GenieDrive: Towards Physics-Aware Driving World Model with 4D Occupancy Guided Video Generation

arXiv 25.12 2025 66.5 method, application

TLDR

GenieDrive uses 4D occupancy and a VAE with Mutual Control Attention to generate physics-aware driving videos, improving forecasting mIoU by 7.2% at 41 FPS.

Reasoning

The paper introduces a novel framework that leverages 4D occupancy as a physics-informed prior for driving video generation, with efficient compression and end-to-end training. Strengths include clear methodology and strong quantitative results, but the abstract lacks explicit real-world dataset details and broader baseline comparisons.

Read-first score

Read-first score 66.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 55.

Recency 8%

86.7

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

Topical relevance 42%

78.6

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=evaluation,experiment

Reproducibility 25%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 210.

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

Using 4D occupancy as a physics-informed foundation for driving video generation
VAE encoding occupancy into a latent tri-plane representation, reducing latent size to 58% of previous methods
Mutual Control Attention (MCA) to model the influence of control on occupancy evolution
End-to-end joint training of VAE and prediction module for forecasting accuracy
Normalized Multi-View Attention for multi-view video generation guided by 4D occupancy

Methodology

GenieDrive first generates 4D occupancy, which captures high-resolution 3D structures and dynamics, as a physics-informed foundation. A VAE compresses this occupancy into a latent tri-plane representation, and a Mutual Control Attention mechanism models how driving controls affect occupancy evolution. The VAE and prediction module are trained end-to-end. For video generation, a Normalized Multi-View Attention module uses the 4D occupancy guidance to produce multi-view driving videos.

Key Results

The method achieves a 7.2% improvement in forecasting mIoU at 41 FPS with only 3.47M parameters, and a 20.7% reduction in FVD for video generation, enabling highly controllable, multi-view consistent, and physics-aware driving video generation.

Resources

Paper Project