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

arXiv 26.3 2026 75 method, system, application

TLDR

FAR-Drive proposes a frame-level autoregressive video generation framework for closed-loop autonomous driving simulation, achieving state-of-the-art on nuScenes.

Reasoning

The paper introduces a novel training strategy (adaptive reference horizon conditioning and blend-forcing) to address long-horizon consistency and autoregressive degradation, and includes system-level optimizations for low-latency inference. However, evaluation is limited to the nuScenes dataset, and generalization to other domains or real-world deployment is not discussed.

Read-first score

Read-first score 75, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 54.

Methodology quality 25%

100

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline,dataset,evaluation,experiment,metric

Recency 8%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Topical relevance 42%

77.1

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=dataset

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 18.

Keyword Scores

world simulator

interactive world model

video world model

generative world model

world model

world dynamics prediction

model-based reinforcement learning world model

Deep Analysis

Innovations

Frame-level autoregressive video generation framework for closed-loop autonomous driving
Multi-view diffusion transformer with fine-grained structured control for geometrically consistent multi-camera generation
Two-stage training strategy combining adaptive reference horizon conditioning and blend-forcing autoregressive training to improve long-horizon consistency and mitigate iterative degradation
System-level efficiency optimizations for low-latency inference on a single GPU

Methodology

FAR-Drive is a frame-level autoregressive video generation framework that uses a multi-view diffusion transformer with structured control to produce geometrically consistent multi-camera outputs. It employs a two-stage training strategy: adaptive reference horizon conditioning to maintain long-horizon consistency, and blend-forcing autoregressive training to reduce degradation under self-conditioning. The model is trained and evaluated on the nuScenes dataset, with baselines from existing closed-loop autonomous driving simulators, and includes system-level optimizations for sub-second latency.

Key Results

FAR-Drive achieves state-of-the-art performance among existing closed-loop autonomous driving simulation approaches on the nuScenes dataset, while maintaining sub-second latency on a single GPU.

Limitations

Evaluation is limited to the nuScenes dataset, so generalization to other driving environments or sensor configurations is not demonstrated
Long-horizon temporal and cross-view consistency may still degrade over very extended sequences despite the proposed training strategy
Autoregressive degradation under iterative self-conditioning is mitigated but not fully eliminated, especially under extreme or out-of-distribution conditions
Low-latency inference is demonstrated on a single GPU, but scalability to multi-GPU or real-time deployment in complex scenarios is not addressed

Resources

Paper