Nvidia Isaac GR00T Reference Humanoid Puts a Blackwell GPU in a 6-Foot Robot

Photo by ZHENYU LUO on Unsplash

Nvidia and Unitree have built a 6-foot, 150-pound humanoid robot with a Blackwell GPU in its chest and 75 total degrees of freedom, and they are selling it to research labs starting in October. The Nvidia Isaac GR00T humanoid robot, officially branded the Unitree H2 Plus, was announced June 1, 2026 at GTC Taipei as the first open humanoid robot reference design built on NVIDIA Jetson Thor and the Isaac GR00T open development platform, according to Unitree's announcement on PR Newswire.

The pitch is simple: take the hardware stack that has been gated behind Figure, 1X, and Tesla's internal teams, and ship a validated version to any robotics lab with a purchase order. CNBC reported that Nvidia selected Unitree as hardware partner for the first robotics system it is selling to researchers, from Stanford to ETH Zurich, with general availability set for October 2026.

A 2,070-Teraflop Brain Inside the Chest Cavity

The headline component is the Jetson AGX Thor T5000 humanoid compute module riding inside the torso. According to Nvidia's official spec page for Jetson Thor, the module pairs 2070 FP4 TFLOPS, 128 GB memory, and 40–130 W power, delivering 7.5× the AI performance and 3.5× the efficiency of AGX Orin. That figure matters because vision-language-action models, the class Nvidia has bet on for humanoid control, are memory-bandwidth bound before they are compute bound.

The T5000's 128 GB unified memory pool is the under-discussed spec. HotHardware's teardown of the Jetson AGX Thor developer kit noted that the memory is a giant pool of unified resources, much like the gobs of unified memory available in Apple Silicon Macs. That means that it doesn't matter if it's a task better suited to the GPU or the CPU, all of the 128GB of memory is available for processing. For a robot running a perception model, a planning model, and a control policy in parallel, that pool eliminates the copy-overhead penalty that limited earlier Jetson Orin platforms.

The other detail worth flagging: the Blackwell GPU architecture employed on Jetson AGX Thor also offers Multi-Instance GPU (MIG) support, which can slice the graphics processor into up to 7 partitions for running multiple models simultaneously. A humanoid robot is not running one model. It is running a vision stack, a language-grounded planner, a low-level motor controller, and safety monitors at the same time. MIG turns that from a scheduling nightmare into a hardware-isolated guarantee.

Nvidia's Isaac GR00T workflow handles the rest of the stack. Nvidia's announcement confirms the Jetson AGX Thor T5000 inside the H2 Plus runs at the same configurable 40–130W envelope, with the same Blackwell GPU and 14-core Arm CPU as the standalone module.

Unitree H2 Plus Specs: 75 Degrees of Freedom and 360 Nm Leg Torque

The chassis is where Unitree, not Nvidia, did the engineering. The published Unitree H2 Plus specs describe a robot nearly 6 feet tall and 150 pounds with 31 degrees of freedom across the body, paired with dual Sharpa Wave tactile five-finger hands at 22 degrees of freedom each, bringing the platform total to 75 DoF.

The actuator numbers are aggressive for a research-tier humanoid:

• Arm torque up to 120 Newton-meters
• Leg torque up to 360 Newton-meters
• Rated arm payload of 7 kg
• Peak payload of 15 kg
• Head-mounted stereo camera with 140° horizontal, 102° vertical field of view
• Wrist cameras for close-range manipulation and an IMU for motion tracking

Leg torque at 360 Nm is the spec that determines whether a humanoid can recover from a slip on a warehouse floor or fall on its face. Most academic platforms in the sub-$100k range have settled for 100–200 Nm. Pairing that with a peak 15 kg payload puts the H2 Plus closer to industrial logistics geometry than to the demo-floor robots that can only pick up a coffee cup.

The hands are the part nobody else in the reference-design space has solved. Sharpa's press release confirms the Singapore-based company is supplying its Wave tactile five-finger hands as the first dexterous humanoid end-effector integrated into the Isaac GR00T reference design. Sharpa's credibility for the spec comes from a public demo: at CES 2026 in Las Vegas, the company's full-body robot "North" played fully autonomous ping-pong rallies against human opponents, establishing independent dexterous-manipulation chops months before this announcement.

Sharpa claims the validated configuration cuts robot skill setup time "from days to hours," and that developers can simulate, train, and evaluate dexterous manipulation policies using Sharpa Wave's tactile parameters within Isaac Sim and Isaac Lab before committing to real-world deployment. That last point is the actual product. The hands are not the differentiator; the simulation parity is.

