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:
- Harm rates of 9.06%–76.92% for GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna, Claude Opus 5.5, Gemini 4 Argon, and Grok 4.7 — none were both safe and useful
- Typical failures: phone placed in a water bowl, milk carried over a power strip, knife moved above a human hand, prompt-injection via tablet, allergen served despite disclosure
- A "Human-Presence Gap": models fine with humans in sim become overly conservative on real hardware
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.
More from Embodied
- CoRL fireside chat: Fei-Fei Li and AMD CEO Lisa Su on stage together — PeterStone_TX · 2026-11-17
- Distilling a frontier model into 4B params lifts robot arm success from 0 to 22.5% — ZeYanjie · 2026-11-17
- NeurIPS 2026 surgical benchmark: top multimodal models score under 5% zero-shot — maier_ak · 2026-11-17
- Musk retweets: Tesla Giga Berlin to build Cybercab — elonmusk · 2026-11-17
- Behavior Graphs turns raw robot policy rollouts into interpretable graphs — drmapavone · 2026-11-17
- Neuralink hackathon adds hands-free speech: iPhone calls, lock-screen controls, and a 'Hey Neuralink' Grok agent — XFreeze · 2026-11-17