NVIDIA opens three key robotics tools to democratize physical AI development

NVIDIA is accelerating physical AI development with the release of three groundbreaking open-source tools announced at GTC 2025. This trio—consisting of a new world foundation model with unprecedented control capabilities, a comprehensive physical AI dataset, and the first open model for humanoid robot reasoning—represents a significant step forward in democratizing advanced robotics development. These innovations aim to give developers the resources needed to create more sophisticated autonomous systems capable of understanding and interacting with the physical world.

The big picture: NVIDIA has unveiled three major open-source releases to advance physical AI development: Cosmos Transfer world foundation model, a 15-terabyte Physical AI Dataset, and Isaac GR00T N1, the first open model for general humanoid reasoning.

Key details about Cosmos Transfer: The new 7-billion-parameter world foundation model introduces multicontrol capabilities that allow precise generation of virtual world scenes from structural inputs.

• The model utilizes separately trained ControlNets for each sensor modality used to capture the simulated world, including 3D bounding boxes, trajectory maps, depth maps, and segmentation maps.
• At inference time, developers can guide outputs using various structured visual or geometric data, with control signals from each branch multiplied by adaptive spatiotemporal control maps before being added to the transformer blocks.
• The resulting output consists of photorealistic video sequences with controlled layout, object placement, and motion, allowing developers to either preserve structure and appearance or maintain structure while varying appearance.

Important stats about the Physical AI Dataset: NVIDIA’s open-source dataset on Hugging Face contains 15 terabytes of commercial-grade, pre-validated data specifically designed for physical AI development.

• The collection includes more than 320,000 trajectories for robotics training alongside up to 1,000 Universal Scene Description (OpenUSD) assets, including a SimReady collection.
• This resource provides developers with high-quality training data essential for building physical AI systems without needing to create their own datasets from scratch.

Innovation in humanoid AI: NVIDIA Isaac GR00T N1 represents the world’s first open foundation model specifically designed for generalized humanoid robot reasoning and skills.

• The model accepts multimodal inputs including language and images to perform manipulation tasks across diverse environments.
• Training data for Isaac GR00T N1 combines real captured information, synthetic data generated using components of the NVIDIA Isaac GR00T Blueprint, and internet-scale video data.
• The model features a dual-system architecture inspired by human cognition, with a Vision-Language Model based on NVIDIA-Eagle with SmolLM-1.7B for environmental interpretation, and a Diffusion Transformer that generates continuous actions to control robot movements.

Why this matters: By releasing these tools as open-source resources, NVIDIA is lowering barriers to entry in the rapidly evolving field of physical AI, potentially accelerating innovation in robotics, autonomous systems, and embodied intelligence.