Cosmos World Foundation Model Platform for Physical AI

By N. Agarwal, Arslan Ali, M. Bala, Yogesh Balaji, Erik Barker, Tiffany Cai, Prithvijit Chattopadhyay, Yongxin Chen, Yin Cui, Yifan Ding and 67 more. arXiv.org, 2025

753 citations recorded by the source corpus.

Physical AI needs to be trained digitally first. It needs a digital twin of itself, the policy model, and a digital twin of the world, the world model. In this paper, we present the Cosmos World Foundation Model Platform to help developers build customized world models for their Physical AI setups. We position a world foundation model as a general-purpose world model that can be fine-tuned into customized world models for downstream applications. Our platform covers a video curation pipeline, pre-trained world foundation models, examples of post-training of pre-trained world foundation models, and video tokenizers. To help Physical AI builders solve the most critical problems of our society, we make Cosmos open-source and our models open-weight with permissive licenses available via https://github.com/nvidia-cosmos/cosmos-predict1.

Subfields: Computer Networks and Communications

View Paper: https://www.semanticscholar.org/paper/b95baf93d1cfb406b50aa370ce5ba18788b5891d

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Networks and Communications

Citations in the corpus

Gen3C: 3D-Informed World-Consistent Video Generation with Precise Camera Control (2025, positive): “For Kubric, we render each dynamic scene from V ∈ [1 , 4] camera trajectories to generate multi-view dynamic videos for creating the 3D cache.”
GR00T N1: An Open Foundation Model for Generalist Humanoid Robots (2025, positive): “In our work, we leverage the recent advancements in video generative models (Agarwal et al., 2025; Wan Team, 2025) to create entire neural trajectories, at a scale that has never been explored before: ∼ 300k neural trajectories which amounts to 827 hours of robot trajectories.”
Aligning Cyber Space With Physical World: A Comprehensive Survey on Embodied AI (2024, neutral): “Thisapproachcansigniﬁcantlyenhancetheaccuracy andfeasibilityoftaskplanning,betteraddressingthechallenges ofreal-worldenvironments.”
Qwen-Image Technical Report (2025, neutral): “Notably, FLUX-VAE (BlackForest, 2024), Cosmos-CI-VAE (Agarwal et al., 2025), and SD-3.”
V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning (2025, positive): “The second baseline we compare with is based on the Cosmos video generation model (Agarwal et al., 2025).”