MetaFine - Grasp Part: leaderboard
Fine-grained robot manipulation from MetaFine (Xu et al., 2026): the policy must grasp an object at the instructed part, such as a mug handle, bottle cap or box lid, and a trial counts only when the part-level or directional constraint is met. Success rate in percent in simulation under nominal conditions, with every policy fine-tuned on the MetaFine demonstrations using its official recipe; higher is better.
Metric: Success rate (%; fine-grained acceptance: grasp at the instructed part (mug handle, bottle cap, cabinet handle, box lid, box handle); nominal simulation conditions; each policy fine-tuned on the MetaFine demonstrations from its released checkpoint with the authors' official recipe). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | π0.5 | 80 | |
| 2 | DP3 | 75 | |
| 3 | π0 | 75 | |
| 4 | ACT | 49 | |
| 5 | OpenVLA-OFT | 37 | |
| 6 | Octo | 35 | |
| 7 | OpenVLA | 35 |
Interactive version: theaggregate.ai/benchmark?slug=metafine-grasp-part · How It Works · Data refreshed daily, snapshot 2026-11-12.