Revolutionary Cooperative Strategy Optimizes Agricultural Robot Swarms for Complex Farmland Environments

Agricultural automation has long been hailed as the cornerstone of addressing the global challenges of rural population aging, labor shortages, and the demand for higher efficiency in precision agriculture. Unmanned farms, equipped with intelligent decision-making and automated control systems, have emerged as the future of agricultural development, yet the complexity of real-world farmland environments—including irregular field shapes, scattered obstacles, and heterogeneous robot performance—has posed persistent barriers to the seamless collaborative operation of agricultural robot swarms. A groundbreaking study led by researchers from Shandong University of Technology has unveiled a novel task allocation and full-area coverage strategy for heterogeneous agricultural robot swarms, designed to rationalize workload distribution, minimize traversal redundancy, and elevate the overall operational efficiency of robot teams in complex farmland settings. This research not only addresses the core pain points of multi-robot collaboration in agriculture but also provides a robust theoretical and technical framework for the large-scale application of unmanned farm systems worldwide.

The research team recognized that the heterogeneity of agricultural robots—stemming from differences in historical workload, energy consumption, failure rates, service quality, and operational efficiency due to equipment aging and wear—is a critical factor often overlooked in traditional multi-robot coverage strategies. Unlike homogeneous robot swarm models that assume identical performance parameters, the new strategy centers on the overall efficiency of the robot team, establishing a scientific task allocation mechanism that selects optimal robots for operation and distributes workloads proportionally based on their actual performance. This approach ensures that each robot’s capabilities are fully utilized, avoiding overburdening of low-efficiency units and underutilization of high-performance ones, a key departure from one-size-fits-all allocation methods that have limited the practicality of previous robot swarm systems.

To adapt to the intricate spatial characteristics of real farmland—such as irregularly shaped fields divided by crisscrossing roads, and scattered obstacles including wind turbines, utility poles, and fixed sensors—the team introduced a two-level partitioning system to simplify the complex working environment. The first-level partitions are defined as individual farmland plots naturally divided by field roads, a design that aligns with the actual layout of agricultural production and eliminates the unrealistic assumption of a single, regular working area in traditional path planning models. For each first-level partition, the research employs grid-based environmental modeling, setting the common working range of the robot swarm as the unit length of grid cells to standardize the spatial representation of farmland. Irregular obstacles are processed using dilation operations in binary morphology, a technique that ensures the robot swarm avoids collision risks by expanding obstacle boundaries to account for physical movement constraints, a critical step for translating virtual path planning into real-world operational safety.

Building on the grid modeling and obstacle dilation, the team establishes second-level partitioning within each first-level partition, along with clear grid partitioning and merging rules. The partitioning process identifies obstacle boundaries with identical horizontal or vertical coordinates to draw partition lines, splits irregular obstacles into rectangles using virtual dividing lines to minimize partition quantity, and merges adjacent partitions with equal common side lengths—prioritizing vertical merging then horizontal merging—to reduce the total number of partitions. After merging, the original shapes of obstacles are restored, and irregular blank areas are merged into adjacent partitions or set as independent ones, ensuring the two-level partitioning system fully preserves the actual spatial features of farmland while simplifying the computational complexity of path planning. This partitioning method bridges the gap between abstract algorithmic design and the real geographic characteristics of farmland, making the robot swarm’s path planning and area coverage far more adaptable to on-site agricultural production conditions.

A major innovation of the study is the development of an improved genetic algorithm (GA) to optimize the traversal order of the two-level partitions, addressing the limitations of traditional GA and simulated annealing algorithms in convergence speed and global optimization capability. The traditional GA suffers from randomness in crossover operations, leading to poor convergence, while simulated annealing algorithms are prone to falling into local optima, resulting in longer traversal paths and higher computational costs. To overcome these issues, the research integrates a hybrid particle swarm optimization (PSO) algorithm into the GA’s crossover operation, where each chromosome crosses with the chromosome of the real-time optimal fitness value during iteration—rather than random pairwise crossover—balancing population diversity and convergence speed. This modified crossover operation ensures that the algorithm retains high-quality genetic structures while continuously exploring better solutions, a key enhancement that accelerates the search for the optimal partition traversal order.

