Reproduce “ADPro: a Test-time Adaptive Diffusion Policy via Manifold-constrained Denoising and Task-aware Initialization for Robotic Manipulation” 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
Zezeng Li, Rui Yang, Ruochen Chen, ZhongXuan Luo, Liming Chen
ADPro is an adaptive diffusion policy framework that enhances robotic manipulation by employing test-time manifold-constrained denoising and task-aware initialization. It improves generalization to novel tasks and environments and boosts sampling efficiency, achieving an average 2.6 percentage point increase in success rate on RLBench and reducing computational time by approximately 25%.
Key insights
- Injecting geometric and task-relevant inductive biases directly into the diffusion inference process significantly improves policy performance and efficiency without requiring model retraining.
- Guiding denoising updates along task-relevant manifolds and constraining the search space with spherical priors stabilizes action generation and minimizes inefficient exploration.
- Initializing the diffusion process with coarse, task-aware estimates from observational data substantially reduces the required denoising steps and improves convergence speed.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.