VLA-Arena Extrapolation: leaderboard
Robot manipulation on combinations never seen in training — unseen objects, novel preposition combinations and rearranged task workflows.
Metric: Success Rate (%). Source: vla-arena.github.io. 30 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Paimon | 68.89 | |
| 2 | MolmoAct2 | 67.63 | |
| 3 | StellaVLA | 64.22 | |
| 4 | ACoT-VLA | 52.82 | |
| 5 | DM0.5 | 52.15 | |
| 6 | GR00T-N1.7 | 41.63 | |
| 7 | LingBot-VLA 2.0 | 39.48 | |
| 8 | LangForce | 37.67 | |
| 9 | LingBot-VLA | 37.11 | |
| 10 | OpenVLA | 36.67 |
Interactive version: theaggregate.ai/benchmark?slug=vla-arena-extrapolation · How It Works · Data refreshed daily, snapshot 2026-10-20.