Six frontier models drive real robot arms into hazards; LEASH guardrail cuts harm to zero

新智元 · wechat · 2026-11-17

A multi-institution team (XJTLU, DKU, NUS, CNRS) released the RoboRisk benchmark and LEASH guardrail (arXiv:2611.14115), finding all six frontier models drive real robot arms into dangerous behavior.

RoboRisk: 92 sim + 13 real tasks across five risk categories (harmful goals, physical danger, environmental damage, injection attacks, privacy/safety), each executed closed-loop — prior benchmarks covered at most two categories.

Findings:

LEASH: model-agnostic, no retraining — task/action/commit gates plus a shared safety policy, each check vision-based and scored by a fast System-1 decision model (default Jev). Results: GPT-6 Astra sim harm 11.23%→0%; four models at zero harm on real hardware; +1.66% runtime, safety gates cost 0.66% of total. Weak instruction-followers (GPT-6 Luna, Grok 4.7) improve least. Benchmark, code, and results are open-sourced.

Original post →

More from Embodied

Embodied channel →