1
Temporal Phase Coupling
Robots must align grasping, pulling, lifting, transporting, lowering, and release despite variable execution timing.
Collaborative Dual-Robot Learning
Robots that Collaborate
Sequential Asymmetric Imitation for Learning Coupled Robot Policies
Dual-Robot Manipulation · Imitation Learning · Mobile Manipulation
Coupled dual-robot manipulation from a single-teleoperator curriculum.
Independent policies can execute local primitives, but fail under phase mismatch, partner delay, and physical interaction conflict. SAI teaches coordination through the data-collection curriculum.
Project Overview
Demo Video
Challenge
Why Is This Important?
Collaborative manipulation fails when two individually competent robots are physically coupled through a shared object.
2
Partner-Contingent Yielding
A robot must slow, pause, or re-time its motion when the partner is delayed, stalled, obstructed, or out of phase.
3
Interaction Conflict
Small timing errors can create large object deformation, internal forces, loss of grasp, or workspace conflict.
Method
Sequential Asymmetric Imitation
SAI decomposes dual-robot coordination into three single-teleoperator stages.
Overview of SAI. Robot A is first bootstrapped from unilateral demonstrations, Robot B is then trained against the deployed Robot-A policy, and Robot A is finally refined through sparse interventions near coordination failures.
1
Bootstrap Robot A
Teleoperate Robot A with a compliant human or passive partner. Train basic task execution from unilateral demonstrations.
→
2
Train Robot B Against A
Freeze and deploy Robot A. Teleoperate Robot B against the learned Robot-A policy to expose B to realistic partner behavior.
→
3
Intervene on Robot A
Deploy both robots and sparsely correct Robot A near coordination failures, such as early pulling, insufficient yielding, or recovery errors.
Core Idea
SAI induces coordination by shifting the partner distribution seen during imitation: compliant support → deployed learned partner → closed-loop robot-robot interaction with targeted corrections. The policies remain decentralized and do not exchange messages, partner states, future actions, or latent embeddings.
Benchmarks
Real-World Task Suite
The experiments cover deformable spreading, shared-workspace collection, and rigid-object transport.
Real-world task suite: bed-throw spreading, tablecloth spreading, laundry collection, and painting transport.
Bed-throw Spreading
Deformable object alignment with handoff, retreat, transport, and placement phases.
Tablecloth Spreading
Coordinated grasping and spreading under cloth tension and partner-delay perturbations.
Laundry Collection
Asynchronous shared-workspace coordination with approach, collection, and delivery.
Painting Transport
Rigid-object cooperative transport requiring synchronized grasping, lifting, motion, and lowering.
Empirical Results
Results
SAI improves both task completion and process-level coupling metrics over Independent Imitation and Partner-Conditioned Imitation.
0%
Best task success
0%
Painting phase sync.
0%
Laundry yield / wait
0pp
Max yield gain
Task Success
Rollout-level task completion across four real-world tasks.
Phase Synchronization
Event-level phase alignment across task-specific checkpoints.
Yield / Wait
Partner-contingent slowing, waiting, and resuming behavior.
Partner Delay Test
When Robot B is paused during tablecloth spreading, Independent Imitation continues pulling and destabilizes the cloth. SAI slows, waits, and resumes after the partner recovers.
Backbone Compatibility
SAI improves over Independent Imitation with both ACT and Diffusion Policy, indicating that the main contribution is the data curriculum rather than a specific action decoder.
Paper at a Glance
Paper Summary
Abstract
We study physically coupled dual-robot manipulation with two bimanual mobile manipulators. SAI is a single-teleoperator curriculum that trains Robot A from unilateral demonstrations, trains Robot B against the deployed Robot-A policy, and refines Robot A through sparse interventions near coordination failures. Across rigid and deformable real-world tasks, SAI improves success, phase synchronization, and partner-contingent yielding.
Contributions
1. A single-teleoperator curriculum for coupled dual-robot imitation learning.
2. A partner-distribution shift mechanism that induces yielding and phase alignment.
3. Real-world validation on deformable spreading, laundry collection, and painting transport.
Citation
BibTeX
Copy the citation directly into your paper, website, or project README.
@article{sai2026,
title = {Sequential Asymmetric Imitation for Learning Coupled Robot Policies},
author = {Yincong Chen and Ranpeng Qiu and Zihao Li and Yanan Zhou and Guoqiang Ren and Weiming Zhi},
journal = {arXiv preprint arXiv:2026.xxxxx},
year = {2026}
}