Diffusion Policy (UNet)
Formulates visuomotor control as conditional denoising diffusion over action sequences, with receding-horizon replanning. Averaged a 46.9% relative improvement over prior state of the art across 15 tasks.
- 76%
- 93K
- 10 Hz
- 5.5K
Benchmarks
Numbers as reported by the authors on the named evaluation suite.
| Evaluation | Suite | Score |
|---|---|---|
15-task average improvement | Robomimic + Push-T + kitchen | 46.9% |
Push-T | Push-T | 91% |
Robomimic Square (PH) | Robomimic | 98% |
Real bimanual tasks | Real Franka | 76% |
Training
Conditional DDPM visuomotor policy
110M
8x A100 40GB
320 accelerator-hours
60 h
5.5K
100-step DDPM (or 10-step DDIM) with FiLM conditioning, EMA weights, receding-horizon control.
Position control beats velocity control; EMA and observation horizon 2 are critical for stability.
Linked datasets, robots & papers
Resolving related models, datasets and papers…