DreamerAD: First Latent World Model Achieves 80x Speedup for Autonomous Driving RL

arXiv·medium signal

DreamerAD is the first latent world model framework enabling efficient reinforcement learning for autonomous driving by compressing diffusion sampling from 100 steps to 1 — achieving an 80x speedup while maintaining visual interpretability. This makes world-model-based RL practical for real-time driving scenarios where previous diffusion-based approaches were too slow. Demonstrates that latent compression can dramatically reduce the inference cost of generative world models without sacrificing quality.