Learning to Avoid Collisions in Mobile Manipulator Reaching Task

Planning a trajectory for a mobile manipulator to complete a task challenges extremely as the robot has high degrees of freedom (DoF), especially in an unstructured environment. In recent years deep reinforcement learning (RL) achieves excellent performance in control fields, this paper applies the state-of-the-art deep RL algorithms to solve whole-body motion planning and control problem for mobile manipulator reaching in different environments. The method generalizes the problem from collision-free environment to those with static or dynamic obstacles by using transfer learning. The experiments are evaluated in different environments and the results verify the feasibility.