Reproduce “AlphaNIST: Emergent Assembly Skills from Binary Task Rewards” 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
Octi Zhang, Iretiayo Akinola, Rosario Scalise, Mateo Guaman Castro, Patrick Yin, Tobias Widmer, Miles Macklin, Chris Amevor and 7 more
AlphaNIST tests whether one shared policy can learn varied assembly tasks without prescribed strategies. It trains across parts and starts using a binary completion reward alongside a control cost. At 64K environments and 25K iterations, shared sparse-reward training reaches 98.2% success, versus 81.4% for separate policies. These are training results; hardware transfer and assembling all 20 pairs in one episode remain untested.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.