Complex Control on complex-control target suite (test)
70AntMaze AUCMSRL + pretrained Fηsrc init
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| MSRL + pretrained Fηsrc initInitialized component=matrix-value branch2026.05 | 70 | 75 | 78 | 76 | 67 | 73 | |
| MSRL from scratchInitialized component=target-trained matrix value2026.05 | 61 | 67 | 70 | 68 | 59 | 65 | |
| TD-MPC-PT+FTInitialized component=latent dynamics/value2026.05 | 60 | 65 | 68 | 66 | 53 | 63 | |
| DreamerV3-PT+FTInitialized component=latent world model/value2026.05 | 58 | 63 | 60 | 64 | 55 | 61 | |
| Traj2Value-Transformer-FTInitialized component=transformer trajectory-value encoder2026.05 | 57 | 62 | 65 | 63 | 54 | 60 | |
| MSRL w/o obstructionInitialized component=target-trained matrix value2026.05 | 57 | 62 | 65 | 63 | 54 | 60 | |
| Subgraph-GNN-to-Value-FTInitialized component=graph value encoder2026.05 | 56 | 61 | 64 | 62 | 53 | 59 | |
| Traj2Value-RNN-FTInitialized component=recurrent trajectory-value encoder2026.05 | 55 | 60 | 63 | 61 | 52 | 58 | |
| MSRL w/o libraryInitialized component=target-trained matrix value2026.05 | 54 | 60 | 63 | 61 | 52 | 58 | |
| DBC/Bisimulation-FTInitialized component=bisimulation encoder/value head2026.05 | 53 | 58 | 61 | 59 | 50 | 56 | |
| SF-GPI-FTInitialized component=successor features2026.05 | 52 | 55 | 58 | 56 | 47 | 53 | |
| UVFA-FTInitialized component=universal value encoder/head2026.05 | 51 | 57 | 60 | 58 | 49 | 55 | |
| DreamerV3Initialized component=none2026.05 | 50 | 57 | 60 | 58 | 49 | 55 | |
| TD-MPCInitialized component=none2026.05 | 49 | 58 | 62 | 60 | 51 | 57 | |
| MBPOInitialized component=none2026.05 | 47 | 48 | 57 | 55 | 46 | 52 | |
| SACInitialized component=none2026.05 | 43 | 50 | 53 | 51 | 42 | 48 | |
| TD3Initialized component=none2026.05 | 41 | 48 | 51 | 50 | 40 | 46 | |
| PPOInitialized component=none2026.05 | 36 | 41 | 45 | 53 | 34 | 40 |