Robot Manipulation on SimplerEnv Google Robot tasks Visual Matching
98.1Pick Coke Can Success RateGreen-VLA
Evaluation Results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Green-VLAAlignment Round=R22026.01 | 98.1 | 50.8 | 61 | 69.9 | — | — | — | — | — | — | |
| ST4VLACo-Train=true2026.02 | 97.3 | 98 | 65.3 | 84.6 | — | — | — | — | 77.8 | — | |
| EcoVLASparsity=25%2026.01 | 95 | 82.1 | 70.8 | 71.7 | 38.9 | 1.45 | 72.65 | 1.44 | — | — | |
| Green-VLAAlignment Round=R12026.01 | 95 | 58.7 | 47 | 66.9 | — | — | — | — | — | — | |
| NORA-1.5 (DPO)Action Model Type=Continuous, Training Strategy=SFT+RFT2026.02 | 94 | 88 | 66.4 | 82.8 | — | — | — | — | — | — | |
| NORA-Long RFT (Ours)Action Model Type=Discrete, Training Strategy=RFT2026.02 | 94 | 84.7 | 58.5 | 79.1 | — | — | — | — | — | — | |
| GAE2025.10 | 94 | 92.5 | 62.5 | 83 | — | — | — | — | — | — | |
| VanillaSparsity=02026.01 | 93.3 | 83.8 | 74.5 | 73.3 | 41.7 | 1.81 | 104.16 | 1 | — | — | |
| EcoVLASparsity=40%2026.01 | 93 | 85.4 | 73.5 | 73.6 | 42.6 | 1.25 | 66.43 | 1.57 | — | — | |
| NORA-1.5Action Model Type=Continuous, Training Strategy=SFT2026.02 | 92.8 | 78.7 | 62.2 | 77.9 | — | — | — | — | — | — | |
| LACBase Model=CogAct2026.01 | 92.3 | 84.2 | 72.7 | 75.5 | — | 1.511 | 37.82 | — | 52.8 | — | |
| SoFAR2025.10 | 92.3 | 91.7 | 40.3 | 74.8 | — | — | — | — | — | — | |
| ThinkActAction Model Type=Continuous, Training Strategy=SFT + RFT2026.02 | 92 | 72.4 | 50 | 71.5 | — | — | — | — | — | — | |
| VLA-CacheBase Model=CogAct2026.01 | 92 | 83.3 | 70.5 | 74.4 | — | 1.496 | 39.63 | — | 51.6 | — | |
| CogACTCo-Train=false2026.02 | 91.3 | 85 | 71.8 | 74.8 | — | — | — | — | 50.9 | — | |
| Vanilla Co-training VLACo-Train=true2026.02 | 91.3 | 75.1 | 55 | 70.2 | — | — | — | — | 59.4 | — | |
| Baseline (CogAct)Base Model=CogAct2026.01 | 91.3 | 85 | 71.8 | 74.8 | — | 1.847 | 54.29 | — | 50.9 | — | |
| Green-VLAAlignment Round=R02026.01 | 90.4 | 61.2 | 62.9 | 71.4 | — | — | — | — | — | — | |
| Vanilla VLACo-Train=false2026.02 | 90 | 69.8 | 52.5 | 66.1 | — | — | — | — | 52.2 | — | |
| NORA-1.5-FASTAction Model Type=Discrete, Training Strategy=SFT2026.02 | 88.6 | 86.4 | 41.2 | 72.1 | — | — | — | — | — | — | |
| VLA-JEPAtraining_data=subsets of the OXE dataset, setting=visual matching2026.02 | 88.3 | 64.1 | 59.3 | 65.2 | — | — | — | — | — | 49.1 | |
| pi_0_starConfig=BF16 uniform2025.01 | 88 | 80.3 | 56 | 70.1 | — | — | — | — | — | — | |
| SpatialVLAEvaluation Protocol=Fine-tuning2025.01 | 86 | 77.9 | 57.4 | 75.1 | — | — | — | — | — | — | |
