AI Daily Brief — 17 March 2026

GTC day-two doubles down on physical AI. NVIDIA unveils Cosmos 3 — billed as “the world’s first fully open omnimodel” — natively translating text, image, video and sound into actions in the physical world. The humanoid demo: Isaac GROOT, a Jetson-Thor reference robot with Unitree articulation. The robotaxi pillar: Alpamayo 2 Super, a 32B-parameter open VLA reasoning model.

Top stories

• Cosmos 3. Fully open omnimodel — translates between text, images, video, sound and physical-world actions. via NVIDIA
• Isaac GROOT reference humanoid. Jetson Thor inside, full NVIDIA software stack, articulated body and hands from Unitree. via Automate.org
• Alpamayo 2 Super (32B VLA). An open reasoning vision-language-action model framed as scaffolding for future autonomous-vehicle infrastructure. via NVIDIA
• Physical AI Data Factory Blueprint. Open reference architecture for real + synthetic data pipelines, with FieldAI, Hexagon Robotics, Linker Vision, Milestone Systems, Skild AI, Uber and Teradyne Robotics signing on. via NVIDIA

By the numbers

• Alpamayo 2 Super: 32B open VLA
• Physical AI Data Factory partners: FieldAI · Hexagon Robotics · Linker Vision · Milestone · Skild AI · Uber · Teradyne

Compiled by AI Feed’s editor from publicly reported announcements on 17 March 2026.