Learning arm movements of target reaching for humanoid robot

The autonomous motor skill learning is crucial for the humanoid robot to adapt to various tasks in complex environments and develop human-like behaviors. In this paper a method of autonomously learning arm movements for target reaching is proposed. To model the dynamics of arm joint trajectories during reaching, the dynamical movement primitives (DMP) model is employed. Based on the DMP representation, reinforcement learning based methods are adopted to learn DMP shape and goal parameters, aiming at not only deriving suitable arm joint trajectories satisfying certain constraints such as energy-saving and collision-free, but also enabling the robot to find goal configurations without complicated inverse kinematics calculations. Furthermore, an adaptive exploration strategy is proposed, which accelerates and improves the arm movements learning. The robot experiments demonstrate the effectiveness of the proposed method.