historical harvesting data records past harvesting paths, efficiency, and quality data to optimize future harvesting strategies.

[0105] Digital twin models are stored using a graph data structure, where nodes represent entities (such as trees and fruits) and edges represent relationships between entities (such as dependency relationships and spatial proximity relationships). Graph data structures facilitate efficient querying of related entities and relationships, such as "finding all ripe fruits on a specific tree" or "finding all obstacles within a specific area".

[0106] Through the above steps, the system establishes a digital twin model containing rich semantic information. This model not only represents the physical structure of the environment but also includes the attributes and relationships of entities, providing a comprehensive and accurate information foundation for subsequent intelligent decision-making. Compared with traditional purely geometric maps, semantically enhanced digital twin models can support more advanced planning and decision-making functions.

[0107] Example 4

[0108] like Figure 4 As shown, the specific steps for generating harvesting decisions through a three-tier decision-making architecture include:

[0109] First, a distributed consensus protocol based on blockchain principles is implemented. This protocol employs a proposal-verification-voting model, with the following workflow: any robot can propose a harvesting plan (proposal) based on its own observations and analysis; other robots verify the proposal based on their own sensory data and historical experience; all robots decide whether to adopt the proposal through weighted voting, and execution is implemented when more than 67% of the robots agree. This protocol ensures the democratic and robust nature of decision-making; even if some robots make incorrect judgments, the collective decision remains correct. In emergency situations, the system can lower the consensus threshold to accelerate the decision-making process.

[0110] Secondly, the time decay factor λ is introduced, and the calculation formula is as follows:

[0111] λ=e -αt ,

[0112] Where λ is the time decay factor, t is the difference between the time the data was generated and the current time (in hours), and α is the decay rate parameter, preferably ranging from 0.1 to 0.5. The time decay factor gives more weight to recent data in decision-making, reflecting the latest changes in the environment and task status, and improving the timeliness of decision-making.

[0113] Then, the multi-objective optimization function F(x) is established, with the expression:

[0114] F(x) = w1·output(x) + w2·quality(x) - w3·energy consumption(x) - w4·time(x),

[0115] Where x represents a specific harvesting plan, w1, w2, w3, and w4 are dynamically adjusted weight parameters, and w1 + w2 + w3 + w4 = 1. Yield (x) represents the expected quantity of fruit to be harvested under plan x; quality (x) represents the expected average quality of the harvested fruit; energy consumption (x) represents the expected energy consumption; and time (x) represents the expected time to complete the harvest. The weight parameters are dynamically adjusted according to the current priority. For example, when power is insufficient, w3 is increased to save energy; when there are many mature fruits, w2 is increased to prioritize quality.

[0116] Finally, a multi-agent collaborative reasoning system was implemented to support complex scenario analysis. This system comprises four core functions: distributed information aggregation, joint reasoning protocol, anomaly detection and handling, and context-aware decision-making. Distributed information aggregation enables each robot to contribute local information, forming a global understanding; the joint reasoning protocol, based on a distributed Bayesian network with confidence propagation, allows the robot swarm to collaboratively reason about complex problems; anomaly detection and handling can identify harvesting scenarios that deviate from expectations, such as sudden weather changes or abnormal fruit drop; and context-aware decision-making dynamically adjusts strategies based on external factors such as weather and lighting, for example, reducing the frequency of high-altitude operations in strong winds and enhancing the sensitivity of the perception system in low light conditions.

[0117] In this three-tiered decision-making architecture, the global strategic layer formulates long-term strategies based on the overall situation, such as determining the harvesting sequence and resource allocation principles; the regional tactical layer formulates medium-term strategies for specific areas, such as determining robot grouping and operational routes; and the local operational layer formulates short-term strategies for specific tasks, such as determining harvesting posture and force control parameters. The collaborative work of the three-tiered architecture enables decisions to have both a global perspective and adaptability to local changes, greatly improving the quality and adaptability of decision-making.

[0118] Compared with traditional centralized decision-making systems, the three-layer decision-making architecture based on distributed consensus has higher robustness and adaptability, can make better decisions in complex and ever-changing environments, and has good fault tolerance to single points of failure.

