Semantic Understanding on AIR-VLA
65.4Base Positioning AccuracySingle Head
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Single Headbase=π0.52026.06 | 65.4 | 44.7 | 82.7 | 11 | 40.2 | |
| Feature-MoEbase=π0.52026.06 | 65 | 45.1 | 96.4 | 10 | 41.2 | |
| Action-Level Soft MoEbase=π0.52026.06 | 62.5 | 39.5 | 46.6 | 5.5 | 32.3 | |
| π02026.06 | 55.7 | 36.1 | 85.1 | 2 | 32.3 | |
| π0-FAST2026.06 | 6.5 | 16.9 | 99.3 | 0 | 15.8 | |
| ACT2026.06 | 0 | 16 | 100 | 0 | 14 | |
| Diffusion Policy2026.06 | 0 | 16 | 100 | 0 | 14 | |
| Action-Level Hard MoEbase=π0.52026.06 | 0 | 19.2 | 100 | 0 | 14.8 |