VLAs (vision-language-action models), which are offspring of LLMs, and their versions that are more suitable to edge devices are being used in self-driving to add common sense to path planners ("don't drive through a police standoff", things like that).