Reproduce “Sparse ActionGen: Accelerating Diffusion Policy with Real-time Pruning” 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, Jianbo Zhou, Yuan Meng, Ye Li, Hanyun Cui, Zhi Wang
Sparse ActionGen (SAG) is an adaptive method that prunes redundant computations in Diffusion Policy models for real-time robot manipulation. It achieves up to 4x speedup on various simulation and real-world tasks by dynamically generating pruning schedules based on observations and employing a "one-for-all" activation reusing strategy, maintaining high task success rates.
Key insights
- Optimal computational sparsity patterns in Diffusion Policies are dynamic and vary significantly across different rollout iterations, necessitating environment-aware adaptation rather than static schedules.
- Significant redundancies exist not only within individual blocks or timesteps but also across different Diffusion Transformer blocks, making a global 'one-for-all' activation reuse strategy highly effective.
- An efficient, observation-conditioned pruner can predict optimal sparsity patterns a priori with minimal overhead, enabling real-time adaptation of computational resources.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.