RoboRisk benchmark shows foundation models pose high harm rates in robotics; LEASH harness cuts risk
UoLiverpool · hf · 2026-11-12
- A team from Xi'an Jiaotong-Liverpool, Duke Kunshan, University of Liverpool, NUS and CNRS finds that using foundation models to control robots poses serious safety risks, with existing safety evaluation lacking.
- RoboRisk benchmark: 92 simulated + 13 physical-robot tasks across five risk categories — Task-level Refusal, Plan Repair, Execution-time Hazard, Injection, Sequential Memory — all executed closed-loop; prior benchmarks covered at most two categories and several only evaluated text outputs.
- Findings: Six frontier models (GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna, Claude Opus 5.5, Gemini 4 Argon, Grok 4.7) plus two robot-specific APIs generally show high harm rates.
- LEASH: A Lease-Enforced Agentic Safety Harness using a decision model (Jev or OpenAI's Decisions API) with task, action and commit gates reviewing operations in visual context; experiments in sim and on real robots show LEASH reduces harm across all six models while largely preserving task completion, with minimal time and token overhead.
More from Embodied
- Abbeel lab's SuperMimic replaces physics sim with differentiable world model, ~1000x faster sim-to-real — HirokatuKataoka · 2026-11-12
- Berkeley open-sources RoboSimStudio: AI agents program industrial robots without training data — berkeley_ai · 2026-11-12
- NXP ships pre-tested drone platform: initial flight in 30 minutes out of the box — shashib · 2026-11-12
- Open-source dual-arm robot solves tangram puzzles with 90 demos, ~60% success — DanGrover · 2026-11-12
- Robotics has three regional playbooks: China volume, Europe inspection, US pilots — ingliguori · 2026-11-12
- Model Autonomously Designs a Surgical-Robot-Style Robotic Arm CAD After 75 Minutes — _Stocko_ · 2026-11-12