FACTR 2: Learning External Force Sensing for Commodity Robot Arms Without Dedicated Sensors

Forum topic · 小凯 · 2026-06-12

Summary

FACTR 2 introduces Neural External Torque Estimation (NEXT), a data-driven method that estimates external joint torques on commodity robot arms without any dedicated force sensors. NEXT trains in just 1 minute using only 10 minutes of free-motion data and achieves estimation accuracy comparable to dedicated joint-torque sensors. Building on NEXT, the authors propose Force-Informed Re-Sampling Training (FIRST), which up-samples pre-contact and contact segments during behavior cloning to improve policy learning. Across five long-horizon manipulation tasks, FIRST outperforms prior force-aware policies by over 17% in task progress. NEXT also enables force-feedback teleoperation on low-cost arms. Together, NEXT and FIRST bring force-aware teleoperation and imitation learning to off-the-shelf robots with no additional sensing hardware. Paper: arXiv 2606.12406 by researchers including Steven Oh, Kenneth Shaw, Ruslan Salakhutdinov, and Deepak Pathak.

Paper Overview

Field: Machine Learning Authors: Steven Oh, Jason Jingzhou Liu, Tony Tao, Philip Han, Kenneth Shaw, Satoshi Funabashi, Ruslan Salakhutdinov, Deepak Pathak Published: 2026-06-10 arXiv: 2606.12406

Abstract

Contact-rich manipulation requires force sensitivity, but many robot arms lack dedicated force sensors due to their high cost. We present Neural External Torque Estimation (NEXT), a data-driven method that estimates external joint torques without needing any dedicated force sensors. NEXT trains in 1 minute from only 10 minutes of free-motion data, yet achieves estimates comparable to dedicated joint-torque sensors. NEXT enables force-feedback teleoperation on low-cost arms and improves policy learning through Force-Informed Re-Sampling Training (FIRST), which up-samples pre-contact and contact segments during behavior cloning. Across five long-horizon tasks, FIRST outperforms prior force-aware policies by over 17% in task progress. Together, NEXT and FIRST bring force-aware teleoperation and policy learning to off-the-shelf robots, with no extra sensing hardware.

Key Takeaways

NEXT (Neural External Torque Estimation): learns to estimate external joint torques from proprioceptive signals alone, without dedicated force/torque sensors.
Efficient training: requires only 10 minutes of free-motion data and trains in about 1 minute.
Sensor-grade accuracy: estimation quality is comparable to dedicated joint-torque sensors.
Force-feedback teleoperation: NEXT enables low-cost robot arms to provide force feedback to human teleoperators.
FIRST (Force-Informed Re-Sampling Training): improves behavior cloning by up-sampling pre-contact and contact segments of demonstrations.
Results: across five long-horizon tasks, FIRST achieves over 17% higher task progress than previous force-aware policies.