Nvidia will ship an open reference robot for humanoid research at the end of 2026. The brain comes from Nvidia, the body from Unitree in Hangzhou. Since July 28, this exact device category can no longer be approved in the US.
Symbolic image · AI-generatedNvidia presented an open reference robot for humanoid research in late May at GTC in Taipei, and this robot is slated for delivery at the end of 2026—in the coming weeks, in other words. The brain comes from Nvidia, the body from Unitree in Hangzhou, plus two tactile five-fingered hands from Sharpa, totaling 75 degrees of freedom.
But on July 28, the American telecommunications regulator FCC placed foreign-manufactured humanoids and robot dogs on its prohibition list, meaning new models in this category will no longer receive device approval in the US. Of the labs Nvidia names as future users in its announcement, three are in the United States, and ETH Zurich is on the same list.
Technically, the package is impressive. The body measures just over 1.80 meters, weighs around seventy kilograms, and comes with thirty-one joints, with the two hands each contributing twenty-two more. Computing happens on board via a Jetson AGX Thor with 128 gigabytes of shared memory, whose power consumption can be adjusted between forty and one hundred thirty watts, and the battery lasts approximately three hours according to the manufacturer. The arms lift seven kilograms in continuous operation and fifteen at peak, which is sufficient for lab work but not for warehousing.
More interesting than the column of numbers is what Nvidia means by open. Open are the models and tools—that is, the GR00T models on GitHub and Hugging Face, the teleoperation software for gathering demonstrations, the simulation environments Isaac Sim and Isaac Lab, as well as the middleware Isaac ROS. The announcement says nothing about design data, bill of materials, or a license for the hardware. Open is the brain; the body remains a product you have to buy, and that's exactly where the problem lies now.
The FCC measure does not prohibit possession or resale. It prevents new models from receiving the approval they need for the American market, while already-approved devices may continue to be sold and operated. For the reference robot, this means: either this model received approval before July 28, or it requires a case-by-case review, for which the American Department of Defense can grant conditional authorization. Neither Nvidia nor Unitree comment on this, and the American distributor advertises an exclusive US launch at the end of 2026 on its product page without mentioning approval in a single sentence. No one on that page names a price either.
Jensen Huang said at the presentation that humanoid robots would bring physical AI to the world's largest industries and open up an economic opportunity in the trillions. Marco Hutter from the Robotic Systems Lab at ETH Zurich put it more soberly, speaking of a platform at the state of the art for collecting data, testing algorithms, and verifying robot behavior. Both statements are true, but they rest on a device whose import into the world's largest research market remains unclear.
In October, Nvidia presented a second reference humanoid at GTC Berlin, the Agile ONE SN from Agile Robots in Munich, which is due in early 2027 and is significantly smaller at 1.40 meters and forty kilograms. Added to this were the general-availability version of Isaac Lab and an open real-time operating system for robots. If you want to see the pattern, here it is: Nvidia gives away the software, sells the computing components, and leaves the body to others—where the choice of body has now become a political one.
For Swiss labs, this creates a situation that no one in Zurich likes to discuss openly. Here, there is no prohibition list for Chinese robots; Unitree devices are freely available and listed with prices in the directory of robothub24.ch. An institute in Zurich or Lausanne can thus order the reference robot as soon as it ships, while a California lab may be waiting for a government agency. That is a location advantage that did not arise from its own strength but from a trade conflict, and it can disappear as quickly as it appeared.
What Nvidia gives away is software, and there is plenty of it. What becomes scarce is the body it runs on.
This article was created with the support of artificial intelligence and editorially reviewed. The article image is an AI-generated symbolic image, not a press photo.