| SpatialVLAAction Model Type=Discrete, Training Strategy=SFT2026.02 | 86 | 77.9 | 57.4 | 73.7 | — | — | — | — | — | — | |
| NORA-Long (Baseline)Action Model Type=Discrete, Training Strategy=SFT2026.02 | 86 | 82.3 | 56 | 74.7 | — | — | — | — | — | — | |
| SpatialVLACo-Train=false2026.02 | 86 | 77.9 | 57.4 | 75.1 | — | — | — | — | — | — | |
| RT-1Training Stage=Converged, Evaluation Protocol=Fine-tuning, Dataset=Fractal [6]2025.01 | 85.7 | 44.2 | 73 | 74.6 | — | — | — | — | — | — | |
| RT-1Co-Train=false2026.02 | 85.7 | 44.2 | 73 | 52.4 | — | — | — | — | 6.5 | — | |
| VLA-JEPA (w/o human videos)training_data=subsets of the OXE dataset, setting=visual matching, ablation=w/o human videos2026.02 | 85.3 | 66.7 | 75.5 | 78.4 | — | — | — | — | — | 86.1 | |
| MagmaCo-Train=true2026.02 | 83.7 | 65.4 | 56 | 52.9 | — | — | — | — | 6.4 | — | |
| villa-xtraining_data=subsets of the OXE dataset, setting=visual matching2026.02 | 81.7 | 55.4 | 38.4 | 44.9 | — | — | — | — | — | 4.2 | |
| SpatialVLAEvaluation Protocol=Zero-shot2025.01 | 81 | 69.6 | 59.3 | 71.9 | — | — | — | — | — | — | |
| RT-2-XEvaluation Protocol=Zero-shot, Dataset=OXE [13]2025.01 | 78.7 | 77.9 | 25 | 60.7 | — | — | — | — | — | — | |
| RT-2-XCo-Train=true2026.02 | 78.7 | 77.9 | 25 | 46.3 | — | — | — | — | 3.7 | — | |
| RoboVLMEvaluation Protocol=Fine-tuning, Dataset=Fractal [6]2025.01 | 77.3 | 61.7 | 43.5 | 63.4 | — | — | — | — | — | — | |
| RoboVLMAction Model Type=Continuous, Training Strategy=SFT2026.02 | 77.3 | 61.7 | 43.5 | 63.4 | — | — | — | — | — | — | |
| RoboVLMstraining_data=subsets of the OXE dataset, setting=visual matching2026.02 | 77.3 | 61.7 | 43.5 | 51.7 | — | — | — | — | — | 24.1 | |
| π0-FASTAction Model Type=Discrete, Training Strategy=SFT2026.02 | 75.3 | 67.5 | 42.6 | 61.9 | — | — | — | — | — | — | |
| π0-FASTCo-Train=false2026.02 | 75.3 | 67.5 | 42.9 | 61.9 | — | — | — | — | — | — | |
| π0-Fasttraining_data=subsets of the OXE dataset, setting=visual matching2026.02 | 75.3 | 67.5 | 42.9 | — | — | — | — | — | — | — | |
| π0-FAST2025.10 | 75.3 | 67.5 | 42.9 | 61.9 | — | — | — | — | — | — | |
| MoTotraining_data=subsets of the OXE dataset, setting=visual matching2026.02 | 74 | 60.4 | 43.1 | — | — | — | — | — | — | — | |
| RoboVLMEvaluation Protocol=Zero-shot, Dataset=OXE [13]2025.01 | 72.7 | 66.3 | 26.8 | 56.3 | — | — | — | — | — | — | |
| π0Action Model Type=Continuous, Training Strategy=SFT2026.02 | 72.7 | 65.3 | 38.3 | 58.7 | — | — | — | — | — | — | |
| π0Co-Train=false2026.02 | 72.7 | 65.3 | 38.3 | 58.8 | — | — | — | — | — | — | |
