Deep Reinforcement Learning for Collision Avoidance of Robotic Manipulators

Abstract

In this paper a real-time collision avoidance approach using machine learning is presented for safe humanrobot coexistence. More specifically, the collision avoidance problem is tackled with Deep Reinforcement Learning (DRL) techniques, applied to robot manipulators with a workspace invaded by unpredictable obstacles. Since the robotic systems are defined in the continuous space, a Normalized Advantage Function (NAF) model-free algorithm has been used. In order to assess the proposal, a robotic system, that is a COMAUSMART3-S2 anthropomorphic robot manipulator, has been considered. The robotic system has been interfaced with external tools for evaluation, control, and automatic training. Simulations carried out on a virtual environment are finally reported to show the effectiveness of the proposed model-free deep reinforcement learning algorithm.