NVIDIA Corp. announced a major push into the humanoid robotics sector on Monday with a standardized developer platform aimed at accelerating academic research while quietly maneuvering around growing geopolitical and cybersecurity concerns.
During a packed keynote address at the Computex trade show in Taiwan, NVIDIA CEO Jensen Huang introduced the Isaac GR00T Reference Humanoid Robot, also known as the H2 Plus. Built specifically for universities and higher education institutions, the open foundation platform aims to eliminate the Frankenstein approach researchers currently face when piecing together disparate hardware and software.
“Humanoid robots will bring physical AI to the world’s largest industries, opening a multitrillion-dollar economic opportunity,” Huang told the audience.
The new 6-foot-tall, 150-pound reference robot is a tri-national collaboration. The chassis comes from Chinese robotics firm Unitree, featuring 31 degrees of freedom. The hands were developed by Singapore-headquartered Sharpa, featuring 22 degrees of freedom and tactile sensing capable of mimicking human dexterity for precise tasks like food preparation or assembly.
Meanwhile, the brain is powered by NVIDIA’s Jetson AGX Thor T5000 onboard compute, utilizing the company’s advanced Blackwell GPU architecture.
Prominent institutions, including Stanford University, UC San Diego, Ai2, and ETH Zurich, have already signed on as early adopters.
The partnership comes at a turbulent time for Unitree. The Chinese firm is currently pursuing a public listing in China but faces intense scrutiny from U.S. lawmakers. Alleging extensive ties between Unitree and the Chinese military, Washington politicians have introduced legislation that would ban researchers receiving U.S. government funding from using the company’s robots.
To mitigate risks, NVIDIA executives revealed to Reuters that the H2 Plus platform directly integrates enterprise-level cybersecurity into the hardware. For the first time, data-center security technologies like “secure boot” and “confidential computing” will be deployed on a humanoid robot.
According to anonymous executives, all software updates meant for the robot’s subsystems must flow through NVIDIA’s silicon, allowing the chip to check the code for authenticity. This ensures malicious code cannot be executed and prevents sensitive research data from being exfiltrated without permission.
“NVIDIA’s reference design shifts humanoid robotics from a hardware and software competition to competing on the development lifecycle. NVIDIA standardizes data capture, simulation, training, and deployment into one pipeline, keeping the body modular while compute and models stay proprietary,” said Mitch Ashley, vice president and practice lead for Software Lifecycle Engineering and AI-Native Software Engineering at The Futurum Group. “Labs adopting the stack shed the bespoke integration that has slowed humanoid research and inherit NVIDIA’s compute and models as the default substrate. Standardizing the lifecycle early sets a multi-year dependency, and the portability that eases adoption is what deepens it.”
Acknowledging the shifting geopolitical landscape, NVIDIA executives confirmed plans to expand similar robotics partnerships to manufacturers in the United States, Europe, and South Korea. These upcoming collaborations remain confidential as the company seeks to build a globally distributed, secure ecosystem for physical artificial intelligence (AI).
While Unitree will handle the commercial availability of the H2 Plus chassis, the combined platform represents a structural shift for AI labs. By providing a unified, secure foundation of hardware and open-source models, NVIDIA is positioning itself as the central operating engine for the next generation of robotic labor.
To that end, the chip maker unveiled Cosmos 3, an open-world foundation model designed to advance physical AI. Built on a mixture-of-transformers architecture, the single system integrates vision reasoning, world generation, and action prediction to help autonomous systems and robots navigate real-world environments.
Billed as the world’s first fully open omnimodel, Cosmos 3 natively processes and generates text, images, video, ambient sound, and physical actions with high physics accuracy. According to the company, this multimodal capability drastically accelerates development, compressing physical AI training and evaluation cycles from months to just days.
NVIDIA also announced expansion of its DRIVE Hyperion ecosystem to accelerate global deployment of commercial robotaxis. The initiative unites global automakers, software developers, and ride-hailing providers to scale Level 4 autonomous vehicle fleets. The DRIVE Hyperion platform serves as a production-ready foundation for the transportation industry. Built on the company’s Halos full-stack safety system, it integrates high-performance DRIVE AGX in-vehicle computing, a safety-certified operating system, a multimodal sensor suite, and specialized AI software.
“Autonomous mobility is entering its industrial scaling moment,” Huang said, emphasizing that future robotaxi fleets will require robust AI infrastructure to operate safely in the real world, positioning Hyperion as the common foundation to transition self-driving cars from pilot programs to everyday transportation.
Additionally, NVIDIA extended its open-source AI lineup with the introduction of NVIDIA Alpamayo 2 Super, a 32-billion-parameter vision-language-action (VLA) model designed to accelerate the development of safe, Level 4 autonomous robotaxis.
The new reasoning-based model builds upon the company’s existing Alpamayo family, which integrates open AI models, simulation frameworks, and physical AI datasets. By enhancing how autonomous vehicles perceive and react to their environments, the computing pioneer aims to provide developers with a more robust, open-source foundation for next-generation self-driving fleets.
