Deep reinforcement learning based proactive dynamic obstacle avoidance for safe human-robot collaboration

Abstract

Ensuring the health and safety of human operators is paramount in manufacturing, particularly in human-robot collaborative environments. In this paper, we present a deep reinforcement learning-based trajectory planning method for a robotic manipulator designed to avoid collisions with human body parts in real-time while achieving its goal. We modelled the human arm as a freely moving cylinder in 3D space and formulated the dynamic obstacle avoidance problem as a Markov decision process. The algorithm was tested in a simulated environment that closely mimics our laboratory environment, with the goal of training a deep reinforcement learning model for autonomous task completion. A composite reward function was developed to balance the effects of different environmental variables, and the soft-actor critic algorithm was employed. The trained model demonstrated a 93% success rate in avoiding dynamic obstacles while achieving its goals when tested on a generated data set.