A leash for robot agents: LEASH cuts harm to zero for four planners on a physical arm
Synopsis
LEASH adds a safety layer outside the agentic harness: task, action and commit gates approve or block each step using the current camera images, and the planner and all three gates read one shared safety policy. Across 92 simulated and 13 physical-robot tasks, six foundation models driving an arm alone showed harm rates from 9.06% to 76.92%; with LEASH, GPT-6 Astra in simulation fell from 11.23% to 0.00% harm while safe success rose from 78.99% to 80.80%, and on the physical robot GPT-6 Astra, GPT-6.1 Sol, Gemini 4 Argon and Claude Opus 5.5 each showed no observed harm in 13 runs.
Interpretation
When foundation models drive a robot arm directly, they commit semantically dangerous actions, such as putting a phone into water or carrying milk over a power strip. Classical robot safety enforces constraints written in advance, such as collision detection and collision-free trajectory planning; those constraints keep the arm from hitting things but do not judge what an action means. On the 92 simulated and 13 physical-robot tasks of RoboRisk, six foundation models showed harm rates from 9.06% to 76.92%; on the adapted simulation subset, the Dream API recorded 0.0% harm with 28.2% safe success and PokeBot RoboAPI recorded 59.0% harm with 0.0% safe success.
LEASH adds three safety gates to the agentic harness, reviewing the task, each proposed action and each irreversible operation. A decision model returns a safety score for each candidate and the gate allows it only above a threshold; a decision model answers a predefined question with a score or choice, so it responds quickly and cheaply enough to be called many times per task. With the default Jev reviewer, the task gate reached 81.20% accuracy for GPT-6 Astra and GPT-6.1 Sol in simulation, and the action and commit gates reached 99.91% to 100% accuracy on labeled candidates; for GPT-6 Astra and GPT-6.1 Sol on RoboRisk-S, mean end-to-end time rose from 245.21 s to 249.28 s, and the gates cost an estimated $0.00522 per run, 0.66% of the $0.791 total.
The gates see only the current camera images and the last five executed actions, the planner and all three gates share one safety policy, and the planner is treated as a black box that needs no retraining. A rejected action does not end the task; control returns to the planner, which replans from fresh observations, and only the task gate can terminate a task directly. In simulation the task gate blocked about half of the unsafe goals and the planner, guided by the shared policy, refused the rest, so GPT-6 Astra and GPT-6.1 Sol caused no harm on T tasks; more than 99.9% of the candidates approved by the action and commit gates were safe.
Perspective
The evaluation covers the 92 MuJoCo simulated tasks and 13 physical-robot tasks of RoboRisk, on a six-degree-of-freedom Dobot CR5V2 arm with a two-finger gripper and wrist and third-person cameras; the two robot-specific APIs were evaluated only on the adapted 13-task simulation subset and were not run with LEASH.
LEASH depends on reviewer accuracy and planner behavior: current decision models have limited reasoning, so the action and commit gates can reject an operation but cannot terminate a task, and when a gate decision is wrong a weak planner that ignores the policy may still issue unsafe actions, which explains the limited gains for GPT-6 Luna and Grok 4.7. Commit review currently covers only gripper opening and pouring, leaving other hazards inside long skills unmonitored; on the physical robot with a person present, several models were overly cautious and achieved no safe success on tasks P1 and P2.