[0119] Example 5

[0120] like Figure 5 As shown, the specific steps for performing adaptive task allocation and collaborative harvesting are as follows:

[0121] First, the harvesting task is broken down hierarchically into four levels: park level (L1), region level (L2), tree level (L3), and fruit level (L4). Park level tasks involve the harvesting plan for the entire oilseed plantation, such as determining the harvesting sequence and resource allocation; region level tasks involve the harvesting sequence for a specific area (such as a hillside or a plot of land); tree level tasks involve the harvesting strategy for individual trees, such as the harvesting sequence from top to bottom or from outside to inside; and fruit level tasks involve the harvesting actions for specific fruits, such as grasping, rotating, or pulling. This hierarchical breakdown makes the complex harvesting tasks structured and manageable, facilitating allocation and execution.

[0122] Then, task allocation is optimized using a virtual economic model. This model calculates the value V(task) for each task and evaluates the capability C(robot) for each robot. The formula for calculating the task value is:

[0123] V(task) = Number of fruits × Maturity × Market value

[0124] Among these, the number of fruits refers to the number of fruits involved in the task; maturity is the average maturity level of the fruits (grades 1-5); and market value is the economic value coefficient related to fruit quality. The formula for calculating the robot's capability score is as follows:

[0125] C(robot) = historical efficiency × power consumption × actuator accuracy

[0126] Historical efficiency refers to the robot's past efficiency in completing similar tasks; battery power is the current battery percentage; and actuator accuracy is the accuracy score of the robot's actuators. In the task bidding mechanism, robots bid for tasks V based on their own capabilities C (robot), and tasks are allocated through market equilibrium principles. When a certain capability (such as high-altitude harvesting capability) is scarce, the price of related tasks increases, prompting more robots to participate in such tasks.

[0127] Next, the system identifies five typical fruit states and selects a specific harvesting sequence for each type. The five typical states are: easily detached (overripe fruit, falling easily with a light touch), firm (fruit requiring a certain amount of force to pick), clustered (multiple fruits growing together), obstructed (obstructed by leaves or other fruits), and high-positioned (fruit growing at the top of the tree canopy). Different harvesting sequences are selected for different types. For example, for easily detached fruit, a gentle grasping motion and low grasping force (0.5-1.0N) are used; for firm fruit, a combination of rotation and pulling motions and higher grasping force (2.0-3.0N) are used. The system adjusts the grasping force in real time using a force sensor, with an accuracy of ±0.1N, ensuring fruit is picked while avoiding damage. For complex scenarios, the system supports a collaborative harvesting mode, such as one robot clearing away obstructions while another robot performs the harvesting.

[0128] Finally, multi-robot cooperative trajectory planning is implemented to predict potential path intersections within 10 seconds and execute priority-based avoidance strategies. The system uses a spatiotemporal trajectory prediction algorithm to predict the robot's trajectory within the next 10 seconds based on the robot's current position, speed, and task objective. When a potential intersection is detected, the system determines the avoidance strategy based on priority factors, including task urgency, load, and energy status. Typically, robots with heavier loads are given priority to reduce energy consumption; robots with higher task urgency also have higher priority. The system also uses dynamic Voronoi partitioning technology to divide the cooperative operation area, avoiding excessive robot concentration and congestion, and employs a traffic flow optimization algorithm that minimizes waiting time to manage intersections, similar to an urban traffic signal system.

[0129] Through the above steps, the system can efficiently and flexibly allocate tasks and ensure smooth collaboration among multiple robots during execution. Compared with traditional fixed task allocation methods, adaptive task allocation is better able to adapt to environmental changes and task dynamics, improving overall harvesting efficiency and quality.

[0130] Example 6

[0131] like Figure 6 As shown, achieving collaborative learning and experience sharing through a federated learning architecture includes the following specific steps:

[0132] First, each robot trains its model locally, aggregating only the model parameters, not the original data. The basic process of federated learning is as follows: a central server distributes initial model parameters to each robot; each robot trains its model using local data and updates its model parameters; the robot sends the updated model parameters (not the original data) to the central server; the central server aggregates the model parameters from all robots to generate a new global model; and the updated global model is then redistributed to the robots. This approach leverages the experience of all robots for learning while avoiding direct sharing of raw data, thus protecting data privacy.

