LEASH: Stay Safe in Agentic Robotics
LEASH reduced harm across six robot-controlling foundation models while largely preserving task completion, the authors report.
Their RoboRisk benchmark runs 92 simulated tasks and 13 physical-robot tasks, covering refusal, plan repair, execution hazards, injection and sequential memory. Every task executes in closed loop, and the six models generally showed high harm rates. LEASH adds gates that review the task, each action and irreversible operations in their visual context under a shared safety policy. The authors report little added time or token overhead, though the abstract gives no numerical results.
HF Daily Papers' note
Their RoboRisk benchmark runs 92 simulated tasks and 13 physical-robot tasks, covering refusal, plan repair, execution hazards, injection and sequential memory. Every task executes in closed loop, and the six models generally showed high harm rates. LEASH adds gates that review the task, each action and irreversible operations in their visual context under a shared safety policy. The authors report little added time or token overhead, though the abstract gives no numerical results.
HF Daily Papers' note
score 5