Residual Rotation Correction using Tactile Equivariance
Yizhe Zhu, Zhang Ye, Boce Hu, Haibo Zhao, Yu Qi, Dian Wang, Robert Platt · 2025
Read Residual Rotation Correction using Tactile Equivariance by Yizhe Zhu, Zhang Ye, Boce Hu, Haibo Zhao, Yu Qi, Dian Wang, Robert Platt (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-07381.txt.
Description
Yizhe Zhu and six co-authors exploit a symmetry of touch to teach robots contact-rich manipulation from very little data. Visuotactile policies add tactile sensing to vision, but collecting tactile demonstrations is expensive, so sample efficiency rules. Their system, EquiTac, first reconstructs surface normals from the raw images of vision-based tactile sensors, arranging matters so that rotating the normal field corresponds exactly to rotating the object in the hand.
An equivariant network built on that rotational symmetry then predicts a residual rotation action that corrects a base policy at test time, with no extra reorientation demonstrations needed. On a real robot, EquiTac generalized zero-shot to unseen in-hand orientations from very few training samples, where baselines failed even with more data. The authors call it the first tactile method to encode tactile equivariance explicitly.
Catalog record
| CREATOR | Yizhe Zhu, Zhang Ye, Boce Hu, Haibo Zhao, Yu Qi, Dian Wang, Robert Platt |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 31 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-07381