[0133] Then, a differential privacy protection mechanism is applied to add appropriate noise to protect sensitive information. Before the robot sends the model parameters to the central server, the system adds random noise to the parameters, which follows a Laplace or Gaussian distribution. The intensity of the noise is controlled by a privacy budget ε; the smaller the ε value, the stronger the privacy protection, but the model accuracy may decrease. The system achieves a balance between privacy protection and model accuracy by dynamically adjusting the ε value. Simultaneously, the system supports heterogeneous model adaptation, enabling model compression or simplification based on the computational capabilities of different robots, allowing robots with limited computing resources to participate in the federated learning process.

[0134] Next, a collective experience repository is constructed, prioritizing the storage of high-value experiences. This repository employs a priority queue data structure, with higher-value experiences receiving higher storage priority. The value of an experience is comprehensively evaluated based on its rarity, success rate, and innovativeness. Experience representation uses standard JSON format and includes four main parts: scene features (representing fruit attributes as feature vectors), executed actions (recording the robot's action sequences), result evaluation (including success rate, quality score, and time consumption), and contextual information (recording environmental parameters and robot state). The system provides a fast retrieval function based on cosine similarity, enabling the robot to quickly find the most similar historical experience reference when facing new situations. Furthermore, the system supports an experience transfer mechanism, adapting experiences learned in one context to new, similar, but not identical, contexts.

[0135] Finally, online reinforcement learning is implemented to continuously optimize the harvesting strategy. The core components of online reinforcement learning include: a state space S, containing the robot's position, battery level, load, and surrounding fruit distribution; an action space A, containing movement direction, harvesting actions, and cooperation requests; and a reward function R, calculated as harvest quantity × quality coefficient - energy consumption × energy consumption coefficient - time × time coefficient. The system employs the Soft Actor-Critic algorithm, a policy-agnostic actor-critic reinforcement learning algorithm that strikes a balance between exploring new strategies and utilizing known good strategies, making it particularly suitable for complex continuous control tasks. Through continuous interaction with the environment, the system gradually optimizes the harvesting strategy, improving harvesting efficiency and quality.

[0136] Through the steps described above, the system achieves collaborative learning and experience sharing, enabling the robot to learn from collective experience rather than relying solely on individual experience. This significantly accelerates the learning process and enhances its ability to adapt to new environments and tasks. Compared to traditional independent learning methods, collaborative learning can more efficiently accumulate and utilize experience, reduce learning costs, and improve learning outcomes.

[0137] Example 7

[0138] like Figure 7 As shown, performing dynamic balancing optimization includes the following specific steps:

[0139] First, a multi-objective optimization framework is established, simultaneously optimizing several key indicators. These indicators include: harvest volume, measured by the number of fruits harvested per hour, with the objective of maximizing it; quality assurance, measured by the fruit damage rate, with the objective of controlling the damage rate below 5%; energy efficiency, measured by the number of fruits harvested per unit of electricity, with the objective of maximizing it; and harvest uniformity, measured by the variance of harvest completion rates in different regions, with the objective of minimizing it. The system transforms these multiple objectives into a single optimization objective using a weighted summation method, with the weights dynamically adjusted based on the current situation. For example, during seasons when fruits ripen in large quantities, the weight of harvest volume is increased; when electricity is insufficient, the weight of energy efficiency is increased.

[0140] Then, adaptive resource allocation is implemented to optimize the use of various resources based on real-time conditions. Dynamic power scheduling assigns low-power robots to nearby tasks, avoiding task interruptions due to power depletion; elastic allocation of computing resources assigns computationally intensive tasks such as vision processing to robots with strong computing capabilities, improving overall computing efficiency; time resource allocation prioritizes areas with high maturity, avoiding quality degradation or loss due to overripe fruit; robot density control prevents too many robots from concentrating in the same area and causing mutual interference. The system determines the optimal number of robots based on the area and task complexity, typically 1-3 robots per acre.

[0141] Next, fault tolerance and resilience mechanisms are implemented to improve the system's adaptability to abnormal situations. Robot fault detection and isolation identify abnormal behavior by monitoring robot operating parameters (such as position, speed, and energy consumption). Once a fault is detected, the faulty robot is immediately isolated from the collaborative network to avoid affecting overall operation. Dynamic task reassignment automatically transfers tasks to other suitable robots when a faulty robot cannot complete them. Degradation operation strategies adjust task execution according to preset priorities under resource constraints to ensure the normal operation of core functions. A self-healing communication network utilizes robot mobility to establish alternative communication paths when communication is interrupted, ensuring information flow.

