Long-horizon robot manipulation on CALVIN
96.3Task Completion Rate (1)OpenVLA-OFT
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| OpenVLA-OFTQuantization setting=Full Precision (FP)2026.02 | 96.3 | 89.1 | 82.4 | 75.8 | 66.5 | 4.1 | |
| AutoQVLAQuantization setting=W8A162026.02 | 95.9 | 88.6 | 81.4 | 72.1 | 63.2 | 4.03 | |
| AWQQuantization setting=W8A162026.02 | 95.7 | 87.5 | 76.8 | 69.5 | 64.3 | 3.97 | |
| LLaVA-VLA w/ UAORreproduced=true, uncertainty-aware observation reinjection=true2026.02 | 95.5 | 84.6 | 72.3 | 60.7 | 49.1 | 3.67 | |
| LLaVA-VLAreproduced=true2026.02 | 94.4 | 82 | 70.8 | 59.4 | 48.2 | 3.55 | |
| Vidman2026.02 | 91.5 | 76.4 | 68.2 | 59.2 | 46.7 | 3.42 | |
| OpenVLA2026.02 | 91.3 | 77.8 | 62 | 52.1 | 43.5 | 3.27 | |
| VLAS2026.02 | 87.2 | 64.2 | 40.9 | 28.1 | 19.6 | 2.4 | |
| pi_0.52026.01 | 86.7 | 73.2 | 61.8 | 54.3 | 45.9 | 3.22 | |
| GR-12026.02 | 85.4 | 71.2 | 59.6 | 49.7 | 40.1 | 3.06 | |
| RoboFlamingo2026.02 | 82.4 | 61.9 | 46.6 | 33.1 | 23.5 | 2.47 | |
| ANY3D-VLATraining data source=simulator point clouds, Point cloud estimation source=Simulator2026.01 | 72.7 | 61.5 | 52.9 | 46.2 | 34.8 | 2.68 | |
| SpatialVLAevaluation protocol=full-tuning2026.01 | 70 | 62.5 | 54.5 | 48 | 32.2 | 2.67 | |
| ANY3D-VLATraining data source=simulator point clouds, Point cloud estimation source=DA32026.01 | 69.1 | 57.1 | 50 | 44.4 | 32 | 2.53 | |
| GraspVLA2026.01 | 56.2 | 47.5 | 40.9 | 30 | 22.5 | 1.97 |