Reproduce “Streaming Diffusion Policy: Fast Policy Synthesis with Variable Noise Diffusion Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sigmund H. Høeg, Yilun Du, Olav Egeland
Streaming Diffusion Policy accelerates diffusion-policy synthesis by retaining partially denoised action chunks in a persistent buffer instead of discarding them. It denoises the imminent chunk more fully, carries later chunks at higher noise, and refines them after observations. On 20 real-world Push-T rollouts, it achieved 0.85 coverage and success at 0.07 s, outperforming Diffusion Policy but not Consistency Policy’s speed.
Key insights
- A robot requires a noise-free immediate action for execution, but subsequent actions in a planned trajectory can remain partially noisy and be refined as they become immediate.
- Reusing partially denoised future actions from a persistent buffer and recursively updating them drastically reduces the computational load per control cycle.
- Optimizing the training noise distribution to match the streaming inference's chunk-wise variable noise levels is crucial for maintaining high policy performance during acceleration.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.