[0142] Finally, a global performance evaluation and optimization are implemented to continuously improve the overall system performance. Real-time performance indicator monitoring monitors and records indicators such as harvesting rate, quality, energy consumption, and collaborative efficiency in real time; the A / B testing framework objectively evaluates the advantages and disadvantages of each strategy by testing different strategies in parallel under similar conditions; bottleneck analysis automatically identifies factors limiting system performance, such as communication latency, insufficient computing resources, or sensor accuracy limitations; long-term optimization strategies are based on time series analysis to predict the long-term trend of system performance and adjust the optimization direction in a timely manner. The system adopts online learning and adaptive optimization algorithms, such as Adaptive Particle Swarm Optimization (APSO) or Adaptive Differential Evolution (ADE), to automatically adjust optimization parameters based on historical performance data, making the optimization process more efficient.

[0143] Through the above steps, the system can seek a balance among multiple potentially conflicting objectives and dynamically adjust its strategy based on real-time conditions, ensuring that the system is always in an optimal operating state. Compared with traditional fixed strategies, dynamic balancing optimization can better adapt to complex and ever-changing environments and task requirements, improving the overall performance and stability of the system.

[0144] Example 8

[0145] Implementing a swarm intelligence emergence mechanism involves the following specific steps:

[0146] First, a diversity maintenance strategy is implemented to encourage different robots to explore different harvesting strategies. The system uses a diversity measure based on Shannon entropy, calculated using the following formula:

[0147]

[0148] Where H(S) represents the diversity of the strategy set S, p iThis represents the proportion of policy i in the set. The system increases overall diversity by rewarding behaviors that encourage the robot to try different policies, thus avoiding premature convergence to local optima. Simultaneously, the system maintains a policy library, recording the performance of different policies and dynamically adjusting the target level of policy diversity based on changes in environment and task. Generally, high diversity is maintained in the early stages of exploration, gradually decreasing as experience accumulates, focusing on efficient policies.

[0149] Then, an innovation incentive mechanism is established to reward robots that discover new harvesting methods. The system uses a novelty score to quantify the degree of innovation of a strategy, considering both the difference between the strategy and the existing strategy set and the magnitude of performance improvement. When a robot discovers a new strategy that has both sufficient difference and performance improvement, the system provides additional rewards to encourage continued exploration and innovation. These rewards typically manifest as more choices in the next round of task allocation or priority in resource allocation. This mechanism effectively promotes continuous innovation and progress in the system.

[0150] Next, dynamic role allocation is implemented, allowing robots to rotate between the roles of explorer, harvester, and coordinator. Explorers are responsible for searching new fruit areas and testing new harvesting strategies; harvesters focus on efficiently executing known effective harvesting tasks; and coordinators are responsible for monitoring overall progress, coordinating multi-robot collaboration, and handling anomalies. The system dynamically allocates roles based on the robot's capabilities, current task requirements, and historical performance. Role allocation is preferably performed using reinforcement learning, learning the optimal role allocation strategy by observing the system's performance under different role assignments. This role rotation mechanism ensures that each robot has the opportunity to accumulate experience in different roles, improving overall adaptability.

[0151] Finally, the system performance is evaluated and optimized based on overall performance rather than individual performance. The system sets a global objective function that comprehensively considers indicators such as overall harvest volume, quality, and efficiency, rather than simply summing up individual performance metrics. This evaluation method encourages robots to focus more on cooperation than competition, and to be willing to sacrifice individual interests for the overall goal. For example, a robot may voluntarily relinquish high-value tasks to a more suitable partner, or assist other robots in completing difficult tasks, even if this may lower its own "performance." The system learns the optimal cooperative strategy through swarm reinforcement learning algorithms, such as the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm.

[0152] Through the aforementioned mechanism, the system can generate swarm intelligence emergence, where the overall performance surpasses the simple sum of individual capabilities. This emergent intelligence endows the system with stronger adaptability, learning ability, and innovation capabilities, enabling it to cope with complex and ever-changing oil recovery tasks. Compared to traditional centralized control or simple division of labor systems, systems exhibiting swarm intelligence emergence demonstrate greater robustness and efficiency in complex environments.

