Reproduce “One View, One Action: Spatially Generalizable VLAs through Observation-Aligned Experts and Cross Graph Consistency” 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

Jinxi Liu, Ruifeng Zhai, Hao Liu, Guangrun Wang

OVOA-VLA adds camera-frame action targets for each view while retaining executable robot-base targets. It transforms demonstrated states and actions using known camera transformations, requiring no new action annotations. In zero-shot real-world tests under spatial shifts, it averages 58.3% success across three tasks, compared with 28.3% for π₀.₅; this result covers the tested shifts, not universal robustness.

Summary by alphaXiv.

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