RT-SAFE Hard Static Collisions per Episode
Safety · 2026-10-07
Mean collisions per episode on RT-SAFE's 36 hard-density routes in static mode, where the simulator freezes during model inference and resumes during action execution, with a mode-specific timing prompt, provider-default reasoning and one simulation seed. Because no contact can occur while the model decides, it is a different protocol from the real-time primary and is kept as a diagnostic; lower is better.
Top models (lower is better)
| Model | Score |
|---|---|
| Inkling | 1.2 |
| GPT-6 Astra | 2.5 |
| Grok 4.6 | 2.7 |
| Gemini 3.8 Flash | 2.8 |
| Claude Fable 5.1 | 3.3 |
| DeepSeek V4 Flash Vision Exp | 4.0 |
| Sonnet 5 | 4.5 |
| GPT-5.6 Sol | 5.4 |