Contact-Rich Object Insertion: Lessons for Zero-Shot Policy Transfer

Samarth Brahmbhatt, Ankur Deka, Andrew Everett Spielberg, Matthias Müller

Keywords: sim-to-real, manipulation, object insertion

TL;DR: We quantitatively examine the impact of various training design choices for transferring contact-rich object insertion policies trained in simulation, directly to a real robot.

Abstract: Robotics simulators have opened up the possibility for contact-rich manipulation policies to be trained entirely or mostly in simulation. Training in simulation can be safer, cheaper, and faster than training with real robots. However, policies for contact-intensive tasks often suffer a large performance drop when transferred to a real robot. In this work, we examine this problem through the task of inserting a plate into a narrow slot. We train a policy with reinforcement learning, propose various steps to make the simulation training more realistic, and report their impact on real robot performance. Our policy not only outperforms baselines and transfers with a negligible sim-to-real performance drop, it also generalizes with a minor modification to inserting a cup and plates of different sizes and weights. Demo videos are available at https://youtube.com/playlist?list=PLdMOXIlbRGoVL0XezrRgk4-LsqXqtICmi.