Reproduce “FastOPD: On-Policy Distillation for Lightweight VLA Deployment” 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
Yoojin Oh, Jeongsol Kim, Yeonwoo Seo, Jangho Park, Seonghyun Jin, Sunwoo Park, Youngmin Kim, Youngjun Jun and 2 more
FastOPD transfers a large robot policy’s behavior to a smaller student using one teacher-supervised state per trajectory and a consistency objective to learn shortcut transitions. On LIBERO, the 451M-parameter student reaches 81.8% success in two inference steps, versus 97.5% for the 3.3B-parameter π0.5 teacher at ten steps. Latency is 66 ms versus 301 ms; real-robot evidence covers one task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.