Two humanoids carry one payload through a narrow corridor: coordination becomes executable code, and a failed physical check rewrites roles

Team coordination is represented as executable code above fixed robot motor skills, separating individual capability from the cooperative strategy instantiated for the current phase. Existing multi-robot frameworks largely emphasize task decomposition, allocation and replanning, with limited support for adapting executable coordination when physical execution deviates from intended behavior. Two Unitree G1 humanoids in MuJoCo carry a freely supported rigid payload through constrained corridors, and the base task requires placement, release and upright robot retreat.

Coordination is instantiated through motion primitives, semantic objectives, temporary robot roles, coordination modes and bounded execution parameters rather than selected from a library of complete strategies. A task orchestrator commits the primitive, objective, mode and temporary roles, and a policy coder supplies bounded numerical parameters, yielding one policy per robot with one allowed action call followed by an explicit stop. The host validates the paired policies as one logical team transaction and dispatches them to separate SONIC deployments whose controllers keep independent observation histories and execution threads, so joint dispatch does not imply identical physical response.

A rejected transition triggers typed revision scoped to the maneuver, the role relation or the execution parameters, with fields outside that scope held fixed. The verifier decides acceptance from measured task progress, payload support and stability, robot configuration and environmental contacts, while a multimodal evaluator only assists scene interpretation and failure diagnosis. Revised policies are evaluated again from the latest accepted simulator checkpoint, and a fingerprint of checkpoint, normalized policy and effective arguments blocks identical rejected transactions at that checkpoint.

Evaluation runs in MuJoCo with two Unitree G1 humanoids carrying a freely supported rigid payload, and the model receives simulator measurements, global and payload-centered observations, and host-supplied route geometry and progress constraints. The base task requires placement, release and upright robot retreat within 80 physical transactions; layout and experience-reuse evaluations end at verified placement and release without retreat. Cross-backbone reuse transfers only strategy records and action exemplars, not simulator state, checkpoints, controller history or trajectories.

The authors report substantial wall-clock cost because model inference, physical execution, verification and recovery recur throughout a run. Rejected executions are recovered by restoring an accepted simulator state rather than a physically executable recovery policy, which limits direct transfer to real robots. The frozen motor interface provides bounded motion intentions rather than precise endpoints, and unequal robot responses, residual motion after stopping and accumulated formation drift can reduce clearance and destabilize the shared payload. Payload stability also depends on finite contact regions, friction and simulator contact dynamics, which are not systematically randomized; all experiments use simulation, privileged physical state information and supplied route references.