ST-BiBench - Sequential Collaborative Manipulation: leaderboard
Metric: Success rate (%) over the 8 sequential collaborative manipulation tasks (hand-overs and multi-stage placements where one arm's step depends on the other's), each run for at least 100 domain-randomized episodes in which the model plans a sequence of seven parameterized bimanual action primitives executed through an API, averaged over the tasks it was scored on; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 28 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 69.38 | #145 |
| 2 | GPT-5 | 59.75 | #91 |
| 3 | Gemini 2.5 Flash | 59 | #237 |
| 4 | Qwen 3 VL 32B Instruct | 50.88 | #276 |
| 5 | Qwen 3 VL 235B A22B Instruct | 50.88 | #264 |
| 6 | Qwen 2.5 VL 32B Instruct | 50.13 | #443 |
| 7 | InternVL3-38B | 49.38 | #395 |
| 8 | Claude Sonnet 4 (20250514) | 46.63 | #211 |
| 9 | GPT-4o | 45.5 | #333 |
| 10 | Claude 3.7 Sonnet (20250219) | 45 | #196 |
| 11 | GPT-4.1 | 42.88 | #240 |
| 12 | Gemini 2.0 Flash | 41.25 | #331 |
| 13 | Qwen 2.5 VL 72B Instruct | 37.25 | #364 |
| 14 | InternVL3-78B | 33.63 | #345 |
| 15 | Llama 4 Scout Instruct | 29.75 | #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-sequential-collaborative-manipulation · How It Works · Data refreshed daily, snapshot 2026-10-11.