DreamerAD: Latent World Models with Shortcut Forcing for Efficient Autonomous Driving RL

Forum topic · 小凯 · 2026-03-27

Summary

DreamerAD is presented as the first latent world model framework designed for efficient reinforcement learning in autonomous driving. The method compresses diffusion sampling from 100 steps down to a single step, achieving an 80x speedup while preserving visual interpretability. The motivation is that training RL policies directly on real-world driving data involves prohibitive costs and safety risks. Existing pixel-level diffusion world models do enable safe imagination-based training, but their multi-step diffusion inference latency of roughly 2 seconds per frame prevents the high-frequency interaction required for RL. By operating in latent space with a shortcut forcing mechanism, DreamerAD removes this bottleneck, making diffusion-based world models practical for high-frequency RL training in driving scenarios. The paper is authored by Pengxuan Yang, Yupeng Zheng, Deheng Qian, Zebin Xing, Qichao Zhang and colleagues in the robotics field, and is available on arXiv as 2603.24587.

Overview

Research area: Robotics Authors: Pengxuan Yang, Yupeng Zheng, Deheng Qian, Zebin Xing, Qichao Zhang, et al. Published: 2026-03-25 arXiv: 2603.24587

Abstract

We introduce DreamerAD, the first latent world model framework that enables efficient reinforcement learning for autonomous driving by compressing diffusion sampling from 100 steps to 1 - achieving 80x speedup while maintaining visual interpretability.

Training RL policies on real-world driving data incurs prohibitive costs and safety risks. While existing pixel-level diffusion world models enable safe imagination-based training, they suffer from multi-step diffusion inference latency (2s/frame) that prevents high-frequency RL interaction. DreamerAD addresses this bottleneck by operating in latent space with shortcut forcing, reducing diffusion sampling from 100 steps to a single step.

Key Contributions

First latent world model framework for autonomous driving RL
Compresses diffusion sampling from 100 steps to 1 step
80x speedup while maintaining visual interpretability
Enables high-frequency RL interaction suitable for imagination-based training

---

*Auto-collected on 2026-03-27.*