Few-shot Robot Manipulation on REASSEMBLE (6 held-out tasks)
82Success Rateπ0.5 fine-tuned (UB)
Evaluation results
| Method | Links | |
|---|---|---|
| π0.5 fine-tuned (UB)m=50, architecture=π0.5, protocol=fine-tuning upper bound2026.05 | 82 | |
| π0.5-primitivem=10, strategy=primitive-aware, architecture=π0.52026.05 | 81 | |
| OpenVLA fine-tuned (UB)m=50, architecture=OpenVLA, protocol=fine-tuning upper bound2026.05 | 79 | |
| OpenVLA-primitivem=10, strategy=primitive-aware, architecture=OpenVLA2026.05 | 78 | |
| π0.5-primitivem=5, strategy=primitive-aware, architecture=π0.52026.05 | 74 | |
| OpenVLA-primitivem=5, strategy=primitive-aware, architecture=OpenVLA2026.05 | 71 | |
| π0.5-primitivem=3, strategy=primitive-aware, architecture=π0.52026.05 | 66 | |
| OpenVLA-primitivem=3, strategy=primitive-aware, architecture=OpenVLA2026.05 | 62 | |
| OpenVLA-flatm=10, strategy=flat, architecture=OpenVLA2026.05 | 61 | |
| π0.5-flatm=10, strategy=flat, architecture=π0.52026.05 | 58 | |
| π0.5-primitivem=1, strategy=primitive-aware, architecture=π0.52026.05 | 44 | |
| OpenVLA-flatm=5, strategy=flat, architecture=OpenVLA2026.05 | 42 | |
| OpenVLA-primitivem=1, strategy=primitive-aware, architecture=OpenVLA2026.05 | 41 | |
| π0.5-flatm=5, strategy=flat, architecture=π0.52026.05 | 39 | |
| OpenVLA-flatm=3, strategy=flat, architecture=OpenVLA2026.05 | 34 | |
| π0.5-flatm=3, strategy=flat, architecture=π0.52026.05 | 31 | |
| π0.5-primitivem=0, strategy=primitive-aware, architecture=π0.52026.05 | 31 | |
| OpenVLA-primitivem=0, strategy=primitive-aware, architecture=OpenVLA2026.05 | 27 | |
| OpenVLA-flatm=1, strategy=flat, architecture=OpenVLA2026.05 | 24 | |
| π0.5-flatm=1, strategy=flat, architecture=π0.52026.05 | 22 | |
| OpenVLA-flatm=0, strategy=flat, architecture=OpenVLA2026.05 | 18 | |
| π0.5-flatm=0, strategy=flat, architecture=π0.52026.05 | 15 | |
| External-planner-only baseline2026.05 | 4 |