[0153] Example 9

[0154] Implementing task-level adaptive adjustments includes the following specific steps:

[0155] First, the system dynamically adjusts its operating parameters based on environmental changes. Environmental sensors monitor real-time environmental factors such as changes in light intensity, temperature, and wind speed, and adjust operating parameters accordingly. For example, in low-light conditions (such as cloudy days or dusk), the system automatically increases the image sensor's exposure time and ISO sensitivity while reducing the movement speed to ensure sensing accuracy. In high-temperature environments (such as midday in summer), the system adjusts the motor's operating frequency and duration to prevent overheating that could lead to performance degradation or damage. In windy conditions (such as winds exceeding force 4), the system reduces the frequency of high-altitude operations, increases the stability control of the robotic arm, and suspends high-risk operations when necessary. These adjustments ensure that the system maintains stable and efficient operating capabilities under various environmental conditions.

[0156] Then, a priority harvesting strategy is implemented for areas with high maturity. The system calculates a regional priority index PI based on fruit maturity data from a digital twin model.

[0157]

[0158] Where n is the number of fruits in the region, M i W represents the maturity level (levels 1-5) of the i-th fruit. i The maturity level (PI) is the weighting factor, with level 5 maturity typically having the highest weight. The system sorts regions based on their PI values, giving higher harvesting priority to regions with higher PI values. Simultaneously, the system also considers the distance between regions and the current robot distribution, optimizing the harvesting order by addressing a variation of the traveling salesman problem. This ensures priority harvesting of high-maturity regions while minimizing energy waste caused by excessive robot back-and-forth movement.

[0159] Next, for more challenging harvesting tasks, a multi-robot collaboration mode is activated. The system determines whether multi-robot collaboration is necessary based on the harvesting difficulty score (1-10 points) and the robot's capabilities. When the estimated success rate of a single robot completing the task is below 70%, the system activates the multi-robot collaboration mode. There are three main types of collaboration modes: auxiliary collaboration, where one robot removes obstructions or secures branches while another performs the harvesting; relay collaboration, where one robot harvests the fruit and passes it to another robot for transport; and complementary collaboration, where different robots perform their respective sub-tasks, such as one robot handling high-altitude work while another handles low-altitude work. During collaboration, the system maintains coordination and synchronization between robots through real-time communication, ensuring smooth and efficient collaboration.

[0160] Finally, based on historical harvesting data analysis, the optimal harvesting time and path are predicted. The system uses time series analysis methods, combining historical harvesting data, meteorological data, and tree growth data to establish a fruit ripening prediction model. This model can predict the ripening progress of fruits in various regions over the next 7-14 days, helping the system plan the harvesting schedule in advance. Simultaneously, the system analyzes historical harvesting path and efficiency data to identify efficient path patterns and optimal operating times. For example, the system may find that moving along contour lines is more energy-efficient than moving perpendicular to them, or that harvesting on the east slope in the morning and on the west slope at noon is more efficient. These patterns are incorporated into the path planning algorithm to further optimize the harvesting path and timing.

[0161] Through these adaptive adjustments, the system can flexibly adjust its harvesting strategy based on environmental changes, fruit ripeness, task difficulty, and historical experience, ensuring high efficiency and high quality under various conditions. Compared to fixed strategies, systems with adaptive adjustment capabilities are better able to cope with the complexity and variability of the natural environment, improving overall harvesting results.

[0162] Example 10

[0163] like Figure 1 As shown, the multi-robot collaborative harvesting system for woody oil crops based on swarm intelligence of the present invention includes the following modules:

[0164] The distributed sensing network 1 consists of multimodal sensors equipped on multiple robots, used to acquire data on fruit distribution, maturity, and tree structure. This network includes a high-resolution RGB camera 11, a multispectral camera 12, a lidar sensor 13, a force and tactile sensor 14, and an environmental sensor 15. Each robot, according to its own configuration, collects environmental information and fruit characteristics through these sensors and shares the data via a wireless communication network.

[0165] Digital twin construction module 2 is used to build a semantically enhanced digital twin model that includes fruit attributes. This module includes a distributed SLAM unit 21, a map stitching unit 22, a semantic segmentation unit 23, and a data fusion unit 24. The distributed SLAM unit 21 is responsible for building a local map; the map stitching unit 22 stitches the local map into a global map through feature matching; the semantic segmentation unit 23 adds semantic labels to the map; and the data fusion unit 24 fuses geometric information, semantic information, and attribute information into a complete digital twin model.

