Reproduce “Action Jets: Learning the Local Shape of the Action-conditioned Data Manifold” 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
Anton Arapin, Chahyon Ku, Bernadette Bucher
Action Jets measure local action-to-image changes once and reuse a third-order approximation to score candidate robot actions without rerunning the source generator. A geometry encoder predicts generator-probed directions, while a correction head fits curvature from recorded transitions. In a four-step, same-hardware test, scoring 16 candidates took 54 ms versus WorldGym’s 17,501 ms; accuracy advantages were strongest at short horizons.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.