Reproduce “stable-envs-v1: Fast and Reproducible Environments with Controllable Variations” 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
Sami BuGhanem, Francesco Capuano, Ahitagni Das, Richard Gao, Randall Balestriero
The authors introduce stable-envs, a library that advances simulator states and renders images on the training GPU rather than moving stored frames from CPU memory. It supports controlled appearance changes from saved states, with explicit visual factors in 78 of 80 classes. Across four tested environments, LeWM training is faster; on PushT, throughput rises from 4.58 to 6.09 steps per second, while planning success averages 94.0% versus 93.3% over three seeds.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.