FAR-Drive: Frame-AutoRegressive Video Generation in Closed-Loop Autonomous Driving

TL;DR

FAR-Drive introduces a novel autoregressive video generation framework for autonomous driving simulation, addressing long-term consistency, multi-view generation, and low-latency inference, enabling more reliable and interactive training environments.

Contribution

The paper presents a multi-view diffusion transformer with structured control and a two-stage training strategy to improve closed-loop autonomous driving simulation.

Findings

Achieves state-of-the-art performance on nuScenes dataset

Maintains sub-second latency on a single GPU

Enhances long-horizon consistency and robustness in simulation

Abstract

Despite rapid progress in autonomous driving, reliable training and evaluation of driving systems remain fundamentally constrained by the lack of scalable and interactive simulation environments. Recent generative video models achieve remarkable visual fidelity, yet most operate in open-loop settings and fail to support fine-grained frame-level interaction between agent actions and environment evolution. Building a learning-based closed-loop simulator for autonomous driving poses three major challenges: maintaining long-horizon temporal and cross-view consistency, mitigating autoregressive degradation under iterative self-conditioning, and satisfying low-latency inference constraints. In this work, we propose FAR-Drive, a frame-level autoregressive video generation framework for autonomous driving. We introduce a multi-view diffusion transformer with fine-grained structured control, enabling geometrically consistent multi-camera generation. To address long-horizon consistency and iterative degradation, we design a two-stage training strategy consisting of adaptive reference horizon conditioning and blend-forcing autoregressive training, which progressively improves consistency and robustness under self-conditioning. To meet low-latency interaction requirements, we further integrate system-level efficiency optimizations for inference acceleration. Experiments on the nuScenes dataset demonstrate that our method achieves state-of-the-art performance among existing closed-loop autonomous driving simulation approaches, while maintaining sub-second latency on a single GPU.