Cosmos-Transfer1

arXiv 25.3 2025 63.2 method, application

TLDR

Cosmos-Transfer is a conditional world generation model using multiple spatial controls for world-to-world transfer, enabling Sim2Real and real-time generation.

Reasoning

Strengths include a novel adaptive spatial conditioning scheme and demonstrated applications in robotics and autonomous driving, with open-source release. Weaknesses are the lack of detailed evaluation metrics and baselines in the abstract, and limited scope to spatial control inputs.

Read-first score

Read-first score 63.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 39.

Recency 8%

86.7

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

Reproducibility 25%

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

Topical relevance 42%

55.7

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

Field roles

FrontierReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 251.

Keyword Scores

world model

generative world model

world simulator

interactive world model

video world model

world dynamics prediction

model-based reinforcement learning world model

Deep Analysis

Innovations

Conditional world generation model accepting multiple spatial control inputs of various modalities (segmentation, depth, edge)
Adaptive and customizable spatial conditioning scheme allowing different weighting of conditional inputs at different spatial locations
Enables world-to-world transfer use cases including Sim2Real
Inference scaling strategy to achieve real-time world generation on NVIDIA GB200 NVL72 rack

Methodology

The model is a conditional world generation model that takes multiple spatial control inputs (e.g., segmentation, depth, edge) and uses an adaptive spatial conditioning scheme that weights different inputs differently at different spatial locations. The model is evaluated on Physical AI applications such as robotics Sim2Real and autonomous vehicle data enrichment, and an inference scaling strategy is demonstrated for real-time generation on an NVIDIA GB200 NVL72 rack.

Key Results

The model demonstrates highly controllable world generation and is applied to Sim2Real and autonomous vehicle data enrichment. Real-time world generation is achieved using an inference scaling strategy on an NVIDIA GB200 NVL72 rack.

Resources

Paper Code