Generative Diffusion Language Model for End-to-End Autonomous Driving: Parallel Planning and Explainable Action

Abstract

This talk examines how generative diffusion language models can overcome the latency and unidirectional reasoning limitations of autoregressive vision-language models for end-to-end autonomous driving. It presents ViLaD for parallel, bidirectional planning and real-world deployment, then discusses masked vision-language-action diffusion with discrete action tokenization and geometry-aware embeddings for faster, more precise, and explainable driving decisions.