VLA Harness CALVIN: leaderboard
Long-horizon language-conditioned manipulation on the CALVIN ABC-to-D split, scored as the average number of sub-tasks completed in a chain of five rather than a success rate.
Metric: Avg Len (subtasks). Source: allenai.github.io. 66 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Xiaomi-Robotics-0 | 4.75 | |
| 2 | AVA-VLA | 4.65 | |
| 3 | FLOWER | 4.54 | |
| 4 | VLA-Adapter | 4.5 | |
| 5 | SDP | 4.49 | |
| 6 | Being-H0.7 | 4.48 | |
| 7 | DreamVLA | 4.44 | |
| 8 | X-VLA | 4.43 | |
| 9 | UniVLA | 4.41 | |
| 10 | MoDE | 4.4 |
Interactive version: theaggregate.ai/benchmark?slug=vla-harness-calvin · How It Works · Data refreshed daily, snapshot 2026-10-20.