Reproduce “Block-wise Adaptive Caching for Accelerating Diffusion Policy” 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
Kangye Ji, Yuan Meng, Hanyun Cui, Ye Li, Jianbo Zhou, Shengjia Hua, Lei Chen, Zhi Wang
The authors introduce BAC, a training-free method that reuses transformer features during robot action denoising. Its scheduler selects refresh times for individual blocks, while a union algorithm coordinates updates to limit propagated errors. On a soft-bag pick-and-release task, BAC reaches 45.1 Hz at 63% success or 39.2 Hz at 71%; these task-specific results do not guarantee losslessness, especially when the base policy performs poorly.
Key insights
- Feature similarities within Diffusion Policy models exhibit non-uniform temporal dynamics and distinct patterns across different architectural blocks (Self-Attention, Cross-Attention, FFNs), necessitating block-wise adaptive caching.
- Direct application of block-level adaptive schedules can lead to an 'error surge phenomenon,' particularly in FFN blocks, due to error amplification caused by upstream caching errors.
- High episode homogeneity within single tasks allows for efficient, offline pre-computation of block-specific optimal caching schedules, incurring minimal real-time computational cost.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.