Nvidia launches Cosmos 3 world model, Alpamayo 2 Super driving model, and Isaac GR00T humanoid robot reference platform at GTC Taipei

Nvidia released three systems for robotics and autonomous vehicles: Cosmos 3 (an open omnimodel processing text, images, video, audio, and action data), Alpamayo 2 Super (a 32 billion parameter model for Level 4 autonomous driving with 75 degrees of freedom), and Isaac GR00T Reference Humanoid Robot (based on Unitree H2 Plus chassis with Jetson AGX Thor T5000 featuring 2,070 FP4 teraflops). Cosmos 3 offers three variants—Super for best quality, Nano for fast inference, and an upcoming Edge model for embedded systems—available under the OpenMDW-1.1 license.

Cosmos 3 uses a mixture-of-transformers architecture where one reasoning transformer analyzes a scene and a second generation transformer produces videos, descriptions, or motion trajectories. Training data included billions of examples spanning text, images, video, audio, and action data. Alpamayo 2 Super outputs meta-actions like 'lane change,' 'stop,' or 'yield' alongside trajectory decisions and includes a 'chain of causation' reasoning chain intended for safety documentation.

Alpamayo 2 Super is positioned as a teacher model for manufacturers to distill into smaller models running on vehicle-grade chips. Nvidia also released AlpaGym (an open-source framework for closed-loop reinforcement learning in simulation) and OmniDreams (a generative model for rare traffic scenarios). Code and weights are expected on GitHub and Hugging Face this summer. Unitree plans to offer the humanoid robot hardware by late 2026; research partners include Ai2, ETH Zurich, Stanford Robotics Center, and UC San Diego ARC Lab.