Implementing Reinforcement Learning on a Smartphone: A Study with OpenBot

Reinforcement learning (RL), like other subfields of machine learning, has gotten significant attention from the research community over the last decade. Most recent RL works involve training in simulation, though. In this study, we investigate using RL for an autonomous navigation task for a mobile robot. The robotic platform we employ consists of an OpenBot chassis, and a smartphone which acts as the processing and perception unit of the robot. Our study investigates if it is feasible to implement RL entirely on a smartphone, and perform the training directly using the robotic platform. The navigation task at hand is to drive across a track for 15 sec. Our study shows that it is feasible to implement RL on a smartphone, with the training conducted using the actual robotic platform. The results show that, in general, after approximately 400 episodes of training, the agent was able to successfully navigate for the aimed 15 sec in half of the attempts.