Development and evaluation of Reinforcement Learning models for the FOSSBot Open-Source educational robot.

Machine learning algorithms play a crucial role in addressing real-world tasks by harnessing data-driven insights to optimize and enhance decision-making processes. Their application in robotics usually relies on simulations as an essential testing ground, allowing us to refine solutions, minimize risks, and fine-tune algorithms before deploying them in practical, physical environments. In this work, we evaluate the performance of two popular reinforcement learning algorithms, namely Proximal Policy Optimization (PPO) and Deep Q Network (DQN): i) on an obstacle detection and avoidance task, and ii) on a navigation task. For this purpose, we rely on the FOSSBot open-source educational robot, which contains ultrasonic sensors, infrared obstacle sensors, an Inertial Measurement Unit, and odometers. More specifically, we work on a simulation of FOSSBot in the CoppeliaSim environment. The evaluation results on the grid environment are very promising for our algorithms since the success rates (i.e. reaching the destination without collision) for both algorithms is above 93%. As for the simulated environment, the results show the superiority of PPO over DQN, with PPO scoring a success rate of 74%, making it a promising navigation algorithm candidate for the real-world FOSSBot.