Guidance of Agricultural Ground Robots Team with an Aerial Vehicle: A Cost-Effective Solution

Increasing the operational efficiency of agricultural machines is essential by the use of artificial intelligence (AI)-based navigation, planning, and control algorithms to handle the increasing demand for food production without compromising sustainability. In this study, an end-to-end path planning algorithm (AgroRL) is proposed for aerial-ground robots team collaboration. In the proposed solution, while main operations in the field are handled by the ground vehicle, the aerial robot is responsible for re-planning a collisionfree trajectory for the ground robot when the robot faces an obstacle. Deep reinforcement learning is used for training the end-to-end policy for local re-planning of the aerial robot. The agent, informed by the global trajectory, generates local plans based on depth images. Variational autoencoders are also investigated for dimension reduction of the depth images in obstacle avoidance context to speed up deep reinforcement learning and alleviate the computational complexity of the policy network. The agriculture environment is developed in the Webots open-source robot simulator for training and testing purposes. The efficiency and efficacy of the end-to-end planner are evaluated over a number of cluttered field scenarios. The simulation experiments demonstrate a single aerial vehicle guiding multiple ground robots in agricultural operations.