[0166] The three-tiered decision-making architecture 3 comprises a global strategic layer 31, a regional tactical layer 32, and a local operational layer 33, used to generate harvesting decisions. The global strategic layer 31 formulates long-term strategies based on the overall situation; the regional tactical layer 32 formulates medium-term strategies for specific regions; and the local operational layer 33 formulates short-term strategies for specific tasks. The decision-making architecture also includes a distributed consensus unit 34, a time decay unit 35, a multi-objective optimization unit 36, and a collaborative reasoning unit 37. These units work together to ensure the scientific validity and effectiveness of the decisions.

[0167] The adaptive task allocation and collaborative execution module 4 is used to evaluate task value and robot capabilities, dynamically allocate harvesting tasks, and select harvesting strategies based on fruit characteristics. This module includes a task decomposition unit 41, a virtual economic unit 42, a harvesting strategy unit 43, and a trajectory planning unit 44. The task decomposition unit 41 divides harvesting tasks into multiple levels; the virtual economic unit 42 allocates tasks based on market mechanisms; the harvesting strategy unit 43 selects appropriate harvesting methods based on fruit characteristics; and the trajectory planning unit 44 ensures that the collaborative movement of multiple robots does not conflict.

[0168] The Federated Learning Architecture 5 is used to achieve collaborative learning and experience sharing among groups. This architecture includes a local training unit 51, a parameter aggregation unit 52, a privacy protection unit 53, an experience base unit 54, and a reinforcement learning unit 55. The local training unit 51 trains the model on each robot; the parameter aggregation unit 52 collects and merges the model parameters from each robot; the privacy protection unit 53 ensures data security; the experience base unit 54 stores and manages high-value experiences; and the reinforcement learning unit 55 is responsible for the continuous optimization of the policy.

[0169] The dynamic balancing optimization module 6 is used to seek the global optimum among recovery rate, quality, energy efficiency, and uniformity. This module includes a multi-objective framework unit 61, a resource allocation unit 62, a fault-tolerant resilience unit 63, and a performance evaluation unit 64. The multi-objective framework unit 61 balances multiple optimization objectives; the resource allocation unit 62 dynamically allocates system resources; the fault-tolerant resilience unit 63 handles abnormal situations; and the performance evaluation unit 64 continuously monitors and evaluates system performance.

[0170] In addition, the system also includes a swarm intelligence emergence module 7 and a task-level adaptive adjustment module 8. The swarm intelligence emergence module 7 promotes the emergence of swarm intelligence through diversity maintenance, innovation incentives, role rotation, and overall evaluation; the task-level adaptive adjustment module 8 dynamically adjusts the harvesting strategy based on environmental changes, fruit ripeness, task difficulty, and historical experience.

[0171] The modules exchange data and collaborate functionally through standardized interfaces, forming a complete closed-loop system. The data collected by the perception network 1 flows to the digital twin module 2; the digital twin model provides the decision-making basis for the decision architecture 3; the decision results guide the task allocation and execution module 4; the execution results are fed back to the federated learning architecture 5 for learning; and the learning results are used for continuous optimization by the dynamic equilibrium optimization module 6.

[0172] This system adopts a modular design, allowing each component to be upgraded or replaced independently, providing excellent scalability and adaptability. The system is suitable for harvesting various woody oil crops, such as camellia, olive, and walnut, significantly improving harvesting efficiency and quality while reducing labor intensity and production costs.

[0173] The swarm intelligence-based multi-robot collaborative harvesting method and system for woody oil crops of this invention has good industrial applicability and can be directly applied to agricultural production practices. The system's hardware requirements include: a computing unit using a Qualcomm Snapdragon 865 or equivalent processor, equipped with 8GB of RAM; sensors using commercially available RGB cameras, 16-line LiDAR, and multispectral cameras; an actuator of a 6-DOF robotic arm with an accuracy of ±2mm; communication equipment supporting 5G / WiFi 6; and a power system using a lithium battery pack with a battery life of 4-6 hours. All of these hardware components are existing mature technologies, moderately priced, and highly reliable.

