Multi-Modal Learning
FlexiTac

LeFlexiTac: Giving Robots a Sense of Touch

Naian Tao, Yifan He*, Wesley Maa*, Binghao Huang, Yunzhu Li

Columbia University

Columbia University RoboPIL Blog (May 2026)

Abstract

SUMMARY (compiled from project sources)

LeRobot has made low-cost robot learning widely accessible, but its policies are still blind to contact, relying entirely on vision. LeFlexiTac adds FlexiTac tactile sensing to the SO10X platform and shows that even a modest amount of touch meaningfully lifts success rates on contact-rich manipulation. The authors evaluate this across four policy families: ACT, Diffusion Policy, Pi0.5, and SmolVLA. The integration is designed as a drop-in tactile path that requires minimal changes to the existing LeRobot pipeline, exposing tactile data through a unified observation.tactile.* data stream. The work targets tasks where vision is limited or where touch is essential and the policy would otherwise be unable to perceive key aspects of the interaction.

Links & Resources