RT-SAFE: Agents Hit 94.1% Task Success but Only 0.7% Safe Success
Lianhuiq · x · 2026-10-13
Researchers introduce RT-SAFE, a real-time safety evaluation method for AI agents in the physical world. Key findings:
- Agents can complete tasks (94.1% task success) but violate safety constraints along the way—safe success is only 0.7%
- Since the physical world doesn't pause while agents think, real-time evaluation exposes 12.3× more dangerous collisions than paused-world tests
The takeaway: conventional paused-world benchmarks drastically underestimate embodied agent safety risks, and evaluation must happen in dynamic environments.
More from Embodied
- Collov Labs launches NewEyes AI, a personal agent with eyes — Scobleizer · 2026-10-13
- Amazon now has over 1 million robot workers. What's next for human employees? — lulzxdxdxd · 2026-10-13
- Chris Paxton on IROS: dexterous, contact-rich manipulation is now robotics' frontier — chris_j_paxton · 2026-10-13
- Scoble hails non-invasive BCI from Phantom Neuro after quadriplegic user demo — Scobleizer · 2026-10-13
- Most iPhone Duo apps look awful for now: old-SDK apps sit centered with empty space until iOS 27 SDK — steipete · 2026-10-13
- Dev to test $13 voice board with Codex and GPT Live in DIY AI device — OpenAIDevs · 2026-10-13