ST-BiBench - Fine-Grained Action Control: leaderboard
Metric: Success rate (%) averaged over 5 atomic bimanual tasks in which the model directly outputs continuous 16-dimensional action streams (7-dimensional pose and 1-dimensional gripper per arm) from images plus ground-truth pose metadata, at least 100 episodes per task; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 66.8 | #91 |
| 2 | Gemini 2.5 Pro | 60.2 | #145 |
| 3 | Gemini 2.5 Flash | 53.6 | #237 |
| 4 | InternVL3-78B | 27.6 | #345 |
| 5 | Claude Sonnet 4.5 | 25.4 | #138 |
| 6 | Qwen 3 VL 235B A22B Instruct | 25.2 | #264 |
| 7 | Gemma 3 27B (IT) | 6.2 | #509 |
| 8 | Llama 4 Scout Instruct | 6 | #526 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=st-bibench-fine-grained-action-control · How It Works · Data refreshed daily, snapshot 2026-10-11.