At GTC Berlin on October 21, 2026, NVIDIA announced a major expansion of its robotics development platform, Isaac. The company added a new research humanoid, the "Agile ONE SN," and made "Isaac Lab 3.0," a framework for training robot motion, generally available. The aim is to provide a development environment that covers everything from training to running on a real robot, easing the burden on researchers who would otherwise have to source and combine hardware and software on their own. However, Agile ONE SN is not expected to ship until early 2027, and work remains to be done on areas such as integrating safety features. In assessing this announcement, it is important to distinguish between the development tools that are already available and the features that will be built out ahead of real-world deployment.
Agile ONE SN: A research humanoid that integrates the robot body and development platform
Agile ONE SN, announced by Agile Robots, is a research humanoid that comes with "Isaac GR00T," NVIDIA's collection of foundation models and development tools for robots, preinstalled, and uses "Jetson Thor" as its onboard computer. It is designed as a reference platform that integrates not only the robot itself, but also the hand mechanisms, the computing platform, and the software needed to move from training to deployment on real hardware. The goal is to reduce the need for researchers to assemble individual components and software from scratch, making it easier to start developing robot motion and testing it on a physical machine.
According to NVIDIA's official announcement, the body of Agile ONE SN is based on the "Agile ONE S," which stands 1.40 m tall and has 31 degrees of freedom. Each arm can handle objects weighing up to 3 kg, and both hands are equipped with the five-fingered "WUJI Hand 2." Each hand has 20 degrees of freedom, and tactile sensors and 1 kHz force control allow it to perform fine movements such as grasping and manipulating objects. Degrees of freedom refers to the number of directions in which joints and other parts can move independently, and is an important measure of how complex a posture or motion a robot can achieve.
To perceive its surroundings, the robot has cameras in its head and wrists, along with LiDAR for measuring distance and an inertial measurement unit for detecting the body's posture and movement. Jetson Thor, which handles onboard computing, offers 2,070 TFLOPS of AI performance at FP4 and 128 GB of unified memory, and is used for processing sensor data, AI-based decision-making, and real-time control. Note, however, that FP4 indicates computing performance using a low-precision number format, not a direct measure of the success rate or precision of actual object manipulation. A robot's performance must be validated with the AI models, sensors, and control software working together.
This is not the first time NVIDIA has announced such a reference design. On June 1, 2026, the company announced a humanoid reference design for academic research that combines Jetson Thor and Isaac GR00T with the Unitree H2 Plus and Sharpa Wave robotic hands. Agile ONE SN extends this effort to Agile Robots' hardware, while strengthening integration with new training environments and execution platforms.
NVIDIA says that adopting a preintegrated hardware-and-software configuration like Agile ONE SN can cut initial setup, which previously took weeks, down to days. This is the company's own projection, however, not the result of verification by a third party under standardized conditions. For research institutions, what matters is not only how fast a robot learns, but whether the time spent connecting equipment and tuning software before experiments can begin is reduced. If the initial setup burden shrinks, that time could be redirected to training and validating motions.
Connecting training to real-robot control with Isaac Lab 3.0 and Isaac OS
The announcement also included significant updates to the software that handles robot training and execution. One is the open-source training framework "Isaac Lab 3.0," now generally available. It is an environment for training and refining "policies," which determine how a robot moves in response to its surroundings, and it supports multi-GPU training and physics engines such as Newton and NVIDIA PhysX. It can scale from desktops to data centers and is intended for repeatedly training robots in virtual spaces on tasks such as object manipulation and posture control.
Meanwhile, the newly announced "Isaac OS" is an open-source runtime for running trained AI models on actual robots. Built on NVIDIA Holoscan, it connects AI inference, sensor input, control processing, and the robot's hardware. To operate a robot reliably, it is not enough for the AI to decide on the right action; the system must also process camera and sensor information and control motors and other components at set timings. With an eye toward the transition from simulation to real machines, Isaac OS aims to execute these processes with predictable timing.
In other words, while Isaac Lab 3.0 is the environment in which robot motions are learned, Isaac OS connects the results of that training to the sensors and control mechanisms of a real robot. Combining the two is intended to simplify the process of transferring motions learned in simulation to a physical robot. However, NVIDIA did not present specific latency figures for Isaac OS in this announcement, nor validation results guaranteeing operation on any robot. In actual deployments, compatibility with the target machine, sensors, and control mechanisms will need to be confirmed.
NVIDIA is also releasing three open-source reference workflows for DGX Station. They cover refining Isaac GR00T 1.7 through reinforcement learning, training dexterous object manipulation in Isaac Lab, and using ChatGPT Astra to further train NVIDIA Cosmos world models for generating synthetic data. A world model is an AI model that learns the state of an environment and how it changes; here it generates data used to train robots. The ChatGPT Astra integration concerns the training-data generation step and does not mean that ChatGPT Astra directly controls a humanoid.
