Scenario-based collision avoidance control with deep Q-networks for industrial robot manipulators

This work proposes a scenario-based Deep Reinforcement Learning (DRL) approach enabling robot manipulators to efficiently execute industrial tasks while avoiding the collision with obstacles. The proposal exploits a DRL-based decision maker trained ad hoc so as to be able to automatically select at any time instant the most appropriate control methodology, in a given set, to execute the required industrial task. The capability of performing the selection automatically is "learnt" by training the system relying on a suitably designed reward function. It takes into account the robot relative distances from the target and the obstacles, the computational cost associated with each methodology, as well as the percentage of task completion obtained by applying the selected methodology. The learning skill is enforced by a properly sized Deep Q-Network (DQN). The proposal is assessed relying on realistic robotic manipulator scenarios reproduced in the CoppeliaSim environment.