Preprint tests execution gate for LLM-directed robot teams

A new preprint describes a robot-orchestration architecture that separates a foundation-model planner from a deterministic execution gate. In drone-UGV simulations and two physical trials using Unitree G1 and Go2 interfaces, the authors report that retrieval improved skill grounding from 51% to 96%, but informed planners still sent 23% to 29% of faulted steps. The enforcement layer refused all eight injected faults before motion, compared with six faults causing robot movement without enforcement.