関節物体操作のための計画アンサンブル上のメタ方策学習

原文アブストラクト

Recent work has shown that complex manipulation skills, such as pushing or pouring, can be learned through state-of-the-art learning based techniques, such as Reinforcement Learning (RL). However, these methods often have high sample-complexity, are susceptible to domain changes, and produce unsafe motions that a robot should not perform. On the other hand, purely geometric model-based planning can produce complex behaviors that satisfy all the geometric constraints of the robot but might not be dynamically feasible for a given environment. In this work, we leverage a geometric model-based planner to build a mixture of path-policies on which a task-specific meta-policy can be learned to complete the task. In our results, we demonstrate that a successful meta-policy can be learned to push a door, while requiring little data and being robust to model uncertainty of the environment. We tested our method on a 7-DOF Franka-Emika Robot pushing a cabinet door in simulation.