An experience-based policy gradient method for smooth manipulation

Policy gradient methods have achieved remarkable success in continuous controlling tasks. However, in robotic control, original policy gradient algorithms depend on the first succeed experience which is usually a suboptimal solution. To improve the performance, we propose an experience-based policy gradient method(EBDDPG) which guides the robot to move in a smooth way. Besides, extra OU-noise is added to the action space to improve exploration. We tested our algorithm on Gazebo simulation environment with Baxter robot. The experimental results show our method guides the robot to manipulate more smoothly and improves success rate of grasping tasks.