Further, the team introduces the concept of chromosome population diversity and designs adaptive crossover and mutation probabilities that adjust based on both the fitness value differences of individual chromosomes and the evolutionary stage of the entire population. For chromosomes with high fitness values—representing near-optimal traversal paths—low crossover and mutation probabilities are set to protect their superior genetic structures; for low-fitness chromosomes, high probabilities are applied to eliminate inferior solutions and generate new ones. Population diversity is quantified using the ratio of the difference between the population’s maximum and average fitness values to the maximum fitness value, with the maximum crossover and mutation probabilities dynamically adjusted according to three diversity stages. This adaptive mechanism prevents the algorithm from falling into local optima during late-stage evolution and ensures sufficient exploration in the early stage, significantly improving the GA’s ability to find the global optimal solution for partition traversal.

The improved GA is paired with a depth-first search (DFS) algorithm to achieve full-area coverage of the farmland by the robot swarm, with a clear division of labor between the two algorithms: the improved GA determines the optimal traversal order of first-level and second-level partitions, while the DFS algorithm handles path search within individual second-level partitions, workload-based area allocation, and specific traversal path planning for each robot. The research establishes customized DFS traversal rules for second-level partitions, including two traversal order matrices that adapt to the starting position (left or right end of the partition) to ensure the continuity of the robot’s working area and reduce the number of turns— a critical design to minimize energy consumption and operational time in real-world operation. The DFS algorithm calculates the real-time traversed area using the Monte Carlo method during path search, allocating working areas to each robot according to their pre-determined workloads and recording paths once the workload requirement is met, then proceeding to the next area.

To further optimize the robot’s traversal path within its assigned working area, the research classifies working areas into vertical and horizontal types based on their aspect ratio, with tailored traversal strategies for each type. For vertical areas (greater vertical span), the robot starts from the bottom-left cell, prioritizes traversal toward the shorter end of each column to minimize backtracking, and searches for un-traversed areas in a left-vertical-right order; for horizontal areas (greater horizontal span), the robot starts from the bottom-left cell and follows a bottom-horizontal-up search order. When the robot encounters a dead end with no adjacent un-traversed cells, the algorithm evaluates new starting points in un-traversed areas based on the principle of minimizing the total traversal area repetition rate of the robot swarm, and plans the path to the new starting point using the A* algorithm combined with an eight-neighborhood search method. This integrated path planning approach ensures that each robot’s traversal path is both continuous and efficient, minimizing redundant movement while guaranteeing full coverage of the assigned working area.

Extensive simulation experiments were conducted using MATLAB 2014 to validate the performance of the improved GA and the full-area coverage strategy, with comparisons to traditional GA and simulated annealing algorithms across different partition scales (10, 20, and 30 partitions). The results demonstrated the superior performance of the improved GA: for 30 partitions, the path length for traversing all partition centroids was reduced by 2.8% compared to traditional GA and 9.3% compared to simulated annealing. More notably, the number of convergence iterations was cut by 69.5% and 19.0% respectively, and the computational time was reduced by 64.2% and 9.9%. The improved GA exhibited a faster convergence speed in the early iteration stage, with no obvious plateau periods— a common issue in simulated annealing that indicates local optima—proving its strong ability to find the global optimal solution for partition traversal.

Additional simulation experiments were designed to test the full-area coverage strategy under varying farmland conditions, with five environmental maps featuring 3, 5, 7, 9, and 11 obstacles (with equal total obstacle area but increasing dispersion) and four irregularly shaped farmlands with equal total working area and identical obstacle settings. The results showed that the total traversal area repetition rate of the robot swarm increased with the number and dispersion of obstacles, reaching 6.3%, 8.9%, 16.7%, 21.7%, and 23.4% for 3 to 11 obstacles respectively— a trend that aligns with real-world farmland complexity, where more scattered obstacles inevitably increase traversal redundancy. For the four irregular farmlands, the total repetition rates were 16.7%, 13.1%, 11.9%, and 6.7%, demonstrating the strategy’s strong adaptability to diverse farmland shapes, including single first-level partitions and multiple partitions divided by vertical or T-shaped roads. In all simulation scenarios, the robot swarm achieved 100% coverage of the farmland working area, confirming the strategy’s reliability in ensuring complete coverage while minimizing traversal redundancy.

Field experiments were conducted on the lawn south of the Library of Shandong University of Technology to validate the practical applicability of the task allocation and full-area coverage strategy, using four information collection agricultural robots selected from a pool of 10 based on the team’s task allocation mechanism. The experimental working area was divided into three sub-regions by virtual roads with eight irregular obstacles, with a total workload requirement of 1 hm² and a time limit of 25 minutes— a realistic operational constraint for agricultural production. The weight factors for the four performance parameters (service quality, historical workload, energy consumption, failure rate) were set to 0.25 each, a neutral setting that reflects the actual decision-making process of farm managers who balance multiple robot performance metrics. The selected robots were assigned proportional workloads based on their operational efficiency, and equipped with RMONCAM HD600 infrared cameras with a 1.5 m ground working range for farm condition surveying, simulating real agricultural data collection tasks.

