The world by now possesses a seemingly vast array of AI models and agents deployed in industrial, research and office settings.
However, the development and scale production of humanoid robots is somewhat elusive and insanely hard, as a senior Nvidia official made clear in talking to reporters in Taiwan on the eve of Computex 2026.
A big problem with humanoids is equipping them with hands with five fingers able to grasp objects. What’s harder is how to build a humanoid brain to control the grasping of objects and to do all the other things a humanoid is asked to do. It is a problem for researchers working to move humanoids beyond their current early state, beyond what Nvidia’s Vice President of Physical AI Rev Lebaredian, termed “franken-robots.”
“In order to create humanoid robots at scale, we have to create a humanoid brain and the only way to do this is with researchers in research labs who need to have all the advance capabilities of a full robot,” he said.
“Extremely important are robotic hands for manipulation of objects. For that we need the most advanced compute and every bit of compute is necessary to push to the limits. But very few of these humanoids are available. They are not being produced at scale and they are kind of franken-robots, so a lab spends time on the basics of a robot to make them work.”
Everything needs to work together, from compute to software stack to robot hands and body. They must work together, at first, in a reference platform, Lebaredian added.
It turns out Nvidia just announced such a reference design in partnership with robotics company Unitree for its H2 Plus humanoid. Lebaredian called it the world’s first open frontier humanoid built on Nvidia’s Isaac GROOT.
It includes Jetson Thor onboard compute with 2070 FP4 Teraflops, a 14-core Arm CPU and 128 GB of memory. It also has five-finger hands with Sharpa Wave dexterous manipulation and the GROOT 1.7 software stack with Unitree whole body control.
In a press release from Nvidia, several leading research institutions were named as preparing to use the H2 Plus reference design to "advance frontier humanoid robotics research." They include AI2, ETH Zurich, Stanford Robotics Center and US San Diego's Advanced Robotics and Controls lab.
Unitree has also announced that technical specs of the H2 Plus are online. Unitree H2 Plus will also be available commercially in late 2026, Nvidia said.
Why a reference design?
Why does a humanoid reference design matter to Fierce Sensors readers? That's because many such readers are developers of Edge AI applications that include robots and humanoids. There are literally dozens or even hundreds of optical and motion sensors in robots and robotaxis and that's not counting the related components that spew data into a system that a robot reference design would use to infer actions such as how to unscrew a cap on a bottle of pickles.
"Humanoid robots will bring physical Ai to the world's largest industries, opening a multi-million-dollar economic opportunity," said Jensen Huang, CEO of Nvidia in a statement. "Unitree H2 Plus is the first humanoid robot reference design bulit on Nvidia Isaac GROOT to help take front robotics research to useful work in factories, warehouses and logistics systems around the world."
The Unitree H2 chassis is nearly 6-feet tall and weighs 150 pounds with 31 degrees of freedom across its body. The Sharpa Wave tactile five-finger hands allow dexterity across 22 degrees of freedom to bring the robot to 75 degrees of freedom across the body and hands.
Multi-view sensing is enabled in the robot, including a head-mounted stereo camera with wide field of view, wrist cameras for close-range manipulation and an inertia measurement unit for motion tracking. The 15Ah battery provides about three hours of life.
Isaac GROOT itself is a complex platform with Isaac Teleop for capturing robot demonstration data for training and open foundation models to support reasoning, learning and multitask work. Isaac Sim and Isaac Lab are used to emulate, train test and evaluate robot policies before real-world deployment. Also, Isaac ROS middleware is used to move trained policies to the actual robots, while Jetson Thor is used to run real-time, on-robot inference and control.
Isaac GROOT will also support the Unitree G1 humanoid robot, Nvidia said.
The research teams working with the Unitree H2 Plus all praised the collaboration with Nvidia. "Unitree H2 Plus gives our students and collaborators an open humanoid reference design with dexterous hands, onboard AI compute and the Nvidia Isaac GROOT development platform for creating, comparing and shareing robot behaviors on physical hardware," said Steve Cousins, executive director of the Stanford Robotics Center, in a statement.
Other Nvidia announcements
Nvidia also announced the launch of Cosmos 3, an omnimodel for physical AI. It can be applied to post-training for vision AI agents used in autonomous robots. A new Cosmos Coalition was also announced of leading AI labs, including Agile Robots, Black Forest Labs, Generalist, LTX, Runway and Skild AI.
Nvidia is also releasing open source agent tools and skills for physical AI to help automate repeatable AI workflows from data to validation. This will help lower the barrier to start building with Nvidia physical AI tools, Lebaredian said.
The tools are available on GitHub and in partnership with Siemens, Cadence, Delta, Foxconn, Agile Robots and Pegatron.
Nvidia separately announced it is working with Foxconn in Taiwan on bringing agentic AI and physical AI to the nation’s initiative to create a “Healthy Taiwan” even as the country already has one of the best healthcare systems in the world. Part of the vision includes agents working as radiologists and nurses as well as robots for hospital services and surgery.
On the robotaxi front, Nvidia announced Foxconn will also build Nvidia’s Drive Hyperion robotaxis in Taiwan , while VinFast and Autobrain will work to launch level 4 autonomous vehicles using Drive Hyperion in southeast Asia. Nvidia said Uber will launch Drive Hyperion robotaxis in Germany and Humain is expected to launch Drive Hyperion robotaxis in the Middle East.
Nvidia also launched Alpamayo 2 Super, an open reasoning model for robotaxis.
Nvidia CEO Huang also made a number of data center-related announcements in a GTC at Computex keynote held in Taipei.
They include a new CPU called Vera, DSX software to optimize token cost and performance and Nematron 3 Ultra to power long running autonomous agents.
A new DGX Station for Windows supercomputer for use at the deskside was also announced to help developers build and run AI agents on Windows.