75 degrees of freedom. 2,070 FP4 teraflops of onboard AI compute. A battery that lasts about three hours. That is what NVIDIA and Unitree are shipping to university robotics labs later this year, and it is the most complete hardware-plus-software reference design the humanoid robotics field has seen from either company.
On June 1, 2026, at NVIDIA GTC Taipei and Computex, Jensen Huang announced the NVIDIA Isaac GR00T Reference Humanoid Robot — a joint platform built with Unitree Robotics. The hardware side is called the Unitree H2 Plus. Together, they form the first open humanoid robot reference design built on NVIDIA's Isaac GR00T development platform. This announcement is part of a broader physical AI push from NVIDIA — including Cosmos 3 and the Vera robotics layer — that the company has been rolling out through Computex 2026.
This is not a product launch. It is an infrastructure decision for the research community. Here is what the announcement actually contains.
What the H2 Plus Actually Is
The H2 Plus is a reference design. That means NVIDIA and Unitree are not just selling a robot — they are selling a starting point. The idea is that a research team at Stanford or ETH Zurich should not have to spend months integrating hardware, writing drivers, and building simulation pipelines before they can start doing actual research. The H2 Plus bundles all of that into one validated system.
Unitree Robotics, based in Hangzhou, China, built the chassis. Sharpa, a Singapore-based company, built the five-fingered hands. NVIDIA contributes the compute module and the full Isaac GR00T software stack. The three components are integrated and shipped as a single unit.
The underlying Unitree H2 robot body (without the H2 Plus additions) is currently listed on Unitree's website at $29,900. No pricing has been announced for the complete H2 Plus system.
Hardware Specs: The Full Breakdown
The H2 Plus chassis stands nearly 6 feet tall and weighs 150 pounds. It has 31 degrees of freedom across the body. Add the Sharpa Wave five-finger hands and that number climbs to 75 degrees of freedom total — 22 per hand, across the body and hands combined.
| Spec | Value |
|---|---|
| Height | ~6 feet |
| Weight | 150 lbs |
| Body degrees of freedom | 31 |
| Total DOF (with hands) | 75 |
| Arm torque | 120 Nm |
| Leg torque | 360 Nm |
| Rated arm payload | 7 kg per arm |
| Peak arm payload | 15 kg per arm |
| Movement speed | Up to 2 m/s |
| Onboard compute | NVIDIA Jetson AGX Thor T5000 |
| AI performance | 2,070 FP4 teraflops |
| Unified memory | 128 GB |
| CPU | 14-core Arm |
| Power range | 40 to 130 watts (configurable) |
| Battery capacity | 15Ah / 0.972 kWh |
| Battery life | ~3 hours |
| Head camera FOV | 140° horizontal, 102° vertical |
| Connectivity | Ethernet, Wi-Fi 6, Bluetooth 5.2, USB |
The sensing suite includes a head-mounted stereo camera (140 degrees horizontal, 102 degrees vertical field of view), wrist cameras for close-range manipulation tasks, and an inertia measurement unit for motion tracking. Microphones and speakers are also included for voice interaction research. An on-remote emergency stop is built in.
The NVIDIA Software Stack
The hardware is only half of it. What NVIDIA is really offering here is the Isaac GR00T development platform — a full-stack set of tools that covers every stage from data capture to real-world deployment.
The stack includes five components. Isaac Teleop handles teleoperation data capture for training. Isaac GR00T open foundation models cover humanoid reasoning and multitask behavior — these are available on GitHub. Isaac Sim and Isaac Lab provide the simulation environment for training and evaluating robot policies before anyone puts a physical robot at risk. Isaac ROS middleware handles the deployment step, moving trained policies onto actual hardware. And the Jetson Thor module runs inference and control on the robot itself, in real time.
The design is modular. Research teams can use the full platform or plug individual components into existing pipelines. Crucially, the announcement states that researchers retain control of their own robot data, training data, telemetry, and logs — NVIDIA does not take the data.
The same Isaac GR00T platform will also support the Unitree G1 — a smaller humanoid that is already widely deployed in research labs. That reference workflow is expected to land on GitHub and Hugging Face, though no specific date has been given.
Who Is This Built For
NVIDIA named four institutions that will use the H2 Plus at launch: Ai2, ETH Zurich, Stanford Robotics Center, and UC San Diego's Advanced Robotics and Controls Laboratory. All four are Western research institutions.
That list is worth noting. Unitree is a Chinese company, and the H2 Plus was announced jointly at an event in Taipei. But the named early adopters are entirely US and European academic labs. CNBC reported that no China-based research institutions were listed as launch partners. Whether that reflects deliberate positioning for the Western research market or something else, the announcement does not say.
