Multi-task robotic manipulation on RoboTwin Simulation 2.0 (test)
95Adjust bottleSDP
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| SDPbackbone=CNN-based Diffusion Policy, MoE layers=1/4 of convolutional layers, num_experts=52026.01 | 95 | 15 | 73 | 35 | 79 | 5 | 42 | 24 | 46 | |
| Di-BMbackbone=CNN-based Diffusion Policy, MoE layers=1/4 of convolutional layers, num_experts=5, beta=0.012026.01 | 83 | 41 | 88 | 60 | 92 | 72 | 78 | 91 | 76 | |
| DPbackbone=CNN-based Diffusion Policy2026.01 | 82 | 6 | 85 | 55 | 95 | 31 | 52 | 83 | 61 | |
| MoDEbackbone=CNN-based Diffusion Policy, MoE layers=1/4 of convolutional layers, num_experts=52026.01 | 82 | 15 | 73 | 59 | 89 | 33 | 56 | 68 | 59 |