Robot Manipulation on LIBERO-V Across Novel Camera Viewpoints (unseen)
97.2Spatial Success Rateπ0.5
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| π0.5Adaptation=One-Shot LoRA Fine-tuned2025.12 | 97.2 | 95.3 | 91.4 | 77.1 | 90.3 | |
| π0.5 (One-Shot FLA)Adaptation=Feature Linear Adaptation, Protocol=One-Shot2025.12 | 96.5 | 97.5 | 91.7 | 77.5 | 90.8 | |
| π0Adaptation=One-Shot LoRA Fine-tuned2025.12 | 95.4 | 96.6 | 79.1 | 63.2 | 83.6 | |
| π0.5 (One Shot FTM)Adaptation=Feature Token Modulation, Protocol=One Shot2025.12 | 95.4 | 95.5 | 87.7 | 70 | 87.2 | |
| GeoAware-VLA (GeoAware BAKU)Policy=BAKU, Backbone=VGGT2025.12 | 94.3 | 98 | 90.7 | 47.3 | 82.6 | |
| GeoAware BAKUViewpoint=Unseen2025.09 | 94.3 | 98 | 90.7 | 47.3 | 82.6 | |
| OpenVLA-OFTProtocol=zero-shot2025.12 | 90 | 15.7 | 81.7 | 13.7 | 50.3 | |
| OpenVLA-OFTViewpoint=Unseen2025.09 | 90 | 15.7 | 81.7 | 13.7 | 50.2 | |
| OpenVLA-OFT-mProtocol=finetuned, Training Data=LIBERO-Plus2025.12 | 67.1 | 72.4 | 72 | 49.2 | 65.2 | |
| GeoAware-VLA (GeoAware VQ-BeT)Policy=VQ-BeT, Backbone=VGGT2025.12 | 54.3 | 99 | 85.7 | 72.7 | 77.9 | |
| GeoAware VQ-BeTViewpoint=Unseen2025.09 | 54.3 | 99 | 85.7 | 72.7 | 77.9 | |
| Evo-0 BAKUViewpoint=Unseen2025.09 | 52 | 98.3 | 88 | 28.3 | 66.6 | |
| VQ-BeTViewpoint=Unseen2025.09 | 41.3 | 38.3 | 83 | 3 | 41.4 | |
| BAKUViewpoint=Unseen2025.09 | 18 | 49.3 | 80.7 | 3.7 | 37.9 |