RT-SAFE Safe Success Rate
Safety · 2026-10-07
Share of RT-SAFE's 108 real-time navigation episodes (36 routes at easy, medium and hard actor density, provider-default reasoning, one simulation seed) that reach the destination with no collision, hazard interaction or traffic violation. It is the paper's strictest safety outcome, but every evaluated model scores between 0 and 3.7%, so it is kept as a floor-saturated diagnostic beside the collisions primary.
Top models (higher is better)
| Model | Score |
|---|---|
| Sonnet 5 | 3.7 |
| GPT-5.6 Sol | 1.9 |
| Claude Fable 5.1 | 0.0 |
| DeepSeek V4 Flash Vision Exp | 0.0 |
| Gemini 3.8 Flash | 0.0 |
| GPT-6 Astra | 0.0 |
| Grok 4.6 | 0.0 |
| Inkling | 0.0 |