OpenDriveLab/DriveLM

Constructs Graph Visual Question Answering datasets (DriveLM-Data) on nuScenes and CARLA that connect perception, prediction, and planning tasks through human-written reasoning chains. The DriveLM-Agent baseline leverages vision-language models for joint end-to-end driving and structured scene reasoning, enabling explainable decision-making via language-guided planning. Provides complete challenge infrastructure with standardized evaluation metrics, including GPT-based scoring and closed-loop simulation support in CARLA.

1,263 stars. No commits in the last 6 months.