Generalization of Reinforcement Learning through Artificial Potential Fields for agricultural UGVs

摘要

The increasing global demand for food, coupled with factors such as the shrinking agricultural workforce and the need for environmentally friendly practices has led to the emergence of Agriculture 4.0. In this context, Unmanned Ground Vehicles (UGVs) have become integral to smart farming, offering efficient and environmentally sustainable solutions compared to traditional machinery. One of the most used path planning algorithms for ground robots is the Artificial Potential Field (APF), whose effectiveness can be compromised by the insurgence of minimum potential points. Hence, a method combining Deep Reinforcement Learning with APF is here proposed. The algorithm is tested on the numerical model of a tracked UGV. Numerous simulations demonstrate its capability to guide the ground robot model through complex environments while ensuring collision avoidance. Specifically, the presented algorithm is able to overcome the generalization limit typical of Reinforcement Learning (RL) and the local minima problem characteristic of APF methods.