STAGE: A Stream-Centric Generative World Model for Long-Horizon Driving-Scene Simulation

IROS 25 2025 59.3 method, application

TLDR

STAGE is an auto-regressive generative world model for long-horizon driving video simulation using hierarchical temporal attention and multi-stage training.

Reasoning

The paper introduces a novel framework with HTFT and multi-stage training to address error accumulation and feature misalignment in long-horizon driving video generation. Strengths include clear methodology and empirical results on Nuscenes; weaknesses are limited discussion of real-world deployment and lack of interactive or RL aspects.

Read-first score

Read-first score 59.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 43.

Recency 8%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Topical relevance 42%

61.4

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

Methodology quality 25%

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

Reproducibility 25%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 214.

Keyword Scores

world model

generative world model

video world model

world simulator

world dynamics prediction

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Hierarchical Temporal Feature Transfer (HTFT) that models temporal and denoising processes separately and transfers denoising features between frames to enhance temporal consistency
Multi-stage training strategy dividing training into three stages for model decoupling and auto-regressive inference simulation to accelerate convergence and reduce error accumulation
Auto-regressive framework with hierarchical feature coordination and multi-phase optimization for sustainable long-horizon video synthesis

Methodology

STAGE is an auto-regressive framework that introduces Hierarchical Temporal Feature Transfer (HTFT) and a multi-stage training strategy. HTFT enhances temporal consistency by separating temporal and denoising processes and transferring denoising features between frames. The multi-stage training divides training into three stages, decoupling the model and simulating auto-regressive inference to accelerate convergence and reduce error accumulation. Experiments are conducted on the Nuscenes dataset.

Key Results

STAGE significantly surpasses existing methods in long-horizon driving video generation, generating 600 frames of high-quality driving videos on the Nuscenes dataset, far exceeding the maximum length achievable by existing methods.

Resources

Paper Project