DriveLM

Why we included this project

DriveLM is worth a look if you're building vision-language models for autonomous driving. It pairs a graph-structured visual question answering dataset built on nuScenes and the CARLA simulator with a baseline VLM agent that can both answer those questions and produce driving decisions, so you get a benchmark and a starting model in one repo. The reasoning is framed as a sequence of perception, prediction, planning, and behavior questions instead of a single end-to-end prediction. Because the project powered an official Autonomous Driving Challenge track, the evaluation setup is well exercised and documented. That makes it a practical place to test how large vision-language models handle structured, scenario-based driving tasks.