Reproduce “Self-Improving Video-to-Action Models” 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
Hyunwoo Ryu, Eric Ming Chen, Max Wilcoxson, Vincent Sitzmann
Mirror learns to map visual plans to actions by training a video-to-action model on actions and observed transitions from attempts to follow those plans. In Minecraft, after about 10 million environment steps without expert action labels or task rewards, it reached the iron-pickaxe milestone in 8.6% of matched-protocol episodes, versus 1.7% for Dreamer V3, but did not reach diamonds through self-improvement alone.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.