Control of Open Mobile Robotic Platform Using Deep Reinforcement Learning

Abstract

AbstractAdvanced control for mobile robotic platforms allows efficient real-time navigation in structured and unstructured environments in various industry applications. Deep reinforcement learning is an emerging control strategy where an agent is trained iteratively according to an optimisation objective by using reward and penalty actions. The agent generates the neural network weights used for computing the robot command towards the reference set point. We present an application for an open hardware mobile robotic platform navigation that integrates the sensing, communication, computing and control functions into a single system for navigation in unstructured environments. Implementation is performed through a dedicated software and communication layer that integrates the hardware platform with the MATLAB environment using standardized Robot Operating System (ROS) libraries. Quantitative testing results are presented, in order to prove the viability of the solution, by defining both simulation and laboratory setting scenarios.KeywordsOpen mobile robot platformDeep reinforcement learningROS