VLA Harness VLABench: leaderboard
Language-conditioned manipulation tasks probing whether a policy actually grounds the instruction rather than pattern-matching the scene layout.
Metric: Success Rate (%). Source: allenai.github.io. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | π_0.5 | 62.2 | |
| 2 | X-VLA | 51.2 | |
| 3 | π_0-FAST | 49.6 | |
| 4 | pi-0 | 44.5 | |
| 5 | pi-0.5 | 42 | |
| 6 | pi-0.₅ | 24.3 |
Interactive version: theaggregate.ai/benchmark?slug=vla-harness-vlabench · How It Works · Data refreshed daily, snapshot 2026-10-20.