Proximity-Based Reward System and Reinforcement Learning for Path Planning

Abstract

Path planning is an important and complex task in the field of robotics and automation. It consists of finding the optimal path given a starting location, obstacles and a final destination. Reinforcement learning is a trial and error approach that has seen success in the field of path planning. Multiple reinforcement learning algorithms such as Q-learning and SARSA exist and have achieved great results. These algorithms typically use a uniform reward system such that every move, collision and goal return a specific reward. We propose a proximity-based reward system for classical reinforcement learning algorithms on path planning scenarios. We compare our reward systems combined with different optimization techniques and algorithms for path planning. These approaches are compared using the total completion rate for the mazes and average training time. We achieved interesting results with our reward systems and optimization techniques allowing us to decrease the training time.