Isaac GR00T Open Humanoid Platform Runs on Jetson Thor T5000

At GTC Taipei, NVIDIA CEO Jensen Huang introduced the NVIDIA Isaac GR00T Reference Humanoid Robot, the company’s first open hardware and software stack for humanoid robotics. Developed alongside Unitree and Sharpa, the platform aims to replace fragmented physical AI pipelines with an end-to-end ecosystem.

Historically, robotics developers spent months assembling custom sensory arrays and writing low-level control loops before training a single behavior policy. The Isaac GR00T stack standardizes the physical layer, providing immediate baseline capabilities for locomotion, perception, and fine motor control.

Hardware Architecture and Compute

The physical platform relies on the Unitree H2 Plus chassis, an upgrade to the prior H1 and H2 series. The robot stands six feet tall, weighs 150 pounds, and supports a peak payload of 15 kg. A 0.972kWh battery provides approximately three hours of operational runtime.

Dexterity is managed by Sharpa Wave tactile hands. The entire system features 75 degrees of freedom, allocating 31 DOF to the core body and 25 DOF to each five-fingered hand.

Onboard inference runs through the NVIDIA Jetson AGX Thor T5000 system-on-module. Utilizing the Blackwell GPU architecture, this module delivers 2,070 FP4 teraflops of compute. High-throughput local inference is critical for robotics, as cloud latency disrupts the real-time feedback loops required for physical balance and manipulation.

Sensing hardware includes a head-mounted stereo camera offering a 140-degree horizontal and 102-degree vertical field of view. Wrist cameras guide the Sharpa Wave hands during close-quarters manipulation, while an inertial measurement unit tracks overall chassis motion.

Simulation and Foundation Models

The reference design serves as a deployment vessel for the Isaac GR00T ecosystem. NVIDIA provides access to the GR00T N1.6 and N1.7 foundation models, which process language, vision, and proprioceptive telemetry natively.

NVIDIA also updated Isaac Lab inside the Isaac Sim 6.0 release to feature the new Newton physics backend. This update allows for kit-less installation and accelerates reinforcement learning without requiring the full Omniverse Kit footprint.

The workflow stack relies on Isaac Teleop for remote demonstration data capture and Isaac ROS for middleware routing. The system also integrates NVIDIA Cosmos 3, a newly released omnimodel that unifies world simulation and action prediction.

Commercial Availability

Early access partners, including ETH Zurich, the Stanford Robotics Center, the Allen Institute for AI, and UC San Diego, are currently testing the platform. The open stack allows institutions to adapt the base design for specific domains, including healthcare robotics research.

Unitree will make the H2 Plus reference robot commercially available for purchase in late 2026. The accompanying foundation models and software packages are hosted on Hugging Face under the OpenMDW-1.1 license.

If you develop autonomous robotic systems, this reference architecture shifts your engineering bottleneck from hardware integration to behavior policy training. Building reliable workflows will now depend on how efficiently you can evaluate and refine physical skills inside Isaac Lab before deploying weights to the Jetson Thor compute module.