Reproduce “HarnessPAI: An Evolving Harness for Physical AI” 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

Xin Wang, Wenhao Wu, Menghao Zhang, Zhi Wang, Kun Shao, Jian Luan, Yang Li, Qing Li and 15 more

HarnessPAI uses fixed executable programs around a frozen action model to ground instructions, plan actions, and check outcomes. Programs can use observations and predefined checks during rollouts but are revised between attempts using execution traces and reusable repair skills. On LIBERO-PRO, π0.5 achieved 96.5% averaged across six suite-by-perturbation cells, versus 34.9% without HarnessPAI; this result covers tested benchmark settings.

Summary by alphaXiv.

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