At GTC Berlin, NVIDIA expanded its Isaac robotics platform. Agile Robots announced Agile ONE SN, a reference humanoid 1.40 meters tall and 40 kilograms, with 31 degrees of freedom and a 3-kilogram payload per arm, dual five-finger hands at 20 degrees of freedom each with 1 kHz force control and tactile sensing, and Jetson Thor onboard compute rated at 2,070 FP4 teraflops with 128 GB of unified memory. It ships preconfigured with Isaac GR00T, is planned to integrate the Halos safety stack, is scheduled for early 2027 and is described as drawing interest from Stanford and ETH Zurich. NVIDIA also released Isaac OS, an open-source runtime built on Holoscan that connects AI, sensors, controls and robot hardware for deterministic real-time operation; Isaac Lab 3.0 as generally available, an open-source learning framework with multi-GPU workflows and physics engines including Newton and PhysX; and a SimReady Central Library of more than 2,000 simulation-ready robotics and factory assets plus a Human 1M dataset with synchronized three-view video and 3D hand tracking, both on Hugging Face, with the full dataset reserved for Cosmos Coalition partners. Three open reference workflows for DGX Station cover refining GR00T 1.7 with reinforcement learning, training dexterous manipulation in Isaac Lab, and post-training Cosmos world models for synthetic data using ChatGPT Astra.
Our takeThe shift is packaging: simulation, training and deployment tools that teams used to stitch together are being collapsed into one runtime plus a reference body, which is where integration weeks usually go. None of the numbers settles the hard question — whether policies trained in Isaac Lab transfer to a 31-degree-of-freedom platform outside it. The Human 1M dataset is the more consequential release, because hand-tracking data at that scale is the bottleneck dexterous policies run into.