UNCAP Framework: Natural Language Communication for Cooperative Autonomous Driving

According to Robohub, a new framework called UNCAP proposes using natural language instead of raw sensor data for cooperative autonomous driving. Vehicles discover neighbors via lightweight heartbeats, filter relevant vehicles, exchange natural language observations with calibrated confidence, use conformal prediction and mutual information to reason about trust, and finally generate driving decisions via a vision-language model. On the CARLA simulator with the OPV2V benchmark, UNCAP reduced communication bandwidth by 63%, improved driving score by 31%, lowered decision uncertainty by 61%, and increased safety margins during near-miss events by 4x compared to traditional cooperation schemes. The research notes that sharing raw images performed worse than sharing language, and improvements were consistent across GPT-4o, GPT-4.1, GPT-4o-mini, and GPT-5. This information comes from the original source and was organized by AI.