Instruction Following on RoboTwin-IF Clean (Held-Out Templates)
79Pick Success Rate (Diverse)QWEN-ROBOTMANIP
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| QWEN-ROBOTMANIPTraining=fine-tuned on RoboTwin Clean, Evaluation=held-out unseen instruction templates2026.06 | 79 | 57 | 42 | 90 | 93 | 72.2 | |
| QWEN-ROBOTMANIP-ContextTraining=fine-tuned on RoboTwin Clean, Evaluation=held-out unseen instruction templates, Context=Included2026.06 | 77 | 71 | 33 | 89 | 90 | 72 | |
| π0.5Training=fine-tuned on RoboTwin Clean, Evaluation=held-out unseen instruction templates2026.06 | 44 | 20 | 15 | 92 | 66 | 49.6 | |
| GR00T-N1.7Training=fine-tuned on RoboTwin Clean, Evaluation=held-out unseen instruction templates2026.06 | 20 | 17 | 0 | 14 | 32 | 16.6 | |
| StarVLATraining=fine-tuned on RoboTwin Clean, Evaluation=held-out unseen instruction templates2026.06 | 11 | 13 | 0 | 49 | 74 | 29.4 |