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

The article introduces Neural External Torque Estimation (NEXT), a method that estimates external joint torques without dedicated sensors, requiring only 1 minute of training from 10 minutes of free motion data. Coupled with Force-Informed Re-Sampling Training (FIRST), which enhances policy learning by focusing on critical contact phases, this approach improves task performance by over 17% compared to existing force-aware methods across five long-horizon tasks. This advancement enables effective force-feedback teleoperation and policy learning on commodity robot arms, making sophisticated manipulation accessible without additional hardware costs.