Two-stage compositional task on Real-robot tabletop suite (closed-loop evaluation)
59.2Success RateVISUALTHINK-VLA
Evaluation results
| Method | Links | |
|---|---|---|
| VISUALTHINK-VLAEvidence=stage-wise routing2026.05 | 59.2 | |
| FULLSOFTEvidence=stage-wise routing2026.05 | 57.4 | |
| BASEVLAEvidence=stage-wise routing2026.05 | 26.5 |