NVIDIA's open-source humanoid blueprint packs Jetson Thor brain

NVIDIA just dropped an open reference design for humanoid robots, effectively handing researchers a standardized blueprint for physical AI.

Revealed at GTC Taipei, the Isaac GR00T Reference Humanoid Robot pairs off-the-shelf robotics hardware with NVIDIA's upcoming onboard superchip. Instead of spending months cobbling together parts from five different vendors, research labs now get a unified body and brain right out of the box.

Unitree body, Blackwell brain

The physical build combines two serious pieces of hardware. The chassis is a Unitree H2 Plus humanoid standing nearly six feet tall and weighing 150 pounds, sporting 31 degrees of freedom across the body.

For dexterous manipulation, NVIDIA added dual Sharpa Wave tactile five-finger hands. Each hand adds 22 degrees of freedom, bringing the total robot to 75 degrees of freedom. The arms deliver up to 120 Newton-meters of torque with a 15-kilogram peak payload, while the legs push 360 Newton-meters.

The real showstopper sits inside: the Jetson AGX Thor T5000 computer. Powered by a Blackwell GPU pushing 2,070 FP4 teraflops, it packs a 14-core Arm CPU and 128GB of unified memory. Running within a configurable 40 to 130-watt power envelope on a 0.972kWh battery, the robot gets around three hours of runtime on a single charge.

NVIDIA is also pairing the robot with its full software loop, from Isaac Teleop demonstration capture to Isaac Sim synthetic training and ROS middleware deployment. Research heavyweights including Stanford, ETH Zurich, Ai2, and UC San Diego have already signed on, with Unitree scheduled to ship the full build in late 2026.

Why it matters

Humanoid robotics has long suffered from an apples-to-oranges problem. Every top lab builds custom rigs, meaning breakthrough research from one group almost never replicates smoothly on another's hardware.

By offering an open, standardized stack, NVIDIA is trying to turn the Unitree-Thor combo into the default reference platform for robotics academia. In our view, this frees researchers from rebuilding low-level infrastructure and lets them focus purely on general-purpose manipulation policies.

Strategically, this is classic Jensen Huang. Just as CUDA locked in a generation of machine learning scientists, getting tomorrow's top robotics engineers hooked on Jetson Thor and Isaac Sim ensures that NVIDIA remains the compute tollbooth when humanoids eventually hit factory floors.

The real bottleneck will be price and delivery. While a lighter workflow for the budget Unitree G1 is heading to GitHub and Hugging Face soon, this flagship 75-DoF rig will command enterprise-level budgets, and research labs still have to wait until late 2026 to get physical units in hand.