Cosmos-Drive-Dreams: Scalable Synthetic Driving Data Generation with World Foundation Models

arXiv 25.6 2025 62.4 method, system, application

TLDR

Introduces Cosmos-Drive-Dreams, a synthetic data generation pipeline using world foundation models to create challenging driving scenarios for improving AV perception and policy learning.

Reasoning

Strengths include scalable synthetic data generation, open-sourcing of models and data, and demonstrated improvements on multiple downstream tasks. Weaknesses are limited details on controllability and evaluation metrics, and potential domain gap between synthetic and real data not addressed.

Read-first score

Read-first score 62.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 44.

Recency 8%

86.7

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

Methodology quality 25%

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

Topical relevance 42%

62.9

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,open source

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 142.

Keyword Scores

world model

generative world model

video world model

world simulator

world dynamics prediction

model-based reinforcement learning world model

interactive world model

Deep Analysis

Innovations

Introduction of Cosmos-Drive-Dreams, a synthetic data generation pipeline specifically designed to generate challenging driving scenarios for autonomous vehicle systems.
Development of Cosmos-Drive, a suite of models specialized from the NVIDIA Cosmos world foundation model for the driving domain, enabling controllable, high-fidelity, multi-view, and spatiotemporally consistent driving video generation.
Open-sourcing of the pipeline toolkit, dataset, and model weights through NVIDIA's Cosmos platform.

Methodology

The methodology involves a synthetic data generation pipeline powered by Cosmos-Drive models, which are derived from the NVIDIA Cosmos world foundation model and fine-tuned for driving domain tasks. The pipeline generates controllable, high-fidelity, multi-view, and spatiotemporally consistent driving videos to scale the quantity and diversity of driving datasets with challenging scenarios.

Key Results

Experimental results demonstrate that the generated synthetic data helps mitigate long-tail distribution problems and enhances generalization in downstream tasks including 3D lane detection, 3D object detection, and driving policy learning.

Resources

Paper Project