Reproduce “FastHarness: Experience-Driven Review Scheduling for Vision-Language-Action Policies” 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
Tao Lin, Yuxin Du, Haoxiao Wang, Zile Zhong, Bo Zhao
FastHarness schedules GPT-6 Astra reviews around a fixed π0.5 robot policy, using remembered stages and outcomes to reduce routine checks and continuously review detected-fault recovery. Across four simulated RoboDojo tasks, recorded tokens fell an average 57.2% and episode time fell from 85.5 to 52.0 minutes between the first evaluated episode and final-five mean; this comparison does not isolate scheduling causally.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.