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 enabling efficient reinforcement learning for autonomous driving. It compresses diffusion sampling from 100 steps 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 carries prohibitive costs and safety risks, while existing pixel-level diffusion world models, though enabling safe imagination-based training, suffer from multi-step diffusion inference latency of roughly 2 seconds per frame, preventing high-frequency RL interaction. By operating in latent space and using shortcut forcing to reduce sampling to one step, DreamerAD removes this bottleneck and supports fast, safe, imagination-based RL training for driving. The paper is authored by Pengxuan Yang, Yupeng Zheng, Deheng Qian, Zebin Xing, Qichao Zhang and colleagues, posted to arXiv (2603.24587) on 2026-03-25 in the robotics domain.

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.

Key points

First latent world model for driving RL: DreamerAD operates in latent space rather than pixel space, enabling efficient world-model-based reinforcement learning for autonomous driving.
1-step diffusion sampling: Compresses diffusion sampling from 100 steps down to 1 via shortcut forcing, yielding an 80x speedup.
Visual interpretability preserved: Despite the aggressive compression, the framework maintains visual interpretability of the imagined trajectories.
Motivation: Training RL policies directly on real-world driving data is costly and unsafe; imagination-based training in world models offers a safe alternative, but prior pixel-level diffusion world models are too slow (~2s per frame) for high-frequency RL interaction.