NVIDIA's decision to build its reference humanoid robot around Unitree's H2 Plus body represents far more than a commercial partnership; it establishes Unitree hardware as the standard platform for embodied AI research worldwide. During his Computex keynote on June 1, 2026, NVIDIA CEO Jensen Huang announced the Isaac GR00T reference humanoid robot, combining Unitree's H2 Plus body with NVIDIA's Jetson Thor compute module and Isaac GR00T foundation models. Leading research institutions including Stanford Robotics Center, ETH Zurich, Ai2, and UC San Diego will use this design, effectively giving Unitree's robot body a technology moat that extends far beyond the Chinese market.

What Makes This Partnership So Significant for Robotics?

The timing of this announcement matters enormously. Unitree is preparing for its initial public offering on the Shanghai STAR Market, with price inquiries scheduled for August 5, 2026, an offer price set for August 6, subscriptions on August 10, and final results on August 14, targeting a valuation near 42 billion yuan (approximately $5.8 billion USD). The NVIDIA partnership provides three layers of validation for that valuation. First, it offers technical credibility by positioning Unitree as a supplier to the world's leading AI computing company. Second, it creates a technology moat independent of the Chinese market at a time when the United States has imposed import curbs on foreign-made robots. Third, it establishes Unitree's hardware as the embodiment standard in NVIDIA's open-source foundation models, meaning researchers globally will train AI on Unitree robots.

In robotics, reference designs function as blueprints that other industry participants adopt and customize. NVIDIA's choice of Unitree reflects a strategic calculation about data collection. During his keynote, Huang stated that for agentic systems, robotic systems, and physical AI, data is the hardest problem. NVIDIA needs widely adopted standard hardware to solve the collection and reuse of embodied AI training data, and Unitree's H2 became the vehicle for that standard.

How Does the Isaac GR00T Robot Actually Work?

• Physical Design: The humanoid stands roughly 1.8 meters tall (nearly six feet) and weighs about 68 kilograms, with 31 degrees of freedom in the body plus 22 in each Sharpa five-fingered tactile hand for a combined total near 75 degrees of freedom, and arms capable of lifting around 33 pounds.
• Computing Brain: The Jetson AGX Thor T5000 module integrates a Blackwell GPU rated at 2,070 FP4 teraflops (a measure of computing speed), a 14-core Arm processor, and 128GB of unified memory, with a 15Ah onboard battery supporting approximately three hours of continuous operation.
• Software Stack: The Isaac GR00T platform includes tools for teleoperation and demonstration data capture, open foundation models for reasoning and multitask behavior, simulation and training environments via Isaac Sim and Isaac Lab, and middleware for deploying trained policies onto the physical robot.

The technical specifications reveal this system's positioning in the global embodied AI value chain. Unitree provides the H2 humanoid hardware where it holds manufacturing advantages, while NVIDIA supplies the AI computing platform, robot foundation models, and simulation system support, with Sharpa providing the five-fingered hands. This split means Unitree operates as a leading supplier at the robot body layer rather than a full-stack platform owner. That positioning is both an advantage, given the cost and scale capabilities in body manufacturing that are hard to replicate, and a constraint, since value capture in software and models accrues primarily to NVIDIA.

Why Open-Source Models Change Everything?

An easily overlooked detail with far-reaching consequences is the open licensing of the software. The Isaac GR00T N1.7 open vision-language-action model and Isaac Sim training pipelines are both available on Hugging Face and GitHub, making them freely accessible to researchers worldwide. Until now, the race to build foundation models for humanoids has been a closed-loop game among well-funded startups like Physical Intelligence, Skild AI, 1X, and Figure, all training on proprietary hardware.

Unitree's hardware is now written into the embodiment taxonomy of NVIDIA's open model through a dedicated UNITREE_G1_SONIC tag for whole-body control. The vision-language-action model predicts compact latent action tokens that a learned whole-body controller decodes into full-body joint commands including legs, arms, and hands. This depth of integration far exceeds an ordinary commercial partnership; it means that every researcher using NVIDIA's open-source model will be training on Unitree hardware specifications.

"The partnership is built on the H2 humanoid hardware where Unitree holds advantages," noted Yolanda Xie, Unitree marketing manager.

Yolanda Xie, Marketing Manager at Unitree

What Does This Mean for Unitree's IPO?

The NVIDIA relationship supports Unitree's 42 billion yuan valuation on multiple fronts. It provides technical validation from the world's most trusted AI infrastructure company, establishes Unitree as the standard platform for embodied AI research at leading universities, and creates a defensible competitive position at a time when geopolitical tensions are reshaping robotics supply chains. For investors evaluating Unitree ahead of its August 10 subscription date, the NVIDIA partnership answers a critical question: why should a Chinese robotics maker command a multi-billion-dollar valuation? The answer is that Unitree has become the hardware standard for the next generation of AI research, a position that transcends any single market or regulatory environment.

The IPO timetable has entered its final phase. Preliminary price inquiries took place on August 5, with the offer price set for August 6, online and offline subscriptions opening on August 10, and payment due August 12. Final results are scheduled for August 14, with the actual listing date awaiting official exchange confirmation, targeting roughly 4.2 billion yuan in proceeds.