ObsDriveBench - Spatial Reliability: leaderboard
Metric: Accuracy (%; dimension average of 2D-3D grounding, spatial relation and interaction reasoning; multiple-choice questions on synchronized camera, LiDAR BEV and radar BEV inputs from K-Radar under fog, rain, snow and low light, answers matched by an LLM judge; official inference settings). Source: arxiv.org. Saturation forecast: Around December 2026. 15 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 88.04 |
| 2 | Qwen 3 VL 32B | 67.1 |
| 3 | GPT-5 | 61.04 |
| 4 | Qwen 2.5 VL 7B | 48.89 |
| 5 | InternVL3-8B | 47.65 |
| 6 | InternVL3-38B | 47.38 |
Interactive version: theaggregate.ai/benchmark?slug=obsdrivebench-spatial-reliability · How It Works · Data refreshed daily, snapshot 2026-09-29.