Double Deep Q-Network for Trajectory Generation of a Commercial 7DOF Redundant Manipulator

Recent studies to solve industrial automation applications where a new and unfamiliar environment is presented have seen a shift to solutions using Reinforcement Learning policies. Building upon the recent success of Deep Q-Networks (DQNs), we present a comparison between DQNs and Double Deep Q-Networks (DDQNs) for the training of a commercial seven joint redundant manipulator in a real-time trajectory generation task. Experimental results demonstrate that the DDQN approach is more stable then DQN. Moreover, we show that these policies can be directly applied to the official visualizer provided by the robot manufacturer and to the real robot without any further training.