Reproduce “Competent Autonomous Driving Through Fine Tuning Vision Language Action Models with Symbolic Traffic Rules” 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
Barkin Dagda, Vinal Asodia, Iman Sharifi, Peng Wei, Saber Fallah
The authors train a driving model to match expert controls while checking actions against traffic rules. They refine it with reinforcement learning, combining rule feedback with a penalty for departing from expert actions. On 3,600 held-out CARLA images, Dr. GRPO reached a 97.59% encoded-rule pass rate, versus 92.93% for the supervised fine-tuned baseline; this offline result covers selected checks, not real-world safety.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.