The field experiment results confirmed the effectiveness and practicality of the proposed strategy: all four robots completed their assigned workloads within the 25-minute time limit, with actual working times ranging from 21.95 to 23.11 minutes, and achieved 100% coverage of their target working areas. The individual traversal area repetition rates of the robots were 5.77%, 4.14%, 6.75%, and 4.85%— far lower than the redundancy rates observed in traditional multi-robot coverage strategies—proving that the strategy effectively minimizes unnecessary movement in real operational conditions. The consistent performance of the robot swarm in both simulation and field experiments validates that the task allocation mechanism and path planning strategy are not only theoretically sound but also translatable to real agricultural production, addressing the gap between laboratory algorithm design and on-site application.

The implications of this research extend far beyond the technical optimization of agricultural robot swarms, as it provides a scalable framework for the development of unmanned farm systems in complex agricultural environments worldwide. By accounting for robot heterogeneity, the strategy makes unmanned farm technology more accessible to real agricultural operations, where robot fleets are rarely composed of identical units due to incremental equipment updates and wear over time. The two-level partitioning system, rooted in the actual layout of farmland, ensures that the algorithmic design is grounded in real agricultural geography, avoiding the over-simplification that has limited the practical use of many previous multi-robot systems. The improved GA, with its adaptive crossover and mutation probabilities and hybrid PSO crossover operation, sets a new standard for optimization algorithms in agricultural robot path planning, offering a faster, more efficient solution for traversal order optimization in complex spatial environments.

In the context of global agricultural transformation, where precision agriculture and unmanned farming are becoming increasingly critical to food security and sustainable agricultural development, this research represents a significant step forward in the industrial application of agricultural robot swarms. The strategy’s ability to balance full area coverage, minimal traversal redundancy, and efficient workload distribution based on robot heterogeneity addresses the core technical challenges of scaling up unmanned farm systems, making it possible to deploy robot swarms in large, complex farmland areas with scattered obstacles and irregular field shapes— the norm in real agricultural production rather than the exception. Future research can build on this framework by integrating real-time environmental sensing and dynamic task reallocation, adapting the strategy to dynamic farmland conditions such as temporary obstacles or sudden changes in workload requirements, and further enhancing the autonomy and flexibility of agricultural robot swarms.

This study also highlights the importance of interdisciplinary collaboration in agricultural automation, combining expertise in mechanical engineering, agricultural engineering, artificial intelligence, and computer science to develop solutions that are both technically innovative and practically relevant. The research team’s partnership with Zibo Hefeng Seed Company to base the farmland environment model on a real digital ecological circular farm ensures that the strategy is designed for real agricultural production needs, a key principle that should guide the development of future agricultural automation technologies. By centering on real-world farmland characteristics and robot performance, the research avoids the ivory-tower design that has plagued some agricultural automation technologies and delivers a solution that is ready for pilot testing and large-scale application in unmanned farms.

In summary, the cooperative operation strategy for agricultural robot swarms proposed by the Shandong University of Technology research team is a transformative advancement in agricultural automation. By addressing robot heterogeneity, complex farmland spatial characteristics, and the limitations of traditional optimization and path planning algorithms, the strategy elevates the overall efficiency and practicality of agricultural robot swarms, providing a vital theoretical and technical foundation for the development of unmanned farm systems in complex environments. As rural labor shortages continue to intensify globally and the demand for precision agriculture grows, this research paves the way for the large-scale deployment of agricultural robot swarms, marking a significant step toward the fully automated, intelligent future of agriculture.

Author Affiliations, Journal and DOI: This research was conducted by Gong Jinliang, Wang Wei from the School of Mechanical Engineering, Shandong University of Technology, Zibo 255049, China, and Zhang Yanfei, Lan Yubin from the School of Agricultural Engineering and Food Science, Shandong University of Technology, Zibo 255049, China. It was published in Transactions of the Chinese Society of Agricultural Engineering in January 2021 (Vol.37, No.2), with the DOI: 10.11975/j.issn.1002-6819.2021.2.002.

(Word count: 3896)