Teaser
EquiBim enforces symmetry-equivariant behavior by ensuring that actions predicted from an observation and its mirrored counterpart transform consistently under reflection.
EquiBim enforces symmetry-equivariant behavior by ensuring that actions predicted from an observation and its mirrored counterpart transform consistently under reflection.
EquiBim is a symmetry-equivariant policy learning framework for bimanual manipulation that enforces consistent transformations between observations and actions. It is model-agnostic and integrates seamlessly into diverse imitation learning pipelines. Evaluations on both simulated and real-world dual-arm systems show improved performance and robustness under distribution shifts.
We visualize representative rollouts across three bimanual manipulation tasks under both training and shifted distributions, followed by closer-look phone recordings from the real system.
@article{zhang2026equibim,
title={EquiBim: Learning Symmetry-Equivariant Policy for Bimanual Manipulation},
author={Zhang, Zhiyuan and Mohan, Aditya and Han, Seungho and Shou, Wan and Wang, Dongyi and She, Yu},
journal={arXiv preprint arXiv:2603.08541},
year={2026}
}