RoboStressBench - Planning: leaderboard
RoboStressBench (2026) tests vision-language models on embodied-scene images under physically grounded visual stress (dark, transparent or specular materials, extreme viewpoints, occlusion and clutter, over- and under-exposure). This board is accuracy in percent on its 821 planning multiple-choice questions, with greedy decoding and at most 64 new tokens; higher is better.
Metric: Accuracy (%) on the 821 planning multiple-choice questions, exact match on the predicted option, greedy decoding with at most 64 new tokens (API models: temperature 0, 64 output tokens), one unified instruction template; images of embodied scenes under physical visual stress (material, viewpoint, lighting, geometry); higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 35B A3B | 92.9 | #482 |
| 2 | Qwen 3.6 35B A3B | 81.1 | #462 |
| 3 | Qwen 3.5 27B | 77 | #359 |
| 4 | Qwen 3.5 4B | 68.8 | #1175 |
| 5 | Qwen 3.6 27B | 68.8 | #341 |
| 6 | Qwen 3.5 9B | 65.5 | #803 |
| 7 | GPT-5.5 | 57 | #49 |
| 8 | Gemini 3.1 Pro (Preview) | 56.3 | #94 |
| 9 | Molmo2-8B | 54.3 | #1483 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=robostressbench-planning · How It Works · Data refreshed daily, snapshot 2026-11-13.