Multi-agent Collaborative Framework for Automated Agriculture

Abstract

The agriculture sector faces numerous challenges in meeting the demands of a growing global population while managing resource constraints, labour shortages and environmental sustainability. This thesis presents a multi-agent collaborative framework for agriculture, integrating Unmanned Aerial Vehicles (UAVs) and Unmanned Ground Vehicles (UGVs) to enhance precision farming practices. The framework aims to automate critical tasks such as surveying farms, monitoring crop health, and coordinating harvesting and pest control activities. An end-to-end pipeline is developed, integrating essential modules from mapping to task identification to deployment, ensuring seamless agent interaction and task execution. The framework's core lies in a novel graph-based decision-maker, which partitions tasks near-optimally while deciding ten times faster than standard solvers. By modelling agents and tasks in a computationally inexpensive manner, the system minimises overhead for decision-making, emphasising onboard and required resources.

Additionally, this research discusses techniques for proximal coverage of non-convex objects, such as trees, enabling effective monitoring and interaction with irregularly shaped agricultural entities. The concept of switching points is introduced to optimise path planning when covering multiple trees in close proximity, leveraging the geometry of the orchard to reduce overall path length significantly, thus reducing energy and resource utilisation. Once waypoints for covering multiple trees are generated, a graph-based partitioning and routing scheme allocates global routes to multiple agents. Simultaneously, reachability analysis ensures that vehicles navigate effectively in cluttered agricultural environments.

To demonstrate the application of the framework and its developed modules, the yield estimation and automated spraying of a mango plantation are presented. This is achieved through a novel method named LocalizeSort, which advances state-of-the-art multiple-object tracking for fruits, resulting in more accurate yield estimations and fruit localisation. The framework is rigorously tested through extensive simulations using ROS and Gazebo in a virtual environment before field deployment. Finally, a comprehensive analysis of the system's architecture, its implementation, and the evaluation of its performance in real-world scenarios is conducted using specialised hardware through rigorous field experiments. The findings demonstrate the potential of such collaborative multi-agent systems in revolutionising modern agriculture, offering scalable, long-term, and adaptable solutions for future farming challenges.