A Collision-Free Path Planning Method Using Direct Behavior Cloning

Abstract

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