Dispatchable Coordination Envelopes for Embodied Collaboration Under Sequential Uncertainty
Abstract
Embodied AI systems can pass planning and still fail in execution: a delay, resource conflict, synchronization change, or candidate invalidation can turn a valid point schedule into broken commitments. We introduce FLECAR, which gives the executor a dispatchable coordination envelope instead of fixed start times. Offline, Flexible Envelope Synthesis builds executable windows, coupling constraints, and a dispatch graph; online, Controllability-Aware Repair updates that envelope after events and repairs only the affected region when dispatchability fails. On generated coordination instances with the same event streams and budgets, FLECAR improves post-event feasibility from 0.3880 to 0.8400, terminal success from 0.5800 to 0.8400, and normalized repair burden from 0.5357 to 0.2964 relative to legacy point-schedule repair. A point baseline with CAR-like repair closes part of the gap, but envelopes still retain higher post-event feasibility and lower repair burden, while pooled one-step survivability remains near parity. The result points to a simple execution principle: after disturbances, live windows, couplings, and alternatives serve the executor better than a point schedule rebuilt after the fact.