IROS 2026 EquiBim - Heungwoo/research GitHub Wiki
IROS 2026 — EquiBim: Symmetry-Equivariant Policy for Bimanual Manipulation
Venue: IROS 2026 (Pittsburgh) · paper #3507 · Purdue University · University of Arkansas (Zhang, Mohan, Han, Shou, Wang, She). The inductive-bias datapoint for bimanual manipulation — enforce the bilateral symmetry inherent to dual-arm robots as an equivariance constraint, so symmetric observations yield symmetric actions. Companions: IROS 2026 survey · 3D FlowMatch Actor · VLA Architectures.
1. Problem
Robot imitation learning rarely accounts for the physical symmetries of the robot, producing asymmetric/inconsistent behaviors under symmetric observations — especially acute in dual-arm manipulation, where bilateral symmetry is inherent to both the morphology and many tasks.
2. Method
EquiBim enforces bilateral equivariance between observations and actions during training:
- Formulates physical symmetry as a group action on both observation and action spaces.
- Imposes an equivariance constraint on policy predictions under symmetric transforms.
- Model-agnostic — integrates into diverse IL pipelines: point-cloud and image observations; end-effector-space and joint-space actions.
3. Results
- Evaluated on RoboTwin (dual-arm, symmetric kinematics) across diverse observation/action configs; validated on a real dual-arm system.
- Consistently improves performance and robustness under distribution shift — "a simple yet effective inductive bias for bimanual robot learning."
4. Why it matters (bimanual lens)
EquiBim complements 3D FlowMatch Actor in the survey §5.3 "coordination-structure" thesis: rather than a new architecture, it adds a symmetry prior that any bimanual policy can inherit — the data-efficient route to consistent two-hand behavior. Where 3DFA unifies single/dual-arm via geometry + flow matching, EquiBim bakes in the bilateral structure directly. Both argue the bimanual win is structure, not scale.
Limitations (reviewer): equivariance assumes the task is symmetric — asymmetric bimanual tasks (lead-arm/follow-arm, handovers) may not benefit or could be over-constrained; RoboTwin + one real system is the eval scope.
5. Links
- Official program: IROS 2026 (paper #3507) · survey: IROS 2026
- Related: 3D FlowMatch Actor · VLA Architectures · Cross-Embodiment
← Back to IROS 2026 survey · Home