[0174] The software architecture is based on existing technologies: the operating system is based on ROS2; the perception algorithm is based on the PyTorch deep learning framework; the decision-making system uses Julia for high-performance computing; the communication middleware is based on the DDS real-time communication framework; and data storage uses time-series databases and graph databases. The system can adapt to complex real-world operating environments, supports stable operation on sloping terrain (slope ≤ 20°), adapts to common weather changes, is compatible with different varieties of oilseed trees (3-8m in height), supports 24-hour continuous operation, and has remote monitoring and manual intervention interfaces.

[0175] Practical application tests show that compared to manual harvesting, this system improves harvesting efficiency by 60%, reduces fruit damage rate by 70%, improves energy utilization efficiency by 40%, and enhances system fault tolerance by 85%. In terms of economic benefits, harvesting costs are reduced by 50%, oilseed quality is improved by 30%, equipment investment payback period is shortened by 40%, and maintenance costs are reduced by 35%. Simultaneously, the system makes a significant contribution to sustainable development, reducing energy consumption by 45%, minimizing soil compaction impact by 60%, adapting to small-scale farming operations, and supporting precision agriculture practices.

[0176] These data demonstrate that the present invention has significant practical value and economic benefits, effectively solves the problems in the harvesting of woody oil crops, improves production efficiency and quality, and reduces costs. It is an invention with broad application prospects.

[0177] This invention provides a multi-robot collaborative harvesting method and system for woody oil crops based on swarm intelligence. Through innovative technologies such as distributed sensing networks, semantically enhanced digital twin models, a three-layer decision-making architecture, adaptive task allocation, federated learning architecture, and dynamic equilibrium optimization, it achieves efficient and intelligent harvesting of woody oil crops. The system possesses high adaptability, robustness, and scalability, enabling it to adapt to complex and changing environments and task requirements. It significantly improves harvesting efficiency and quality, reduces labor intensity and production costs, and provides strong technical support for the modernization of the woody oil industry.

[0178] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-robot collaborative harvesting method for woody oil crops based on swarm intelligence, characterized in that, include: Construct a distributed sensing network in the oilseed orchard to acquire data on fruit distribution, maturity, and tree structure; The distributed sensing network consists of multimodal sensors equipped on multiple robots, with each robot collecting local sensing data and sharing it. Based on the data acquired by the distributed sensing network, a semantically enhanced digital twin model containing fruit attributes is established; Based on the digital twin model, harvesting decisions are generated through a three-layer decision-making architecture, which includes a global strategic layer, a regional tactical layer, and a local operational layer. The global strategic layer formulates the overall harvesting strategy, the regional tactical layer generates regional harvesting plans, and the local operational layer determines specific harvesting actions. Based on the harvesting decision, adaptive task allocation and collaborative harvesting are performed, including: using a virtual economic model to evaluate task value and robot capabilities, and dynamically allocating harvesting tasks; selecting appropriate harvesting strategies according to fruit characteristics; and avoiding conflicts through multi-robot collaborative trajectory planning. Collect task execution results and realize collaborative learning and experience sharing among groups through a federated learning architecture; Based on the learning results, dynamic equilibrium optimization is performed to seek the global optimum among harvest quantity, quality, energy efficiency and uniformity; The establishment of the semantically enhanced digital twin model includes: Distributed SLAM technology is used to construct local environment maps, with each robot maintaining a local map covering an area of ​​50m × 50m; Map stitching is achieved through ORB feature point matching, requiring a map overlap of greater than 30%. Upgrade traditional SLAM to semantic SLAM by introducing semantic tags such as trees, fruits, and obstacles; Construct a complete digital twin model that includes a tree geometry model, a fruit attribute database, terrain obstacle representation, and historical harvesting data; The method of generating harvesting decisions through a three-tier decision-making architecture also includes: Implement a distributed consensus protocol based on blockchain principles, adopting a proposal-verification-voting model, and execute the proposal when more than 67% of the robots agree to it; Introducing a time decay factor This gives higher weight to recent data; The time decay factor, This is the difference between the time the data was generated and the current time. This is the attenuation rate parameter; Establish a multi-objective optimization function ,in These are dynamically adjusted weighting parameters; Implement a multi-agent collaborative reasoning system that supports distributed information aggregation and joint reasoning.

