NVIDIA Expands Isaac, Releases Robotics Learning Platform and Datasets

NVIDIA announced an expansion of its robotics development platform, NVIDIA Isaac, on October 21, 2026. It released more than 2,000 simulation assets for robotics and factories, as well as the Human 1M Dataset of human motion, and began general availability of the Isaac Lab 3.0 learning framework. The company also added Isaac OS, a runtime environment that connects AI, sensors and control systems with robot hardware.

In NVIDIA's announcement, Isaac Lab 3.0 handles training, while Isaac OS connects AI, sensors and control systems with robot hardware. For robotics developers, the platform offers a way to link training processes and physical-robot runtime environments that have previously been built separately. However, access to the full Human 1M Dataset is limited to Cosmos Coalition partners, making the scope of available data a factor in development.

The focus of this expansion is not the announcement of a specific robot, but the presentation of data, simulation, training and runtime environments as a connected development pipeline. More than 2,000 assets and the Human 1M Dataset provide material for training and evaluation; Isaac Lab 3.0 handles policy training; and Isaac OS connects to sensors and control systems on physical robots. The expansion can be seen as a development platform spanning the process of getting a robot to operate, rather than just a model. Access is not uniform, however. Isaac Lab 3.0 and Isaac OS are described as open source, but access to the full Human 1M Dataset is limited to Cosmos Coalition partners. Broad availability of data is not the same as being able to reproduce training and evaluation under the same conditions using the complete dataset. The scope of data available to developers will affect the platform's practical utility. The inclusion of assets from FANUC and Universal Robots in the SimReady Central Library also creates a way to incorporate equipment from real-world robot manufacturers into simulation environments, but the range of supported assets alone does not necessarily eliminate differences between simulated and real-world behavior. Coverage by this newspaper of AWS unveiling a toolchain for robot development also addressed tools for robotics developers. The specific features and scope of the two offerings have not been confirmed to be the same, but they show that development environments are being built out not only for robot hardware, but also for software and training processes. Meanwhile, the challenge of enabling robots to improve themselves in the physical world, covered in leaderobot.com's report on the paper “Who Improves the Robot?”, will not be resolved by improving training environments alone. How to verify capabilities gained in simulation in real-world settings and feed the results back into retraining remains an open question. Agile ONE SN is a reference example of physical hardware to be combined with Isaac's software and data platform. Its planned availability in early 2027 will be a milestone for assessing how much of the development setup will be usable. However, the announcement does not specify the robot's real-world performance, evaluation methods, dataset licensing and usage terms, or the range of devices supported by Isaac OS. Assessing whether the design linking training to physical deployment can be replicated across different robots and applications will require examining these conditions and operational results.

The SimReady Central Library contains more than 2,000 simulation-ready assets and is available on Hugging Face. The Human 1M Dataset includes a wide range of human activities, synchronized videos from three viewpoints and 3D hand tracking. The dataset is also available on Hugging Face, but access to the full version is limited to Cosmos Coalition partners. Isaac Lab 3.0, now generally available, is an open-source learning framework that supports multi-GPU workflows and physics engines including Newton and NVIDIA PhysX. Isaac OS is an open-source runtime environment that connects AI, sensors, control systems and robot hardware from simulation to physical deployment. The announcement also included Agile Robots' reference humanoid, Agile ONE SN, which stands 1.40 meters tall, weighs 40 kilograms, has 31 degrees of freedom and can carry 3 kilograms per arm. Jetson Thor delivers 2,070 FP4 teraflops of compute performance and has 128GB of unified memory. Availability is planned for early 2027.