Robotic Manipulation on RoboTwin (Downstream Task Set)
61Rank-Blk Success RateCreFlow
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| CreFlowType=Video RL post-training, Base=Vidar, Reward=Binary compositional2026.05 | 61 | 24 | 49 | 75 | 54 | 61 | 57 | 44 | 53.1 | |
| VidarType=Policy-level reference, Reference Scale=Absolute2026.05 | 52 | 3 | 22 | 39 | 25 | 42 | 36 | 24 | 30.4 | |
| DiffusionNFTType=Video RL post-training, Base=Vidar, Reward=Binary compositional2026.05 | 52 | 14 | 36 | 58 | 41 | 53 | 49 | 36 | 42.4 | |
| DanceGRPOType=Video RL post-training, Base=Vidar, Reward=Binary compositional2026.05 | 48 | 5 | 20 | 38 | 22 | 41 | 35 | 26 | 29.4 | |
| EVAType=Video RL post-training, Base=Vidar, Reward=Continuous IDM-based2026.05 | 45 | 5 | 20 | 44 | 39 | 56 | 50 | 15 | 34.3 | |
| π0.5Type=Policy-level reference, Reference Scale=Absolute2026.05 | 25 | 11 | 33 | 37 | 9 | 49 | 46 | 31 | 30.1 |