NVIDIA Isaac OS and Isaac Lab 3.0
At GTC Berlin 2026 NVIDIA expanded the Isaac platform in a big way: Isaac Lab 3.0 is generally available, Isaac OS is an open-source Holoscan-based runtime that wires AI, sensors, controls and robot hardware into one path with deterministic real-time behaviour from simulation to real robots, and it ships with 2,000+ simulation-ready assets, the Human 1M dataset and three open-source reference workflows on DGX Station
Tool Interface
Interactive tool will be available soon
Features
- ✓ Isaac Lab 3.0 is now GA: an open-source, agent-ready robot learning framework that scales from a desktop GPU to multi-GPU data centre workflows, with both Newton and NVIDIA PhysX physics engines
- ✓ Isaac OS is an open-source runtime framework built on NVIDIA Holoscan that connects AI inference, sensor input, control loops and robot hardware, so simulation and physical robots share the same deterministic real-time behaviour
- ✓ The SimReady Central Library offers more than 2,000 simulation-ready robotics and factory assets from vendors including FANUC, Franka Robotics, Universal Robots and LG Electronics, all hosted on Hugging Face
- ✓ The Human 1M dataset contains diverse human activities, synchronised three-view video and 3D hand tracking, with the full set open to Cosmos Coalition partners for motion generation and imitation learning
- ✓ Three DGX Station reference workflows: refine GR00T 1.7 models with reinforcement learning, train dexterous manipulation skills in Isaac Lab, and post-train Cosmos world models with ChatGPT Astra for synthetic data
- ✓ The companion Agile ONE SN reference humanoid stands 1.40 m and weighs 40 kg with 31 degrees of freedom, dual WUJI Hand 2 five-finger hands and Jetson Thor; NVIDIA says it cuts setup from weeks to days, with availability planned for early 2027
How to Use
- Start at developer.nvidia.com/isaac and pick up Isaac Lab 3.0 and Isaac OS, checking that your GPUs and drivers meet the requirements
- Pull assets from the SimReady Central Library on Hugging Face and assemble your scene from the existing factory and robot models rather than building from scratch
- Choose a training path for the skill you want: distributed multi-GPU workflows in Isaac Lab, reinforcement learning to refine GR00T 1.7, or a separate environment for dexterous manipulation
- Deploy trained policies to real hardware through Isaac OS, and put safety logic on Halos or your own stack rather than pushing simulation policies straight to production
FAQ
Is Isaac Lab 3.0 available now?
Yes, it is generally available. NVIDIA's 21 October 2026 announcement describes it as an open-source, agent-ready robot learning framework with multi-GPU workflows from desktop to data centre, using physics engines including Newton and NVIDIA PhysX. Source: the Isaac platform release on nvidianews.nvidia.com.
What is Isaac OS, and how is it different from Isaac Lab?
Isaac OS is the runtime framework, built on NVIDIA Holoscan, that connects AI, sensors, controls and robot hardware for deterministic real-time execution. Isaac Lab is the learning framework that trains and refines robot policies. One runs skills in production, the other learns them, and together they form the data-to-deployment path NVIDIA describes.
What are the SimReady Central Library and Human 1M?
SimReady Central Library is more than 2,000 simulation-ready robotics and factory assets from vendors including Comau, FANUC, ForgeCAD, Franka Robotics, LG Electronics, Robotiq, Standard Bots, Universal Robots and Vertiv. Human 1M is a dataset of diverse human activities with synchronised three-view video and 3D hand tracking. Both are on Hugging Face, with the full Human 1M available to Cosmos Coalition partners.
What do the three DGX Station workflows actually do?
Per NVIDIA: refine GR00T 1.7 models with reinforcement learning, train dexterous manipulation skills in Isaac Lab, and post-train NVIDIA Cosmos world models with ChatGPT Astra for synthetic data generation. All three are open source.
When can I get the Agile ONE SN reference humanoid?
Availability is planned for early 2027. Specs are 1.40 m tall, 40 kg, 31 degrees of freedom and a 3 kg payload per arm, with dual WUJI Hand 2 five-finger hands at 20 degrees of freedom each with 1 kHz force control and tactile sensing. Onboard Jetson Thor provides 2,070 FP4 teraflops and 128 GB of unified memory, and NVIDIA plans Halos safety integration.