Robot Manipulation on LIBERO (All four suites combined)
99.2Spatial Success RateTwinBrainVLA
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| TwinBrainVLATraining Mode=Combined task suites2026.01 | 99.2 | 99 | 96.8 | 95.4 | 97.6 | — | |
| π0.5Training Mode=Combined task suites2026.01 | 98.8 | 98.2 | 98 | 92.4 | 96.9 | — | |
| π0.5Embodied pretraining (Emb. PT.)=true2026.04 | 98.8 | 98.2 | 98 | 92.4 | 96.9 | — | |
| PFDEmbodied pretraining (Emb. PT.)=false, Evaluation protocol=partial fine-tune2026.04 | 98.6 | 99.2 | 99.2 | 95.4 | 98.1 | — | |
| LingBot-VAEmbodied pretraining (Emb. PT.)=true2026.04 | 98.5 | 99.6 | 97.2 | 98.5 | 98.5 | — | |
| SDPmode=supervised fine-tuning2026.01 | 98.3 | 99.8 | 95.6 | 93.8 | 96.9 | — | |
| Fast-WAMEmbodied pretraining (Emb. PT.)=false, Evaluation protocol=released2026.04 | 98.2 | 100 | 97 | 95.2 | 97.6 | — | |
| MPCoTEvaluation Protocol=One policy for all 4 suites2026.06 | 98.2 | 99.7 | 98.6 | 98.9 | 98.9 | — | |
| OpenVLA-OFTEvaluation Protocol=One policy for all 4 suites2026.06 | 97.7 | 98 | 96.1 | 95.3 | 96.8 | — | |
| AVA-VLAEvaluation Protocol=One policy for all 4 suites2026.06 | 97.4 | 99.4 | 97.4 | 97.6 | 98 | — | |
| PFDEmbodied pretraining (Emb. PT.)=false, Evaluation protocol=adapter-only, Ablation=theta' = empty2026.04 | 97.2 | 98.8 | 96.6 | 93.8 | 96.6 | — | |
| Fast-WAMEmbodied pretraining (Emb. PT.)=false, Evaluation protocol=reproduced2026.04 | 97 | 99.4 | 96.6 | 94.8 | 96.95 | — | |
| π0Training Mode=Combined task suites2026.01 | 96.8 | 98.8 | 95.8 | 85.2 | 94.1 | — | |
| π0Embodied pretraining (Emb. PT.)=true2026.04 | 96.8 | 98.8 | 95.8 | 85.2 | 94.1 | — | |
| MotusEmbodied pretraining (Emb. PT.)=true2026.04 | 96.8 | 99.8 | 96.6 | 97.6 | 97.7 | — | |
| π0Evaluation Protocol=One policy for all 4 suites2026.06 | 96.8 | 98.8 | 95.8 | 85.2 | 94.2 | — | |
| Ours-πPerturbation Type=Agent Camera2026.03 | 96.4 | — | — | 92.5 | 94.5 | 94.8 | |
| π0-FASTEvaluation Protocol=One policy for all 4 suites2026.06 | 96.4 | 96.8 | 88.6 | 60.2 | 85.5 | — | |
| UnifiedVLAEvaluation Protocol=One policy for all 4 suites2026.06 | 95.4 | 98.8 | 93.6 | 94 | 95.5 | — | |
| UniVLAmode=supervised fine-tuning2026.01 | 95.2 | 95.4 | 91.9 | 87.5 | 92.5 | — | |
| Ours-OVPerturbation Type=Agent Camera2026.03 | 95.2 | — | — | 77.8 | 85.6 | 83.8 | |
| π0.5Perturbation Type=Agent Camera2026.03 | 92.4 | — | — | 39.9 | 67.9 | 71.4 | |
| π*0.5Perturbation Type=Agent Camera2026.03 | 87.6 | — | — | 86.2 | 87.2 | 87.9 | |
| WorldVLAEvaluation Protocol=One policy for all 4 suites2026.06 | 87.6 | 96.2 | 83.4 | 60 | 81.8 | — | |
| OpenVLATraining Mode=Combined task suites2026.01 | 87.4 | 88.4 | 79.2 | 53.7 | 76.5 | — | |
| GeoAwareVLAPerturbation Type=Agent Camera2026.03 | 87.1 | — | — | 84 | 86.1 | 87.2 | |
| OpenVLA-OFTPerturbation Type=Agent Camera2026.03 | 85.2 | — | — | 46.2 | 62.1 | 54.9 | |
| OpenVLAmode=supervised fine-tuning2026.01 | 84.7 | 88.4 | 79.2 | 53.7 | 76.5 | — | |
| OpenVLAEmbodied pretraining (Emb. PT.)=true2026.04 | 84.7 | 88.4 | 79.2 | 53.7 | 76.5 | — | |
| TraceVLAEvaluation Protocol=One policy for all 4 suites2026.06 | 84.6 | 85.2 | 75.1 | 54.1 | 74.8 | — | |
| Octomode=supervised fine-tuning2026.01 | 78.9 | 85.7 | 84.6 | 51.1 | 75.1 | — | |
| MDTmode=supervised fine-tuning2026.01 | 78.5 | 87.5 | 73.5 | 64.8 | 76.1 | — | |
| DiffPolicymode=supervised fine-tuning2026.01 | 78.3 | 92.5 | 68.3 | 50.5 | 72.4 | — | |
| MaILmode=supervised fine-tuning2026.01 | 74.3 | 90.1 | 81.8 | 78.6 | 83.5 | — | |
| UniActionsmode=supervised fine-tuning2026.01 | 65 | 78 | 68 | 47 | 64.5 | — |