Abstract
Robust and generalizable autonomous driving (AD) demands intelligent, human-like, and safe decision-making. While modern end-to-end systems rely on imitation learning (IL) from vast expert data, recent work shows the power of (self-play) reinforcement learning (RL)—alone or combined with IL—to discover diverse skills and achieve robust out-of-distribution performance. Meanwhile, generative world models and Vision-Language-Action (VLA) models enable controllable, realistic closed-loop simulation, which together with GPU-accelerated multi-agent training facilitates efficient sim-to-real transfer. This workshop brings together key researchers to discuss these developments and challenge assumptions on the path toward robust autonomy.
