Reproduce “NarrativeFlow: Flow-Based Vision-Language-Action Model Using Robot Velocity Fields” 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

Shota Kobayashi, Koki Seno, Daichi Yashima, Komei Sugiura

NarrativeFlow generates language-conditioned image-plane motion using continuous velocity fields rather than sparse point displacements and trains an intermediate representation with descriptions of task-relevant scene changes. On Bridge V2 it achieves an average displacement error of 30.59 pixels versus 50.73 for FLIP. Across three real-robot tasks it averages 55% success versus 42% for Im2Flow2Act, with policies fine-tuned on demonstrations.

Summary by alphaXiv.

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