FACTR 2: Learning External Force Sensing for Commodity Robot Arms Improves Policy Learning

TL;DR

We propose NEXT (Neural External Torque Estimation) for estimating external forces on commodity robot arms without force sensors, and FIRST (Force-Informed Re-Sampling Training) to leverage these estimates for improved policy learning.

Problem

Contact-rich manipulation requires force sensitivity, but dedicated force sensors are expensive and rarely installed on most robot arms. A method is needed to accurately estimate external forces without sensors and use them to enhance policy learning.

Approach

NEXT is a data-driven external joint torque estimator trained in 1 minute from only 10 minutes of free-motion data. Using the learned estimator, we propose FIRST, which up-samples pre-contact and contact segments during behavior cloning based on force information.

Results & Contribution

Across five long-horizon tasks, FIRST outperforms prior force-aware policies by over 17% in task progress. NEXT and FIRST enable force-aware teleoperation and policy learning on off-the-shelf robots without additional sensing hardware.