Goal and Force Switching Policy for DMP-Based Manipulation

Abstract

Effective solving of manipulation tasks may be executed using classical or learning-based control methods. Classical methods are accurate; however, they require complicated tuning. Learning-based control provide obtaining the parameters of the process model while training, but this model is rough. We ap- ply a combined approach to solving manipulation tasks, where the robot moves to the target vicinity under learning-based control and then op-erates the target under simplified classical control. On the first stage, control system generates reference trajectory for execution using dynamic move-ment primitives (DMP). The parameters of the DMP are determined by out-put of a neural network and trained via policy optimization. On the second stage the forcing term of the DMP is set to zero, while goal is defined by the simplified predictive control model. We evaluate our approach on the tasks of reaching target point by end effector and pushing elevator button with UR5 collaborative manipulator.