NVIDIA's GTC Berlin announcements: NVIDIA wants to be a robot platform company
NVIDIA used its GTC Berlin conference on October 20, 2026 to announce a wave of robotics platform updates that represent the company's most aggressive push yet into physical AI infrastructure. The announcements span new silicon, updated foundation models, a revamped safety framework, and a major commercial robot built on NVIDIA's full stack.
Agile Robotics ONE SN: the first full-stack NVIDIA humanoid
Agile Robotics unveiled the ONE SN, a humanoid designed from the ground up around NVIDIA's Jetson Thor compute module. Standing 1.4 meters tall and weighing between 34 and 40 kilograms, the robot carries 31 degrees of freedom in its body and is equipped with the WUJI Hand 2, a 20-DOF hand on each arm with 1kHz force control. The combination gives it dexterous manipulation capabilities that Agile says are suitable for both industrial and service environments.
The ONE SN ships with the Jetson Thor SoC running at 2,070 FP4 TFLOPS with 128 GB of unified memory.
NVIDIA says the compute headroom is intentional: it allows operators to run multiple AI models simultaneously without offloading to the cloud. Agile is targeting a Q1 2027 commercial release and has told partners it expects to produce between 1,000 and 2,000 units during calendar year 2027.
The robot will run Isaac Lab, GR00T, and Halos natively, making it the first commercially available humanoid to ship with the complete NVIDIA robotics software stack pre-integrated.
Isaac Lab 3.0 goes generally available
NVIDIA announced Isaac Lab 3.0 as generally available at GTC Berlin. The simulation environment, which runs on Omniverse and supports large-scale parallel training, now includes expanded physics fidelity for contact-rich manipulation tasks and tighter integration with the GR00T model family.
The update also ships with a new teleoperation pipeline that allows operators to collect training demonstrations directly inside the simulator, reducing the cost of generating manipulation data.
NVIDIA says Isaac Lab 3.0 has been validated across more than 200 robot configurations contributed by hardware partners.
GR00T N1.7 and N2: better generalization, faster inference
NVIDIA released GR00T N1.7 in early access. The model was trained on 20,854 hours of demonstration data, significantly more than its predecessor, and shows improved generalization across object categories and lighting conditions that were not present in training.
NVIDIA also previewed GR00T N2, which uses a revised architecture that doubles task success rates on the company's internal benchmarks compared to N1.5. The company says N2 will reach general availability by the end of 2026.
Bear Robotics, which makes the KINO-2 food service robot, published performance numbers alongside the GTC announcements. The company says GR00T N1.7 inference on its hardware runs in 60 milliseconds, down from 500 milliseconds with the previous generation: an 8-times speedup. Bear's own proprietary model saw a 4.2-times improvement on the same hardware update.
The H2 Plus reference design, built with Unitree
NVIDIA and Unitree announced a joint reference humanoid at GTC Berlin. The H2 Plus reference design is built on Jetson Thor and is intended to give system integrators a validated starting point for building application-specific humanoids.
NVIDIA is positioning it as a platform rather than a product: companies license the design and adapt it for their use case rather than buying finished units from NVIDIA directly.
Halos safety stack gains new partners
NVIDIA's Halos safety framework, which provides runtime monitoring and intervention for humanoid robots, added 5 new hardware partners at GTC Berlin:
- Agile Robotics
- Agility Robotics
- Dexory
- FORT Robotics
- KION Group
The additions expand Halos from its initial focus on logistics environments to include general-purpose service and warehouse automation.
Halos works by running a parallel safety model alongside the primary task model. If the safety model detects an anomaly, it can override the task model's output before the robot acts. NVIDIA says the framework adds less than 10 milliseconds of latency on Jetson Thor hardware.
What the platform strategy means for the market
NVIDIA's approach to robotics is consistent with how it entered the data center market: build the compute, build the software stack, and let hardware makers build on top. GTC Berlin made clear that this strategy is accelerating.
Jensen Huang told the audience that general-purpose robotics is advancing at extraordinary speed, and that the ChatGPT moment for robotics is fast approaching.
The competitive implication is significant.
Robotics companies that build on NVIDIA silicon and software can reach customers faster and with lower engineering overhead. Companies that build proprietary stacks face a higher bar to compete on software quality alone. The announcements at GTC Berlin suggest NVIDIA is deliberately raising that bar.
Sources
- NVIDIA GTC Berlin robotics announcements - NVIDIA Newsroom, October 2026
- Agile Robotics ONE SN product page - Agile Robotics, October 2026
- Bear Robotics GR00T N1.7 performance report - Bear Robotics, October 2026
- Isaac Lab 3.0 release notes - NVIDIA Developer, October 2026
Written by
Managing Director, Next Events & Conferences Ltd. Power Transmission & Distribution · Future of Energy for Industry · Humanoid Technology — the technology bringing Humanoids to Life.