Robot Manipulation on Robot manipulation tasks m=3
82Success RateFull fine-tune (UB, 50 demos)
Evaluation Results
| Method | Links | |
|---|---|---|
| Full fine-tune (UB, 50 demos)Base Model=π0.5, Demonstration count (m)=50, Training paradigm=Fine-tuning2026.05 | 82 | |
| Full fine-tune (UB, 50 demos)Base Model=OpenVLA, Demonstration count (m)=50, Training paradigm=Fine-tuning2026.05 | 79 | |
| Primitive few-shotBase Model=π0.5, Demonstration count (m)=3, Training paradigm=Few-shot2026.05 | 66 | |
| Primitive few-shotBase Model=OpenVLA, Demonstration count (m)=3, Training paradigm=Few-shot2026.05 | 62 | |
| Octo-style demo-conditionedBase Model=π0.5, Demonstration count (m)=3, Training paradigm=Few-shot2026.05 | 44 | |
| Octo-style demo-conditionedBase Model=OpenVLA, Demonstration count (m)=3, Training paradigm=Few-shot2026.05 | 41 | |
| Diffusion Policy (CNN-based)Base Model=OpenVLA, Demonstration count (m)=3, Training paradigm=Zero-shot2026.05 | 36 | |
| Flat few-shotBase Model=OpenVLA, Demonstration count (m)=3, Training paradigm=Few-shot2026.05 | 34 | |
| Zero-shot primitive sequencingBase Model=π0.5, Demonstration count (m)=3, Training paradigm=Zero-shot2026.05 | 31 | |
| Flat few-shotBase Model=π0.5, Demonstration count (m)=3, Training paradigm=Few-shot2026.05 | 31 | |
| Zero-shot primitive sequencingBase Model=OpenVLA, Demonstration count (m)=3, Training paradigm=Zero-shot2026.05 | 27 |