NVIDIA GTC 2025: Cosmos Transfer 7B WFMs, Physical AI Dataset, and Isaac GR00T N1 open humanoid model

AI Impact Summary

NVIDIA unveils Cosmos Transfer, a 7B World Foundation Model with multicontrols plus per-sensor ControlNets, enabling tightly guided, photoreal world scene generation and synthetic data pipelines via Omniverse. The Physical AI Dataset provides 15 TB of diverse robotics trajectories and OpenUSD assets on Hugging Face, accelerating post-training fine-tuning and data-coverage for Cosmos Predict and related models. The Isaac GR00T N1 open foundation model introduces cross-embodiment humanoid reasoning with multimodal inputs, supported by a 2B variant on Hugging Face and dual-system architecture (Vision-Language System 2 and Diffusion Transformer System 1), signaling a shift toward open, extensible robotics capabilities; teams should evaluate these assets for piloting manipulation and planning tasks while considering licensing and integration complexity.