Technical Summary: Sequential Asymmetric Imitation for Learning Coupled Robot Policies

Problem Statement

Collaborative mobile manipulation involving two robots physically coupled through shared objects (rigid or deformable) presents a coordination challenge that transcends individual local skills. Success requires precise alignment of task phases, adaptation to partner delays, and yielding during interaction conflicts.

Existing approaches face significant bottlenecks:

• Model-based planning struggles with deformable objects, intermittent contacts, and unmodeled dynamics.
• Multi-agent reinforcement learning suffers from sparse rewards and simulation-to-real transfer difficulties.
• Standard imitation learning for dual-robot tasks typically requires synchronized joint demonstrations, where two operators control both robots simultaneously with precise temporal alignment. This data collection is expensive, requiring multiple teleoperation channels and calibrated interfaces.

The core question addressed is: Can coupled dual-robot policies be learned from single-teleoperator demonstrations without synchronized dual-operator data, reinforcement learning, or explicit inter-robot communication?

Methodology: Sequential Asymmetric Imitation (SAI)

The authors propose Sequential Asymmetric Imitation (SAI), a staged curriculum that learns physically coupled behaviors using a single teleoperator. The method decomposes coordination learning into three asymmetric stages, shifting the "partner distribution" seen by the robots from compliant support to autonomous interaction.

The Three-Stage Curriculum

The data collection sequence is defined as: $D^{solo}_A \rightarrow D^{demo}_{B|A} \rightarrow D^{int}_A$.

1. Stage 1: Bootstrapping Robot A (Solo with Human Support)

   • Setup: Robot A performs the task while a compliant human or passive partner supports the other side of the shared object.
   • Technique: Robot A learns basic task competence (grasping, pulling, aligning) via behavioral cloning.
   • Augmentation: To prevent overfitting to the human partner's visual appearance, partner-masking augmentation is applied. The supporting partner's region in frames is replaced with randomized perturbations (color jitter, blur, noise), forcing the policy to focus on the object state rather than the partner's identity.
   • Outcome: An initial policy $\pi^{(1)}_A$ capable of task execution with external support but lacking coordination with an autonomous robot.

2. Stage 2: Training Robot B Against Deployed Robot A

   • Setup: $\pi^{(1)}_A$ is frozen and deployed on Robot A. A single operator teleoperates Robot B while interacting with the deployed Robot A.
   • Technique: Robot A is executed with bounded variations (speed scaling, action noise, timing offsets) to expose Robot B to realistic deviations.
   • Outcome: Robot B learns to compensate for Robot A's timing, motion profile, and residual errors via behavioral cloning, producing $\pi^{(2)}_B$. This stage is asymmetric: Robot B learns against the actual learned behavior of Robot A, not an idealized or synchronized human partner.

3. Stage 3: DAgger-Style Intervention Fine-Tuning

   • Setup: Both $\pi^{(1)}_A$ and $\pi^{(2)}_B$ are deployed together.
   • Technique: A single operator performs sparse interventions on Robot A only when coordination failures occur (e.g., premature pulling, failure to yield, phase mismatch). This follows a DAgger-style approach where the learner induces states, and corrections are provided near failure boundaries.
   • Refinement: The intervention dataset $D^{int}_A$ is aggregated with the original solo demonstrations. Interventions are upweighted to teach Robot A partner-contingent behaviors (waiting, yielding, slowing) while preserving original task skills. Optional action-dimension masking allows correcting specific components (e.g., base motion) without altering others.
   • Outcome: A refined policy $\pi^{(3)}_A$ that has adapted to the autonomous behavior of Robot B.

Policy Instantiation

SAI is a curriculum-agnostic framework. In the experiments, it is instantiated with a decentralized ACT-style (Action Chunking Transformer) policy:

• Inputs: Local RGB observations (top-view and wrist cameras), proprioception, and previous actions.
• Architecture: Features are extracted via a frozen DINOv3 backbone, fused with state features, and encoded by a GRU history window ($H=30$ steps).
• Output: A cascaded action head predicts base commands ($v_x, v_y, \omega$) first, followed by arm commands conditioned on the base prediction. This ensures base-arm consistency.
• Constraint: No explicit inter-robot communication, privileged partner states, or centralized coordination signals are used.

Key Contributions

1. Sequential Asymmetric Imitation (SAI): A single-teleoperator curriculum for learning physically coupled dual-robot policies without synchronized dual-operator data, RL, or explicit communication.
2. Partner-Distribution Curriculum: Demonstration that shifting the training partner from compliant support $\to$ frozen learned policy $\to$ closed-loop robot-robot execution induces phase alignment and partner-contingent yielding.
3. Real-World Validation: Empirical evidence showing SAI improves task success, phase synchronization, and yielding behavior on contact-rich tasks involving rigid and deformable objects, outperforming independent imitation and curriculum-ablation baselines.

Experimental Results

The method was evaluated on four real-world dual-robot tasks: Bed-throw Spreading, Tablecloth Spreading, Laundry Collection, and Painting Transport.

• Task Success & Coordination: SAI significantly outperformed Independent Imitation (training robots separately) and Partner-Conditioned Imitation (training B against A, but without Stage 3 intervention).
  • Example (Laundry Collection): SAI achieved 70.0% success vs. 50.0% for Independent Imitation and 63.3% for Partner-Conditioned Imitation.
  • Yield/Wait Metric: SAI showed a marked improvement in partner-contingent yielding (e.g., 71.7% for Laundry vs. 31.7% for Independent Imitation), indicating the robots learned to slow or pause when the partner was delayed.
• Partner Delay Robustness: In controlled delay experiments (Tablecloth Spreading), Independent Imitation continued pulling despite the partner being paused, causing object misalignment. SAI successfully slowed and waited for the partner to recover.
• Policy Components: Ablation studies confirmed that temporal history tokens and cascaded base-arm heads improve execution reliability (reducing premature release and lowering), but the cooperative gains are primarily driven by the SAI data curriculum.
• Backbone Compatibility: SAI improved performance across different imitation backbones (ACT and Diffusion Policy), confirming it is a data-curriculum contribution rather than an artifact of a specific decoder.

Significance and Claims

The paper claims that physically coupled collaboration can be learned through the structure of the imitation curriculum rather than through synchronized multi-operator demonstrations or explicit coordination mechanisms.

• Implicit Policy Coupling: The results suggest that "partner-responsive coordination" can emerge from local observations and shared physical interaction when the training distribution is structured to include realistic partner behaviors (delays, mismatches).
• Practicality: By requiring only a single teleoperator and avoiding complex synchronization hardware or reinforcement learning, SAI offers a more scalable path to deploying collaborative mobile manipulators.
• Limitations: The authors note that SAI assumes robots can learn useful standalone primitives and that failures can be corrected via single-agent interventions. The current intervention dataset is relatively small (~30% of baseline size), suggesting that recovery behaviors have not yet saturated and could benefit from more targeted data collection around rare failure modes.