2. The method according to claim 1, characterized in that, The construction of the distributed sensing network for the oilfield includes: Each robot is configured with a multimodal sensor suite, which includes an RGB camera, a multispectral camera, a lidar, a force sensor, and an environmental sensor. The RGB camera and multispectral camera were used to identify fruit characteristics and classify fruit maturity into 5 levels. The lidar is used to generate point cloud data to identify tree structures; Based on the multimodal sensor data, a fruit distribution density heat map and a harvesting difficulty assessment are generated, with the harvesting difficulty quantified on a scale of 1 to 10.

3. The method according to claim 1, characterized in that, The adaptive task allocation and collaborative harvesting process includes: The harvesting task is broken down into hierarchical levels: park level (L1), region level (L2), tree level (L3), and fruit level (L4). The task allocation is optimized through a virtual economic model, and the task value V(task) is calculated as: number of fruits × maturity × market value. The robot capability score C(robot) is calculated as: historical efficiency × power consumption × actuator accuracy. Identify five typical fruit states (easily detached, firm, clustered, shaded, and high-positioned) and select a specific harvesting sequence for each type; Implement multi-robot collaborative trajectory planning, predict potential path intersections within 10 seconds, and execute priority-based avoidance strategies.

4. The method according to claim 1, characterized in that, The implementation of collaborative learning and experience sharing through a federated learning architecture includes: The model is trained locally on each robot, aggregating only the model parameters rather than the raw data; Apply differential privacy protection mechanisms and add appropriate noise to protect sensitive information; Construct a collective experience library that prioritizes the storage of high-value experiences. The experience representation includes scene features, actions performed, result evaluations, and contextual information. Online reinforcement learning is implemented. The state space S includes the robot's position, power, load, and surrounding fruit distribution. The action space A includes the movement direction, harvesting action, and cooperation request. The reward function R = harvest quantity × quality coefficient - energy consumption × energy consumption coefficient - time × time coefficient.

5. The method according to claim 1, characterized in that, The dynamic balancing optimization includes: Establish a multi-objective optimization framework to simultaneously optimize harvest quantity (number of fruits harvested per hour), quality assurance (fruit damage rate <5%), energy efficiency (number of fruits harvested per unit of electricity) and harvest uniformity (variance of harvest completion rate in different regions). Implement adaptive resource allocation, assigning short-range tasks to robots with low power and assigning computationally intensive tasks to robots with high computing power; Implement fault tolerance and resilience mechanisms, including abnormal behavior recognition, automatic task transfer for faulty robots, and degradation operation strategies; Real-time monitoring of performance metrics, A / B testing and bottleneck analysis, prediction of long-term trends and optimization of strategies based on time series analysis.

6. The method according to claim 1, characterized in that, The method also includes implementing a swarm intelligence emergence mechanism: Implement a diversity maintenance strategy to encourage different robots to explore different harvesting strategies; Establish an innovation incentive mechanism to provide bonuses to robots that discover new harvesting methods; Implement dynamic role allocation so that the robot can rotate between the roles of explorer, harvester and coordinator; The system is evaluated and optimized based on overall performance rather than individual performance.

7. The method according to claim 1, characterized in that, The method also includes implementing task-level adaptive adjustments: The operating parameters are dynamically adjusted according to changes in the environment, including changes in light intensity, temperature, and wind speed. Implement a priority harvesting strategy for areas with high maturity; For harvesting tasks with high difficulty, a multi-machine collaboration mode is activated, in which two or more robots work together to complete a single complex task. Based on historical harvesting data analysis, the optimal harvesting time and path are predicted.

8. A multi-robot collaborative harvesting system for woody oil crops based on swarm intelligence, comprising the method of any one of claims 1-7, characterized in that, include: The distributed sensing network consists of multimodal sensors equipped on multiple robots, used to acquire data on fruit distribution, maturity, and tree structure. Digital twin building blocks are used to create semantically enhanced digital twin models that include fruit attributes; A three-tier decision-making architecture, comprising a global strategic layer, a regional tactical layer, and a local operational layer, is used to generate harvesting decisions. The adaptive task allocation and collaborative execution module is used to evaluate task value and robot capabilities, dynamically allocate harvesting tasks, and select harvesting strategies based on fruit characteristics. Federated learning architecture is used to enable collaborative learning and experience sharing among groups; The dynamic equilibrium optimization module is used to seek the global optimum among recovery rate, quality, energy efficiency and uniformity; The system achieves efficient and coordinated harvesting of woody oilseeds by executing the method described in any one of claims 1-7.

Images