Robot Manipulation on LIBERO Object, Spatial, and Goal Suites
96.5Object Success RateVLA-AD (with-Qwen)
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| VLA-AD (with-Qwen)Teacher architecture=π0.5-4B, VLM supervision=Qwen, Number of trials=200 trials/cell, Action chunking (K)=52026.05 | 96.5 | 93 | 94 | 94.5 | |
| VLA-AD (no-Qwen)Teacher architecture=π0.5-4B, VLM supervision=none, Number of trials=200 trials/cell, Action chunking (K)=52026.05 | 94.5 | 94 | 90 | 92.8 | |
| π0.5 (teacher)Role=Teacher, Teacher architecture=π0.5-4B, Number of trials=200 trials/cell2026.05 | 90 | 87 | 94.5 | 90.5 |