Robot Manipulation on VL-Think
62.5Arrow ScoreVLA Grounder (OpenVLA)
Evaluation results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| VLA Grounder (OpenVLA)VLA Backbone=OpenVLA, Language-Conditioning Policy Model=Qwen3.5-9B, Optimization=GRPO2026.07 | 62.5 | 82.8 | 37.5 | 65.6 | 74.5 | 63 | 61.5 | 63.9 | |
| GEPAVLA Backbone=OpenVLA, Optimization=GEPA2026.07 | 51.3 | 73.2 | 47 | 48.5 | 59.5 | 49.4 | 57.2 | 55.2 | |
| TextGradVLA Backbone=OpenVLA, Optimization=TextGrad2026.07 | 44 | 61.2 | 19.1 | 25.6 | 54.2 | 35.8 | 30.4 | 38.6 | |
| VLA Grounder (pi-0)VLA Backbone=pi-0, Language-Conditioning Policy Model=Qwen3.5-9B, Optimization=GRPO2026.07 | 41.7 | 40.1 | 25 | 43.8 | 43.2 | 32.3 | 44.8 | 38.7 | |
| OpenVLA (Qwen3.5-9B w/o GRPO)VLA Backbone=OpenVLA, Language-Conditioning Policy Model=Qwen3.5-9B, Optimization=none2026.07 | 21.4 | 63.5 | 21.9 | 27.1 | 49 | 28.1 | 31.3 | 34.6 | |
| OpenVLA (orig)VLA Backbone=OpenVLA, Conditioning=original instructions2026.07 | 13.5 | 55.7 | 15.6 | 21.4 | 27.6 | 16.7 | 20.3 | 24.4 | |
| pi-0 (Qwen3.5-9B w/o GRPO)VLA Backbone=pi-0, Language-Conditioning Policy Model=Qwen3.5-9B, Optimization=none2026.07 | 4.7 | 25 | 10.4 | 14.6 | 21.4 | 8.3 | 14.6 | 14.1 | |
| pi-0 (orig)VLA Backbone=pi-0, Conditioning=original instructions2026.07 | 4.2 | 32.3 | 12.5 | 9.4 | 12 | 5.7 | 12 | 12.6 |