Reproduce “GeoAAC: Geometry-Based Adaptive Action Chunking from Denoising Trajectories in VLA Policies” 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
Xin Chen, Sen Chen, Yujuan Ding, Jian Liu, Guoqing Wang, Wei Ye, Heng Tao Shen, Yi Bin
GeoAAC adapts a flow-based robot policy’s action horizon by measuring how denoising revisions change across prefixes of one proposed action chunk and choosing an execution boundary. It temporally weights these changes and applies a motion-aware lower bound without additional training. On RoboCasa365 it reaches 75.1%, versus 66.4% for Fixed-12, within evaluated flow-based policies and settings.
Key insights
- Intermediate denoising geometry correlates with prefix-level uncertainty, with reported Spearman correlations generally above 0.5 against leave-one-out variance estimates.
- Adaptive horizons reflect the manipulation primitive: alignment and placement receive shorter average chunks, whereas transport, pushing, and turning use longer ones.
- Examining the denoising process can provide a more targeted reliability estimate than comparing multiple final action samples, especially when several action modes are valid.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.