NVIDIA unveils Cosmos 3 physical AI agent skills, Alpamayo 2 Super, and Isaac Sim 6.0 at CVPR

At CVPR 2026, NVIDIA unveiled a sweeping set of physical AI agent skills — runtime tools that let researchers stitch together world models, simulators, training and evaluation into a single agent-driven workflow. The skills are powered by NVIDIA Cosmos 3, the open frontier world foundation model NVIDIA introduced earlier in the week, and span autonomous vehicles, robotics and vision AI. They ship alongside a new 32-billion-parameter open driving model called Alpamayo 2 Super and a refreshed NVIDIA Isaac Sim 6.0.

What's new

• Cosmos 3-powered agent skills for AV, robotics, and vision AI. NVIDIA describes Cosmos 3's architecture as a mixture-of-transformers — a reasoning transformer analyzes observations and feeds instructions to a generation tower that produces physically grounded virtual worlds.
• Autonomous-vehicle stack: Neural Reconstruction skills, NVIDIA Omniverse NuRec, InstantNuRec, Harmonizer and HiGS for scene reconstruction; AlpaGym, an open-source closed-loop RL framework that scales across thousands of GPUs; OmniDreams, an action-conditioned generative world model; and Alpamayo 2 Super, an open 32-billion-parameter reasoning vision language action (VLA) model targeting level-4 development.
• Vision-AI stack: new NVIDIA Metropolis skills for synthetic visual scenario generation and pseudo-labeling, a Defect Image Generation skill for visual inspection, and the Metropolis VSS Blueprint plus TAO and Video Augmentation skills for video AI.
• Robotics stack: NVIDIA Isaac Sim 6.0 with agent-friendly skills and connectors, Isaac Lab agentic workflows for sim-to-sim and sim-to-real transfer, and Cosmos-H-Surgical-Simulator for surgical-robotics data generation.

Context

This is the second pillar of NVIDIA's CVPR-week physical AI push. Cosmos 3 itself dropped earlier in the week as what NVIDIA calls the world's first full omnimodel unifying vision reasoning, world and action generation, and leads the open public leaderboards for physical AI. NVIDIA also says its broader NVIDIA Physical AI Dataset has now surpassed 15 million downloads on Hugging Face. The new agent-skills layer plus Alpamayo 2 Super sit on top of an existing NVIDIA stack — Omniverse, Isaac Sim, OSMO orchestration — that NVIDIA has been building out for several years.

Why it matters

For research labs, the operational story matters more than any individual model. Scene reconstruction, scenario generation, policy training, evaluation and the orchestration glue between them have historically been four fragmented tools; the new agent-skills surface turns that into a single agent-callable workflow. For autonomous-vehicle developers, Alpamayo 2 Super is notable for its scale — 32 billion parameters — and for being openly released; it is one of the most capable open-source VLA driving models published to date, and a credible alternative to closed proprietary stacks. And for NVIDIA, the strategic effect of pairing Cosmos 3 with an agent-orchestration layer is to deepen the company's position as the de-facto research platform for physical AI heading into the second half of 2026.