Research map / Image generation / Fast diffusion sampling
One-Step Diffusion Policy: Fast Visuomotor Policies via Diffusion Distillation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Distills a pre-trained diffusion policy into a single-step visuomotor action generator via reverse-KL score-difference distillation, achieving comparable or higher success with ~42x faster inference and ~20x faster convergence than Consistency Policy.
Problem
Diffusion policies achieve strong imitation learning performance but require many iterative denoising steps, making action generation too slow (about 1.5 Hz) for real-time, dynamic robotic control.
Approach
Adapts score distillation (SDS/VSD-style) to diffusion policies: a one-step generator is trained by minimizing reverse KL between its action distribution and the pre-trained policy's, applied over diffused action samples using a generator score network initialized from the teacher, with a deterministic variant that drops the score network.
Key results
- OneDP raises action prediction frequency from 1.5 Hz to 62 Hz, with single-step generation (7 ms vs 660 ms for DDPM 100 steps on a V100).
- OneDP-S achieves the highest average success across six simulation tasks under DDPM, slightly above the pre-trained DP, while single-step DP scores 0%.
- OneDP converges about 20x faster than Consistency Policy under the EDM setting, and OneDP-S outperforms CP's single-step and 3-step settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Distill Diffusion Policy into a single-step action generator | Diffusion Policy · Real-world inference | action prediction time: 66 -> 7 ms | Helped |
| Distill Diffusion Policy into a single-step action generator | Diffusion Policy · Real-world action prediction | frequency: 1.5 Hz to 62 Hz | Helped |
| Use OneDP-S instead of Diffusion Policy | Diffusion Policy · Robomimic and PushT simulation tasks | success rate or target-area coverage | Helped |
| Use OneDP-S instead of OneDP-D (ablation) | OneDP · Robomimic and PushT simulation tasks | success rate or target-area coverage | Helped |
| Use OneDP-S instead of Diffusion Policy | Diffusion Policy · EDM simulation tasks | success rate | No clear effect |
| Use OneDP-D instead of Diffusion Policy (ablation) | Diffusion Policy · EDM simulation tasks | success rate | Hurt |
| Use OneDP-S instead of Consistency Policy | Consistency Policy · EDM simulation tasks | success rate | Helped |
| Use OneDP instead of Consistency Policy | Consistency Policy · EDM distillation | convergence speed: 20× | Helped |
| Use OneDP instead of Diffusion Policy | Diffusion Policy · pnp-milk-move | success rate | Helped |
| Use OneDP instead of Diffusion Policy | Diffusion Policy · Real-world tasks | task completion time | Helped |
| Use OneDP-S instead of OneDP-D (ablation) | OneDP · Real-world tasks | success rate | Helped |
| Set generator score network learning rate to 2e-5 (ablation) | OneDP · Ablation study tasks | performance | Helped |
Surprising findings
- Single-step OneDP-S slightly outperforms the multi-step pre-trained Diffusion Policy, which the authors attribute to cumulative errors in iterative denoising.
- Naively running the pre-trained Diffusion Policy in a single step yields a 0% success rate on all tasks.
Limitations
- Not tested on long-horizon real-world tasks.
- Real-world control limited to 20 Hz for stability, underutilizing the speed gain.
- KL-based distillation may not be optimal for distribution matching; discriminator term suggested as improvement.
Experimental setup
Datasets: Robomimic PushT real-world Franka demonstrations (pnp-milk, pnp-anything, coffee)
Benchmarks: Robomimic PushT real-world Franka pick-and-place and coffee tasks
Metrics: success rate inference latency (ms) action frequency (Hz) task completion time convergence speed
Baselines: Diffusion Policy (DDPM 100 steps) Diffusion Policy (DDIM 10 steps) Consistency Policy
Builds on: Diffusion Policy Consistency Policy Variational Score Distillation Diffusion-GAN EDM
Closest papers by meaning
- Two-Steps Diffusion Policy for Robotic Manipulation via Genetic Denoising (NeurIPS 2025 · similarity 0.92)
- Improving Generative Behavior Cloning via Self-Guidance and Adaptive Chunking (NeurIPS 2025 · similarity 0.91)
- Diffusion Policy Policy Optimization (ICLR 2025 · similarity 0.90)
- Prediction with Action: Visual Policy Learning via Joint Denoising Process (NeurIPS 2024 · similarity 0.90)
- Score Regularized Policy Optimization through Diffusion Behavior (ICLR 2024 · similarity 0.90)
- Model-based Diffusion for Trajectory Optimization (NeurIPS 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.