Robot Policy Inference Efficiency on NVIDIA RTX 4090 simulation (inference)
0.6Inference Time (ms)DMPO
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| DMPOVision=light ViT, Action=MLP, Params=1.78M, Size=~28MB, Steps=12026.01 | 0.6 | 1,770 | 694 | |
| DMPOVision=light ViT, Action=MLP, Params=1.78M, Size=~28MB, Steps=22026.01 | 1.1 | 946 | 371 | |
| ShortCutVision=light ViT, Action=MLP, Params=1.78M, Size=~28MB, Steps=52026.01 | 2.5 | 396 | 155 | |
| DMPOVision=light ViT, Action=MLP, Params=1.78M, Size=~28MB, Steps=52026.01 | 2.6 | 386 | 151 | |
| MP1Vision=PointNet, Action=UNet, Params=255.80M, Size=~4GB, Steps=12026.01 | 4.1 | 244 | 96 | |
| CPVision=ResNet-18×2, Action=UNet, Params=284.86M, Size=~4.4GB, Steps=12026.01 | 5.4 | 187 | 73 | |
| MP1Vision=PointNet, Action=UNet, Params=255.80M, Size=~4GB, Steps=22026.01 | 7.7 | 130 | 51 | |
| ReFlowVision=light ViT, Action=MLP, Params=1.78M, Size=~28MB, Steps=202026.01 | 8.4 | 119 | 46 | |
| MP1Vision=PointNet, Action=UNet, Params=255.80M, Size=~4GB, Steps=52026.01 | 18.6 | 54 | 21 | |
| DPVision=ResNet-18×2, Action=UNet, Params=281.19M, Size=~4.4GB, Steps=10 (DDIM)2026.01 | 40.3 | 24.8 | 10 | |
| DPVision=ResNet-18×2, Action=UNet, Params=281.19M, Size=~4.4GB, Steps=16 (DDIM)2026.01 | 63.7 | 15.7 | 6 | |
| DPVision=ResNet-18×2, Action=UNet, Params=281.19M, Size=~4.4GB, Steps=100 (DDPM)2026.01 | 391.1 | 2.6 | 1 |