The target use cases cited are factories, warehouses, and logistics systems — not consumer or home applications. Humanoid robot deployments have so far been largely limited to warehouse environments, and NVIDIA's framing here stays within that range.
What Is Not Confirmed Yet
A few things in the announcement are forward-looking and worth separating from what is confirmed.
Unitree says the H2 Plus will be available in the second half of 2026 — this is an expected timeline, not a confirmed ship date. The complete system price has not been announced. The Isaac GR00T reference workflow for the Unitree G1 is described as "expected soon" on GitHub and Hugging Face, with no specific release date. Jensen Huang's reference to a "multitrillion-dollar economic opportunity" for humanoid robotics is a market projection, not a reported figure.
My Take
The fragmentation problem NVIDIA is solving here is real. Any researcher who has tried to build a humanoid robotics pipeline from scratch knows how much time gets eaten by integration work that has nothing to do with the actual research question. Buying a robot, writing drivers, setting up simulation, building a data collection workflow — that is months of effort before the science starts. If the H2 Plus delivers on the "plug in and start training" promise, it is genuinely useful infrastructure.
That said, this is a press release, and the modular software stack that "covers every stage from data capture to deployment" is a description of components, not a proof of integration quality. Whether Isaac Teleop, Isaac Sim, and Isaac ROS actually work smoothly together in practice — or whether each tool requires its own wrestling match — we will only know when researchers start publishing results from it.
The three-hour battery life is the number I keep coming back to. For a research platform that costs this much and targets this kind of serious use, three hours of operation is thin. That one detail tells you something about the current state of the hardware.
- H2 Plus is the first humanoid robot reference design built on NVIDIA Isaac GR00T — announced June 1, 2026 at Computex Taipei.
- Hardware: ~6 feet tall, 150 lbs, 75 total degrees of freedom, 2,070 FP4 teraflops onboard via Jetson AGX Thor T5000.
- Software stack covers data capture, simulation, training, evaluation, and deployment — all open, researchers keep their own data.
- Launch partners include Stanford, ETH Zurich, Ai2, and UC San Diego. No China-based institutions listed.
- H2 Plus availability is expected in late 2026 — no confirmed ship date or system price announced.
FAQ
What is the Unitree H2 Plus?
The Unitree H2 Plus is a full-size humanoid robot designed for academic and research use. It is built on the NVIDIA Isaac GR00T development platform, combining Unitree's H2 chassis, Sharpa Wave five-finger dexterous hands, and NVIDIA Jetson Thor onboard compute into a single integrated system.
What is NVIDIA Isaac GR00T?
Isaac GR00T is NVIDIA's open development platform for humanoid robots. It includes tools for teleoperation data capture (Isaac Teleop), simulation and training (Isaac Sim and Isaac Lab), open foundation models for reasoning and behavior, and middleware for deploying policies onto physical robots (Isaac ROS). The platform is designed to cover the full development workflow from data collection to real-world deployment.
When will the Unitree H2 Plus be available?
Unitree has announced the H2 Plus will be available in late 2026. This is the expected timeline stated in the official announcement — a confirmed ship date has not been provided. Pricing for the complete H2 Plus system has also not been announced.
What is NVIDIA Jetson Thor?
Jetson AGX Thor T5000 is NVIDIA's onboard compute module used in the H2 Plus. It features an NVIDIA Blackwell GPU with 2,070 FP4 teraflops of AI performance, a 14-core Arm CPU, 128GB of unified memory, and a configurable 40 to 130-watt power range. It handles real-time sensor processing and robot inference directly on the robot hardware.
Which universities are using the H2 Plus?
NVIDIA named four institutions as launch partners: Ai2, ETH Zurich, Stanford Robotics Center, and UC San Diego's Advanced Robotics and Controls Laboratory. All are based in the United States or Europe.
Does the H2 Plus support the Unitree G1?
The NVIDIA Isaac GR00T platform will also support the Unitree G1 humanoid robot, extending the same development approach to a robot already used widely in research labs. The reference workflow for G1 is expected to be available on GitHub and Hugging Face, though no release date has been given.
Conclusion
The H2 Plus is a notable infrastructure commitment, not a capability breakthrough. NVIDIA is betting that the biggest bottleneck in humanoid robotics research right now is not the algorithms — it is the overhead of standing up a working development environment. That bet might be correct.
What we will find out in the second half of 2026, when hardware actually ships to Stanford and ETH Zurich, is whether this integration holds together under real research conditions. The specs are solid. The software stack is comprehensive on paper. The research institutions named are credible validators.
The question no press release can answer: does it actually save researchers time, or does it just move the integration headache to a different layer? That answer is coming.
Sources: Unitree press release, PR Newswire, June 1, 2026 | NVIDIA Newsroom, June 1, 2026
0 Comments