Nvidia unveils Agile ONE SN reference humanoid robot at GTC Berlin

Nvidia unveiled Agile ONE SN, a reference humanoid robot developed with the German company Agile Robots, at its GTC Berlin conference on Wednesday, October 21, 2026. Chief executive Jensen Huang presented the robot during a keynote of roughly two hours at the Tempodrom in Berlin, alongside new robot-training software and an open-source system meant to keep AI agents from reaching beyond what they are allowed to touch. Nvidia and Agile Robots say the humanoid will be available to developers in early 2027.

Humanoid robots have become one of the most heavily funded corners of artificial intelligence, but the field is fragmented. Each builder tends to stitch together its own motors, hands, cameras, chips, simulation tools and training data, which makes it slow to move from a lab demonstration to a machine that works on a factory floor. A reference design is Nvidia's answer to that problem. It is the same approach the company has long used with graphics cards and servers: publish a complete working blueprint so that partners can build on it instead of starting from scratch. Agile Robots says the platform can cut the time needed to set up a development system from weeks to days. That is the company's claim; it has not yet been tested by independent users.

Two technical terms explain what is being offered. "Degrees of freedom" counts the separate ways a robot's joints can move; more of them allow finer, more human-like motion, which matters most in the hands, where a robot must grip a screw, a cable or a cup without crushing it. "Simulation" means training the robot's software inside a detailed virtual world before it ever touches a real object. Isaac Lab, Nvidia's simulation framework, lets developers run thousands of practice attempts in software, which is far cheaper and safer than learning by trial and error on real machines. The 2,000 ready-made virtual objects released this week are meant to make those virtual worlds quicker to build.

The robot was one part of a wider message. Huang used the Berlin keynote to urge European industry to move from buying AI to producing it, a pitch Nvidia frames as "sovereign" AI: companies and countries running open models, such as Nvidia's own Nemotron family, trained further on their own data and hosted on their own infrastructure. Nvidia said it works with more than 12,000 partners in Europe. Among the other announcements, it said Volvo Cars is adopting its Hyperion computing platform for automated driving and that it is working with Siemens Healthineers on AI-assisted radiation therapy. It also showed DSX POD, a design for building AI data centres inside modular containers, which it says can cut on-site deployment to 14 days once power and cooling are in place.

OpenShell addresses a different worry. AI "agents" are programs that do not just answer questions but act: they open files, send messages and run code. Nvidia describes OpenShell as a policy-controlled runtime that sits between an agent and the systems it can affect, using sandboxing at the level of the operating system's kernel to fence off what the agent can reach. For Indian companies building agents on top of open models, the release adds another free tool for limiting what such software can do on their own networks.

Short-term: Robot developers, university labs and start-ups can plan around a common hardware and software design, and begin training robot skills in simulation now, before the physical humanoid ships in early 2027.

Long-term: If many builders adopt the same blueprint, humanoid development could speed up and costs could fall, while Nvidia's chips and software become even more deeply embedded in the robotics supply chain. That concentration is the trade-off: easier entry for newcomers, but heavier dependence on one supplier.

Who is affected: Robotics companies and researchers, including India's growing group of robotics start-ups and engineering colleges; manufacturers considering automation; and firms deploying AI agents that need guardrails.

Questions and Answers