Bimanual role-adaptive decomposition

BiRoADLearning Shared and Role-Adaptive
Representations for Bimanual Manipulation

Yan Shen1,4,*Yuchen Liu1,*Feng Jiang1Hangtian Hu1Xiaoqi Li1,4Shu Chen3Ruihai Wu2Hao Dong1,4,†

1 CFCS, School of Computer Science, Peking University2 UC Berkeley3 HKUST (Guangzhou)4 PrimeBot Research Institute

* Equal contribution   ·   † Corresponding author

Read the paper Watch videoReal-world demos
Two robot arms opening a bottle in the first scene configurationScene configuration A
The paired bottle-opening demonstration with a different arm-role assignmentScene configuration B
Same task. Different roles.Sharing coordination across arms while adapting their roles to the scene.Watch the pairs

01 / Overview

Two arms. One coordinated policy.

A modular representation framework for scene-conditioned role adaptation.

BiRoAD · Overview and experimental demonstrations · 3 min 51 sec

Abstract

Bimanual manipulation requires policies that coordinate two arms while adapting their functional roles to scene geometry, object configuration, and task context. Uneven role distributions in demonstrations can limit generalization to underrepresented arm–role configurations.

BiRoAD decomposes paired arm trajectory or action-token features into swap-symmetric components for shared coordination and swap-antisymmetric components for role-specific distinctions. The components are recomposed as residual updates to the original representations.

This modular transformation preserves the policy inputs and imitation-learning objective, and requires no manually defined role labels. Experiments across balanced and imbalanced role distributions show improved robustness over corresponding base policies, with notable gains on underrepresented configurations.

Read the full paper
BiRoAD overview: diverse bimanual tasks, imbalanced training configurations, and shared and role-specific residual feature updates
Figure 1: Overview of BiRoAD. (a) Bimanual tasks in which scene initializations induce different functional roles for two arms. (b) Under imbalanced training distributions, a base policy may favor frequent arm–role configurations and perform less reliably when the scene induces a less common role assignment. (c) BiRoAD decomposes paired arm features into swap-symmetric shared coordination and swap-antisymmetric role-specific components, then recomposes them as residual updates to encourage structured cross-arm sharing while preserving scene-dependent role adaptation.

02 / Method

Share what stays.
Adapt what changes.

Decompose paired features, learn complementary components, and recompose them as residual updates.

01

Decompose

Map left- and right-arm features to their mean and signed difference.

m = ½ (fL + fR)
d = ½ (fL − fR)
02

Share & specialize

A symmetric branch models shared coordination; an antisymmetric branch models role-specific variation.

Shared · swap-invariantRole-specific · sign-reversing
03

Recompose

Add the shared component to both arms and the role-specific component with opposite signs.

ΔfL = WSzS + WAzA
ΔfR = WSzS − WAzA
BiRoAD module and policy integration: symmetric and antisymmetric branches refine contextualized and pre-output action features inside an existing action decoder
BiRoAD integrates into the action-generation pathway of 3DFA and a π₀-style policy. It imposes a local feature-level structure; the full policy is not required to be strictly equivariant. Figure 2.
No manual role labelsUnchanged policy interfaceUnchanged training objective

03 / Real-world experiments

See the roles adapt.

Five tasks. Two scene configurations per task.
All clips are shown at 3× recorded speed.

A pair of Franka Research 3 arms performs the same task under different object configurations. Compare the paired demonstrations to see how the arms change their functional roles.

01

Bouquet handover

3× speed
Scene configuration A
Scene configuration B
02

Put banana into drawer

3× speed
Scene configuration A
Scene configuration B
03

Open bottle cap

3× speed
Scene configuration A
Scene configuration B
04

Sweep to dustpan

3× speed
Scene configuration A
Scene configuration B
05

Pour water

3× speed
Scene configuration A
Scene configuration B

Paired clips start together and retain their individual durations. Each demonstrates the same task in a different scene.

04 / Citation

BibTeX

@inproceedings{shen2026biroad,
  title  = {{BiRoAD}: Learning Shared and Role-Adaptive
            Representations for Bimanual Manipulation},
  author = {Shen, Yan and Liu, Yuchen and Jiang, Feng and
            Hu, Hangtian and Li, Xiaoqi and Chen, Shu and
            Wu, Ruihai and Dong, Hao},
  booktitle = {Proceedings of the Conference on Robot Learning (CoRL)},
  year   = {2026}
}