Reproduce “SuperGlance: Driving at a Glance with a Vision-Language Model” 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
Jieyuan Pei, Xin Yan, Mingwei Xu, Haiyi Zhang, Siyu Zou, Zhangyi Hu, Zhuoxuan Li, Yuguang Yang and 4 more
SuperGlance uses a vision-language model’s hidden representation to score a fixed library of driving trajectories in one forward pass, then combines its 64 most probable candidates. This enables exact reward optimization over the finite set. In AlpaSim, with GPU time charged as delay, it recorded 1.248 km per at-fault incident versus Qwen-Drive-1.0’s 0.317; this is simulation evidence, not real-world driving.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.