Reproduce “Concertina: Embodying Manipulation and Cognition in One Diffusion Vision-Language Model” 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
Chengcheng Wang, Wei He, Wenhao Li, Hongguang Li, Kai Han, Jianyuan Guo, Chang Xu
Concertina puts continuous action representation and generation inside a shared vision-language model, without a separate action expert or discrete codes. It encodes each timestep’s action vector as one token and generates blocks through masked reconstruction, supervised in physical coordinates. Checkpoint probes attribute cross-time response to the shared backbone. Regular LIBERO success averages 98.15% across four suites versus π₀.₅’s 97.5% in closed-loop simulation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.