RT-SAFE Collisions per Episode
Safety · 2026-10-07
RT-SAFE (Cornell and UC San Diego) runs vision-language model agents along 36 pedestrian navigation routes on five Unreal Engine 5 SimWorld city maps while pedestrians, vehicles, movable objects and hazards keep moving during model inference. This row is mean collisions per episode in real-time mode at provider-default reasoning, averaged with equal weight over easy, medium and hard actor densities (108 episodes, one simulation seed), counting contacts during both inference and action execution and including failed episodes. Events are logged by the simulator without a judge; lower is better.
Top models (lower is better)
| Model | Score |
|---|---|
| Sonnet 5 | 19.6 |
| Claude Fable 5.1 | 20.9 |
| Inkling | 21.9 |
| GPT-6 Astra | 22.4 |
| GPT-5.6 Sol | 25.6 |
| Gemini 3.8 Flash | 27.7 |
| DeepSeek V4 Flash Vision Exp | 51.4 |
| Grok 4.6 | 59.5 |