Why an NVIDIA Humanoid Robot for Research Is Different This Time

The overlooked angle is the institutional list. According to Unitree's release, the four confirmed early adopters are Ai2 in Seattle, ETH Zurich, the Stanford Robotics Center, and UC San Diego's Advanced Robotics and Controls Laboratory. No Chinese institutions appear on the list, despite Unitree being a Chinese company.

That omission is not accidental. CNBC noted that Unitree disclosed more than 40% of its revenue already comes from markets outside China, and the announcement coincides with Unitree pursuing a 4.2 billion yuan (~$620 million USD) IPO on Shanghai's STAR Market. Selling a humanoid platform co-marketed with Nvidia, into U.S. and European labs, is a geopolitical hedge as much as a product launch.

Rev Lebaredian, Nvidia's VP of physical AI simulation, framed the access argument directly. CNBC quoted him saying the platform is about "taking frontier humanoid research out of the hands of only the world's largest tech companies and AI unicorns, and putting it in reach of every lab". He added that H2 Plus will be available in October and said "anyone can buy it". Neither Nvidia nor Unitree has published a unit price.

Jensen Huang's keynote framing was characteristically maximal. Per Nvidia's newsroom, he said: "Humanoid robots will bring physical AI to the world's largest industries, opening a multitrillion-dollar economic opportunity." The more revealing line, also from the keynote: "This platform runs the new Thor, and our entire software stack, data generation stack, data simulation stack, the runtime, all integrated into a robot that is designed for everyone to use."

The Isaac GR00T Reference Design Is Really a Software Lock-In

The hardware gets the press photo. The lock-in is in the workflow.

Unitree CEO Xingxing Wang's statement is unusually direct about this: "H2 Plus combines Unitree's humanoid with NVIDIA Jetson Thor and the NVIDIA Isaac GR00T development platform, giving teams a validated starting point for creating robot skills." According to the Nvidia announcement, the Isaac GR00T reference design covers data capture and generation, robot model evaluation, and deployment workflows in a single chain.

The practical translation: a Stanford grad student who trains a manipulation policy in Isaac Sim against the H2 Plus URDF can deploy that policy onto physical hardware without rewriting the perception or control layers. Every lab that builds a thesis on this stack ships its trained weights, its dataset format, and its simulation assumptions into Nvidia's tooling. The reference workflow for the smaller Unitree G1 is also coming soon to GitHub and Hugging Face, per the Nvidia announcement.

That is how reference designs work, and it is the same pattern Nvidia ran with CUDA in the early 2010s and with Jetson in the autonomous-vehicle research stack before that. The hardware is the foot in the door. The simulation, training, and deployment pipeline is the moat. AMD, Qualcomm, and the smaller robotics-compute startups have nothing approaching parity at this layer, which is why the conversation in AI infrastructure keeps returning to the same question of inference lock-in.

The Limitation Nobody Is Advertising

Humanoid robots are not solved. The press release acknowledges this obliquely, framing the H2 Plus as a research platform rather than a production system, and the CNBC reporting is blunter: humanoid deployments of AI-powered robots have largely been limited to warehouses so far, with safety and reliability constraints still unsolved.

That is the asymmetric reality. A 150-pound, 6-foot machine with 360 Nm of leg torque is not safe to operate near humans without engineered cell containment, even with the wrist cameras and IMU feeding the safety stack. Every lab buying an H2 Plus in October will need to write its own e-stop protocols, its own contact-force limits, and its own fall-recovery procedures. Nvidia is shipping the brain and the body. The institutional risk acceptance is on the buyer.

This is not a complaint about the product. It is a clarification of what "available to every lab" actually means. The four confirmed early adopters all run dedicated robotics safety programs with full-time engineers handling exactly this overhead. A startup buying an H2 Plus expecting plug-and-play deployment will discover the gap quickly. For broader context on where the robotics and drones category sits in 2026, the H2 Plus extends the research-platform tier rather than crossing into commercial autonomy.

What to Watch Next

The specific date is October 2026, when Unitree begins shipping H2 Plus units. Three concrete things will determine whether this reference design moves the field or becomes another well-funded demo:

1. The Hugging Face drop of the Isaac GR00T reference workflow for the Unitree G1, which Nvidia says is coming soon and will signal how openly the company actually intends to publish its training pipelines.
2. The Shanghai STAR Market review of Unitree's IPO application, which CNBC reported is scheduled the same week as the announcement and will determine whether Unitree has the capital to ship at volume.
3. The first peer-reviewed paper from one of the four adopter institutions, expected in the 2027 ICRA or CoRL submission cycles, which will reveal whether H2 Plus enables manipulation results that the previous generation of academic humanoids could not produce.

If the answer to that third item is yes, Nvidia will have done to humanoid robotics what it did to deep learning in 2012: standardized the platform before anyone realized the platform decision was the important one.