LEASH reduced dangerous robot plans, but some harmful executions remained

A language-model planner can produce a syntactically valid sequence that is unsafe for a nearby person or physical object. LEASH inserts a task-specific safety reviewer and action gates between the planner and a robot execution interface, using a separate suite of risky robotics tasks.

The controlled evaluations report fewer harmful outcomes with the reviewer but include residual failures, including five harmful cases among thirteen selected Grok trials. The study covers simulations and selected physical-robot tests; it is not certification for arbitrary hardware, people or environments. The reliable question is whether the actuator refuses an unsafe motion under its actual physical constraints.

A safety reviewer improves a planner’s screening but does not remove the need for an independent physical interlock.

A refusal to propose a hazardous command cannot replace actuator constraints, collision control or emergency stop logic. This benchmark tests selected planned tasks and simulator/robot conditions; it neither proves a general factory safety certification nor measures real injuries avoided. The key deployment question is whether the final physical effect remains constrained after an imperfect model decision.