Reproduce “ODEAction: From Predicting Chunks to Evolving Actions” 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

Qishen Wang, Pengxiang Ding, Shifeng Bao, Mingyang Sun, Hengtao Li, Yi Gao, Baohua Yin, Yang Liu and 9 more

ODEAction generates robot actions recursively by learning an action transition field that evolves the current action over time, rather than predicting chunks. It trains on demonstrations, combining progress along the demonstrated path with correction toward it. On EBench with the π⁰·⁵ base policy, success rose from 15.28% to 24.72% against a matched chunk baseline. This shows tested gains, not a runtime advantage.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.