| π0training_data=subsets of the OXE dataset, setting=visual matching2026.02 | 72.7 | 65.3 | 38.3 | — | — | — | — | — | — | — | |
| pi_0Mode=Fine-tune2026.01 | 72.7 | 65.3 | 38.3 | 58.8 | — | — | — | — | — | — | |
| π02025.10 | 72.7 | 65.3 | 38.3 | 58.8 | — | — | — | — | — | — | |
| RT-1Training Stage=15%, Evaluation Protocol=Fine-tuning, Dataset=Fractal [6]2025.01 | 71 | 35.4 | 56.5 | 60.2 | — | — | — | — | — | — | |
| GR00T N1.5Action Model Type=Continuous, Training Strategy=SFT2026.02 | 69.3 | 68.7 | 35.8 | 52.4 | — | — | — | — | — | — | |
| RT-1-XEvaluation Protocol=Zero-shot, Dataset=OXE [13]2025.01 | 56.7 | 31.7 | 59.7 | 53.4 | — | — | — | — | — | — | |
| RT-1-XAction Model Type=Discrete, Training Strategy=SFT2026.02 | 56.7 | 31.7 | 59.7 | 53.4 | — | — | — | — | — | — | |
| RT-1-XCo-Train=false2026.02 | 56.7 | 31.7 | 59.7 | 42.4 | — | — | — | — | 21.3 | — | |
| RT-1-X2026.01 | 56.7 | 31.7 | 59.7 | 53.4 | — | — | — | — | — | — | |
| RT-1-X2025.10 | 56.7 | 31.7 | 59.7 | 53.4 | — | — | — | — | — | — | |
| HPTEvaluation Protocol=Zero-shot, Dataset=OXE [13]2025.01 | 56 | 60 | 24 | 46 | — | — | — | — | — | — | |
| GR00T N1.5Co-Train=false2026.02 | 51.7 | 54 | 27.8 | 35.2 | — | — | — | — | 7.4 | — | |
| GR00T-N12025.10 | 47 | 70 | 18.1 | 45 | — | — | — | — | — | — | |
| TraceVLAEvaluation Protocol=Zero-shot, Dataset=OXE [13]2025.01 | 28 | 53.7 | 57 | 42 | — | — | — | — | — | — | |
| TraceVLAAction Model Type=Discrete, Training Strategy=SFT2026.02 | 28 | 53.7 | 57 | 42 | — | — | — | — | — | — | |
| OpenVLACo-Train=false2026.02 | 18 | 56.3 | 63 | 34.3 | — | — | — | — | 0 | — | |
| Octo-BaseEvaluation Protocol=Zero-shot, Dataset=OXE [13]2025.01 | 17 | 4.2 | 22.7 | 16.8 | — | — | — | — | — | — | |
| Octo-BaseAction Model Type=Continuous, Training Strategy=SFT2026.02 | 17 | 4.2 | 22.7 | 16.8 | — | — | — | — | — | — | |
| Octo-Base2025.10 | 17 | 4.2 | 22.7 | 16.8 | — | — | — | — | — | — | |
| OpenVLAEvaluation Protocol=Zero-shot, Dataset=OXE [13]2025.01 | 16.3 | 46.2 | 35.6 | 27.7 | — | — | — | — | — | — | |
| OpenVLAAction Model Type=Discrete, Training Strategy=SFT2026.02 | 16.3 | 46.2 | 35.6 | 27.7 | — | — | — | — | — | — | |
| OpenVLA2026.01 | 16.3 | 46.2 | 35.6 | 27.7 | — | — | — | — | — | — | |
| OpenVLA2025.10 | 16.3 | 46.2 | 35.6 | 32.7 | — | — | — | — | — | — | |
| RT-1Training Stage=Begin, Evaluation Protocol=Fine-tuning, Dataset=Fractal [6]2025.01 | 2.7 | 5 | 13.9 | 6.8 | — | — | — | — | — | — | |
| GR00T N1training_data=subsets of the OXE dataset, setting=visual matching2026.02 | 0.7 | 1.9 | 2.9 | 1.4 | — | — | — | — | — | 0 |