A Collision-Free Path Planning Method Using Direct Behavior Cloning

摘要: An effective path planning approach based on deep learning for robotic arms is presented in this paper. Direct behavior cloning applied to extract the obstacle avoidance policy collision-free paths generated by reliable motion planners, such as RRT* algorithm our case. Behavior simplest form of imitation learning, also known Learning from Demonstration (LfD), where an agent tries learn a recover expert’s action with respect state environment. The designed paper gives scene knowing pose obstacle, initial and goal configurations. taken each time changes thus method able achieve online regardless whether environment static or dynamic. We build simulation V-REP Python client program collect state-action dataset validate trained policies. Policy models without visual input are constructed tested same experiment setting determine best solution. Results show that model accurate handles issue well.