Combining these technologies should make it easier for researchers to carry out data preparation, simulation, training, and real-robot execution on a common development platform. Compared with adopting each tool independently, this could reduce the effort needed to set up environments and connect functions. Still, real robots face problems that simulation alone cannot fully verify, such as sensor accuracy, control latency, and safety, so an integrated development platform does not by itself guarantee practical deployment.
Isaac Lab is generally available; Agile ONE SN is due in early 2027
This expansion of the Isaac platform mixes software that is already available with hardware and features that will arrive later. Isaac Lab 3.0 is now generally available, whereas Agile ONE SN is scheduled to ship in early 2027, and integrating its safety features will require further development. The terms of use for training data and simulation assets also differ from item to item.
Based on NVIDIA's October 21 announcement, the role and availability of each component can be summarized as follows.
| Development stage | Main component | Role and availability |
|---|---|---|
| Building virtual environments | SimReady Central Library | Includes more than 2,000 robot and factory simulation assets. Available on Hugging Face |
| Human motion data | Human 1M Dataset | Contains synchronized three-view video and 3D hand-tracking data. Access to the full data is limited to Cosmos Coalition partners |
| Robot training | Isaac Lab 3.0 | Supports policy training using multiple GPUs and physics engines. Generally available |
| Real-robot execution | Isaac OS | Announced as an open-source runtime connecting AI, sensors, and control mechanisms |
| Humanoid research | Agile ONE SN | Integrates Isaac GR00T and Jetson Thor. Scheduled for early 2027 |
| Safety integration | Agile ONE SN and NVIDIA Halos | Integration of on-board safety features with Halos is planned |
There is a time lag between the release of training environments and simulation data and the availability of an actual research robot. For example, the "SimReady Central Library," which contains more than 2,000 simulation assets, is available on Hugging Face, but access to the full "Human 1M Dataset" is limited to Cosmos Coalition partners. Human 1M includes a variety of human motions, with synchronized three-view video and 3D hand-tracking information, but the existence of a public page does not mean anyone can use the entire dataset. Nor can the number of videos it contains be inferred from the "1M" in its name.
On the safety side, integration with "Halos," which NVIDIA is developing, is planned. Halos is a safety platform for robots and autonomous driving systems built on NVIDIA IGX Thor, and for Agile ONE SN the idea is to combine it with the robot's onboard safety features. NVIDIA also highlighted SICK's evaluation of Halos for improving the dynamic safety of industrial robots, as well as work with companies involved in certification. However, companies taking part in developing and evaluating safety technology is not the same as individual robots having obtained safety certification and being ready for use in real-world settings.
Likewise, it has not been announced that safety integration for Agile ONE SN is complete. The safety measures required vary with the operating environment, and beyond the accuracy of learned motions, risks such as contact with people and surrounding equipment must be evaluated. Even when adopting a research reference platform, it is important to check how safety features have been implemented and what the verification results are.
Adoption grows among European companies, but real-robot validation is the next focus
NVIDIA's Isaac platform is being used in robot development beyond humanoids. ABB Robotics has integrated NVIDIA Omniverse libraries into "RobotStudio HyperReality," creating a way to validate industrial robot motions in a virtual environment before the robots are installed. Sereact has trained and further trained "Cortex," a model for handling goods, using Isaac GR00T and Cosmos Transfer, and runs it on Universal Robots industrial arms. These examples show that NVIDIA is aiming not just for a development environment for specific humanoids, but for deploying an AI foundation across a broad range of fields, including industrial robots.
Such efforts could also create new demand for NVIDIA's computing platforms. StockTitan's October 21 analysis views the string of announcements in Berlin from two angles: expanding semiconductor supply capacity and widening the uses for those chips. The view is that if NVIDIA's technology is adopted in robots, autonomous vehicles, and communications networks, demand for the corresponding computers may also grow. But as the outlet points out, the number of additional chips these uses would require has not been disclosed. NVIDIA's expansion of semiconductor supply capacity cannot be directly tied to robotics demand or investment amounts.
As for who might adopt Agile ONE SN, NVIDIA says research institutions such as Stanford University and ETH Zurich have shown interest. It has not announced, however, that these institutions have confirmed purchases or deployments. The selling price was also not disclosed in this announcement, so research institutions will need to check pricing, delivery timing, and support arrangements when procuring the robot.
Looking toward the early-2027 launch of Agile ONE SN, the key question is how well motions learned in simulation can be reproduced on a real machine. Also at stake are whether it can operate stably with safety features integrated, and whether researchers can repeatedly test different motions and control methods. By packaging hardware and software into a shared reference configuration, researchers may be able to redirect time once spent building development environments toward research on object manipulation and whole-body control. Whether NVIDIA's push to unify its robot development platform accelerates practical adoption will depend heavily on future validation results with real machines.