Hybrid Policy Learning for Dexterous Manipulation

Anonymous

Keywords: hierarchical controller, residual policy learning, dexterous manipulation

TL;DR: We propose a hierarchical controller where a state-machine sequences manipulation primitives, and a residual policy provides torque corrections to impedance controllers.

Abstract: In Phase 2, we solved the tasks of manipulating the cube to random goal positions with the real robot. Each task is decomposed into three stages: moving the fingers to the cube to grasp it, reposing the cube with three fingers, and flipping the cube with two fingers. These grasping, lifting, and flipping primitives are sequenced with a state-machine. As a baseline, we designed model-based controllers to execute each of these primitives. To improve performance of the controllers on the real robot, we trained a residual policy, which adds joint-torque corrections to torques output by the controller to mitigate discrepancies between the simulated and real robot.