Reinforcement learning of dual-arm cooperation for a mobile manipulator with sequences of dynamical movement primitives

This paper presents a state-of-art reinforcement learning strategy to enable a human-like dual arm mobile robot to deal with some complicated tasks with dual arm cooperation. A complex movement task of robot can be divided into action phases and subgoals, each of which corresponds to a dynamic motion primitive (DMP). During control of the humanoid-like dual arm mobile robot, there are some noises that affect the precision of the movement of the robot. To deal with those uncertain perturbations while handling varying manipulation dynamics for grasping motion, a reinforcement learning (RL) algorithm with sequences of dynamical motion primitives strategy is proposed in this paper. To verify the effectiveness of the proposed strategy, a two-phase planning has been considered in the experiment, namely, the online redundancy resolution based on the neural-dynamic optimization algorithm to obtain the initial joint trajectories on the first trial, and the reinforcement learning of DMP in the learning process, where DMP is used to model the joint trajectories, and then reinforcement learning is employed to adjust the model to suppress uncertain perturbations.