Nvidia pushes 'physical AI' stack
Nvidia used GTC 2026 to announce a full-stack push for ‘physical AI’—new foundation models (Cosmos, OSMO), digital twins, and operator tooling aimed at closing sim-to-real gaps for robots. The company framed the move as more than chips: it wants to own data curation, synthetic-data generation, and orchestration for embodied AI deployments. (blogs.nvidia.com)
Cosmos ships in 2B and 14B parameter flavors and is marketed to produce up to 30‑second predictive video “worlds” from text, image or video inputs for closed‑loop simulation and edge‑case generation. (nvidia.com) The Cosmos‑Predict2 codebase documents video‑to‑world and action‑conditioned generation pipelines and offers model variants (including 720p/16fps presets) plus transfer/post‑training hooks for robot‑centric datasets. (github.com) OSMO was published as an open‑source, cloud‑native orchestration framework that lets teams describe entire sim‑to‑real pipelines in declarative YAML and automates scheduling across heterogeneous compute and hardware‑in‑the‑loop stages. ( ) NVIDIA said ABB, FANUC, KUKA and Yaskawa are integrating Omniverse libraries and Isaac simulation into their toolchains — a group whose combined install base the company cites at over two million industrial robots. ( ) NVIDIA’s Isaac GR00T N1.6 is available on GitHub as an open robot foundation model for vision‑language‑action workflows, and NVIDIA previewed a GR00T N2 architecture at GTC claiming more than a two‑fold improvement on novel tasks versus prior VLA baselines. ( ) On autonomous vehicles, BYD, Geely, Isuzu and Nissan announced they will build Level‑4 programs on NVIDIA DRIVE Hyperion, and NVIDIA said a full‑stack robotaxi deployment with Uber will span 28 cities by 2028 with initial launches in Los Angeles and San Francisco in H1 2027. (investor.nvidia.com) NVIDIA also published an Omniverse DSX Blueprint and a Physical AI Data Factory blueprint at GTC to standardize digital‑twin simulations across thermals, rack power, networks and mechanical systems for factory‑scale pre‑deployment validation. ( )
