Reproduce “VICTR: Value-Informed In-Context Learning for Vision-Language-Action Models” 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
Luca Macesanu, Tianluo Zhang, Aditya Patil, Archit Sharma, Kevin Angers, Christopher M. Clark, Chen Feng, Juexiao Zhang
VICTR retrieves demonstration chunks by estimated task progress rather than visual similarity, using IC-VFE with a π₀.₅ in-context policy whose weights remain fixed at deployment. Across eight unseen tasks in one physical environment, it reached 55.5% progress and 23.8% binary success, versus 42.1% and 15.0% for visual-retrieval RICL-π₀.₅. Progress gives partial milestone credit; demonstrations are provided at inference.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.