Reproduce “Human-like autonomy emerges from self-play and a pinch of human data” 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

Daphne Cornelisse, Julian Hunt, Zixu Zhang, Waël Doulazmi, Kevin Joseph, Jaime Fernández Fisac, Eugene Vinitsky

The authors train a driving policy through self-play, using 30-minute human demonstrations as a behavioral anchor. The policy learns from 20-billion simulated transitions with a penalty that keeps its actions near the anchor. On 10,000 held-out scenarios, it records a 0.7% at-fault collision rate against human-replay vehicles, versus 2.1% for unregularized self-play; this is a simulated proxy, not on-road evidence.

Summary by alphaXiv.

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