Physical‑AI stack updates
NVIDIA rolled out a suite of 'physical AI' tools for teaching machines to sense and act in simulated and real worlds — including Isaac GR00T open models, Cosmos world models, the Newton 1.0 physics engine, plus Isaac Sim 6.0 and Isaac Lab 3.0. The announcements also pointed to partner projects using those tools for autonomous humanoids and surgical arms, showing the company expanding its robotics and simulation ecosystem. ( )
Robotics developers train machines in virtual worlds before they trust them in factories, hospitals, or warehouses, and NVIDIA just expanded that software stack with new models and simulators announced on March 16. (nvidianews.nvidia.com) NVIDIA said the release includes new Isaac GR00T and Cosmos open models, plus Isaac Sim 6.0 and Isaac Lab 3.0 for building, testing, and training robots. The company paired those launches with partner announcements from ABB Robotics, Agility Robotics, CMR Surgical, Figure, Medtronic, Universal Robots, and Yaskawa. (investor.nvidia.com) The basic problem is the “simulation-to-reality” gap: a robot that works in software can fail when friction, lighting, or object contact changes in the real world. NVIDIA’s Isaac Sim documentation describes the product as a robotics simulation and synthetic-data system, while Isaac Lab is its open-source framework for training robot policies at scale. (docs.isaacsim.omniverse.nvidia.com) (developer.nvidia.com) Cosmos is NVIDIA’s “world foundation model” family, which generates predictive video worlds from text, images, or video and can turn simulated scenes into more photoreal outputs for robot training. NVIDIA says those models can be paired with Isaac Sim to create more varied synthetic data and edge cases before a robot is deployed. (nvidia.com) (developer.nvidia.com) GR00T is NVIDIA’s robot-model effort, aimed at giving humanoid and other general-purpose robots reusable perception, reasoning, and action skills. NVIDIA’s developer site describes it as a platform for general-purpose robot foundation models, and the company has been open-sourcing parts of that work since at least March 2025. (developer.nvidia.com) (nvidianews.nvidia.com) Newton is the physics layer underneath that stack, meant to model contact-rich tasks such as grasping, pushing, and walking. NVIDIA said in 2025 that Newton was being developed with Google DeepMind and Disney Research, and in its March 2026 update said Newton 1.0 adds locomotion and manipulation features with integration points for Isaac Lab, Isaac Sim, MuJoCo, and OpenUSD. (developer.nvidia.com 1) (developer.nvidia.com 2) Isaac Sim 6.0 is not yet broadly packaged as a standard download. NVIDIA’s release notes label it an “Early Developer Release” that currently requires building from source, with general-availability binaries and containers still to come. (docs.isaacsim.omniverse.nvidia.com 1) (docs.isaacsim.omniverse.nvidia.com 2) NVIDIA has been building toward this stack in stages. At the Consumer Electronics Show in January 2026, it released new physical-artificial-intelligence models, Isaac Lab-Arena for evaluation, and OSMO workflow software; at the Conference on Robot Learning in September 2025, it said Newton had landed in Isaac Lab alongside the GR00T N1.6 model. (investor.nvidia.com 1) (investor.nvidia.com 2) The partner list shows where NVIDIA wants this software used: humanoids, industrial arms, and surgical systems that need both vision and precise motion. The company’s March 2026 announcement framed the push as “production-scale physical AI,” with software meant to move from synthetic training data to real machines. (nvidianews.nvidia.com) (investor.nvidia.com)
