HorizonDrive: Self-Corrective Autoregressive World Model for Long-horizon Driving Simulation

Conglang Zhang^1,*, Yifan Zhan^2,*, Qingjie Wang³, Zhanpeng Ouyang³, Yu Li⁴, Zihao Yang⁵, Xiaoyang Guo⁶, Weiqiang Ren³, Qian Zhang³, Zhen Dong¹, Yinqiang Zheng², Wei Yin^3,‡, Zhengqing Chen^3,†

¹ Wuhan University   ² The University of Tokyo   ³ Horizon Robotics   ⁴ Tsinghua University   ⁵ University of Science and Technology of China   ⁶ The Chinese University of Hong Kong

* Equal contribution   ‡ Project lead   † Corresponding author

🌍 Overview

HorizonDrive is an anti-drifting training-and-distillation framework for minute-scale autoregressive driving simulation. Through self-corrective teacher training and teacher rollout long-horizon distillation, HorizonDrive enables minute-scale, action-controllable autoregressive video generation of complex driving scenarios on a single GPU, and supports closed-loop interactive simulation.

✨ Key Features

• Controllable driving scene generation.
• Stable minute-scale autoregressive rollout.
• Interactive AR rollout for closed-loop driving simulation.
• Generalizable across diverse driving scenes and scenarios.
• No reliance on explicit 3D representations.

🧪 Abstract

Closed-loop driving simulation requires real-time interaction beyond short offline clips, pushing current driving world models toward autoregressive (AR) rollout. Existing AR distillation approaches typically rely on frame sinks or student-side degradation training. The former transfers poorly to driving due to fast ego-motion and rapid scene changes, while the latter remains bounded by the teacher’s single-pass output length and thus provides only a limited supervision horizon. A natural question is: can the teacher itself be extended via AR rollout to provide unbounded-horizon supervision at bounded memory cost? The key difficulty is that a standard teacher drifts under its own predictions, contaminating the supervision it provides. Our key insight is to make the teacher rollout-capable, ensuring reliable supervision from its own AR rollouts. This is instantiated as HorizonDrive, an anti-drifting training-and-distillation framework for AR driving simulation. First, scheduled rollout recovery (SRR) trains the base model to reconstruct ground-truth future clips from prediction-corrupted histories, yielding a teacher that remains stable across long AR rollouts. Second, the rollout-capable teacher is extended via AR rollout, providing long-horizon distribution-matching supervision under bounded memory, while a short-window student aligns to it with teacher rollout DMD (TRD) for efficient real-time deployment. HorizonDrive natively supports minute-scale AR rollout under bounded memory; on nuScenes, HorizonDrive reduces FID by 52% and FVD by 37%, and lowers ARE and DTW by 21% and 9% relative to the strongest long-horizon streaming baselines, while remaining competitive with single-pass driving video generators.