NVIDIA GTC 2025 Physical AI Releases: Cosmos Transfer, Physical AI Dataset, and Isaac GR00T N1

NVIDIA has unveiled three major open-source releases at GTC 2025 to accelerate the development of physical AI: the Cosmos Transfer world foundation model, a comprehensive Physical AI Dataset, and the Isaac GR00T N1 humanoid reasoning model. These tools provide developers with the capabilities to generate high-fidelity synthetic environments, access commercial-grade robotics data, and implement generalized reasoning for humanoid robots.

Cosmos Transfer: Controllable World Foundation Model

Cosmos Transfer is a 7-billion parameter world foundation model (WFM) designed to generate photorealistic video sequences with precise control over layout, object placement, and motion. It enables the creation of high-fidelity virtual world scenes from structural inputs, which is critical for scaling synthetic data generation for autonomous vehicles and robotics via the NVIDIA Omniverse platform.

Technical Implementation and Control Mechanisms

Cosmos Transfer utilizes a multicontrol architecture where individual ControlNets are trained separately for each sensor modality. During inference, developers can guide the output using several structured visual or geometric data types, including:

  • Spatial and Geometric Inputs: 3D bounding box maps, depth maps, and LiDAR scans.
  • Motion and Layout Inputs: Trajectory maps, segmentation maps, edge maps, and human motion keypoints.
  • Environmental Data: HD maps.

Control signals from these branches are multiplied by adaptive spatiotemporal control maps and summed before being integrated into the base model's transformer blocks. This allows developers to either strictly preserve structure and appearance or maintain structure while varying the appearance, such as changing weather conditions or environments.

Open Physical AI Dataset

NVIDIA has released a commercial-grade, pre-validated Physical AI Dataset on Hugging Face to support the post-training of foundation models, such as Cosmos Predict.

The dataset includes:

  • Robotics Trajectories: 15 terabytes of data representing over 320,000 trajectories.
  • 3D Assets: Up to 1,000 Universal Scene Description (OpenUSD) assets, including a specialized SimReady collection.

NVIDIA Isaac GR00T N1: General Humanoid Reasoning

NVIDIA Isaac GR00T N1 is the first open foundation model designed for generalized humanoid robot reasoning and skills. The NVIDIA Isaac GR00T-N1-2B model is a cross-embodiment model that processes multimodal inputs (language and images) to perform manipulation tasks across diverse environments and different robot hardware, such as the 1X Neo and Fourier GR-1.

Dual-System Cognitive Architecture

Isaac GR00T N1 employs a dual-system architecture inspired by human cognition to separate high-level reasoning from low-level execution:

  • System 2 (Vision-Language Model): Based on NVIDIA-Eagle and SmolLM-1.7B, this system handles methodical thinking. It interprets vision and language instructions to reason about the environment and plan necessary actions.
  • System 1 (Diffusion Transformer): This action model translates the high-level plan from System 2 into precise, continuous robot movements and actions.

Capabilities and Training

GR00T N1 demonstrates robust generalization in tasks such as grasping and manipulating objects with one or both arms, transferring items between arms, and executing multi-step tasks that require sustained contextual understanding. These capabilities are targeted at applications in material handling, packaging, and inspection.

The model was trained on a diverse dataset comprising real captured data, internet-scale video data, and synthetic data generated via the NVIDIA Isaac GR00T Blueprint.

Sources