Announced at GTC Taipei, the NVIDIA Isaac GR00T Reference Humanoid Robot packages hardware and compute from three separate vendors into a single baseline for physical AI and loco-manipulation research. The reference architecture aims to give laboratories an open alternative to proprietary humanoid systems, combining physical hardware with NVIDIA's simulation, training, and deployment software stack.

Distributed hardware architecture

The physical assembly divides locomotion, manipulation, and onboard processing across distinct commercial partners:

  • Chassis and locomotion: Unitree provides the H2 Plus humanoid chassis. The frame stands nearly 6 feet tall, weighs 150 pounds, and incorporates 31 degrees of freedom across the body. NVIDIA reports arm torque reaching up to 120 Newton-meters, leg torque up to 360 Newton-meters, a rated arm payload of 7 kilograms, and a peak arm payload of 15 kilograms.
  • Tactile manipulation: Sharpa supplies dual Wave five-finger dexterous hands, adding 22 degrees of freedom per hand to bring total articulation across the body and hands to 75 degrees of freedom.
  • Onboard compute: NVIDIA provides the Jetson AGX Thor T5000 module for real-time sensor processing and robot inference. The board features an NVIDIA Blackwell GPU delivering a vendor-claimed 2,070 FP4 teraflops of AI performance, a 14-core Arm CPU, 128GB of unified memory, and a configurable 40- to 130-watt power envelope.
  • Sensing and connectivity: Integrated perception includes a head-mounted stereo camera (140 degrees horizontal by 102 degrees vertical field of view), wrist cameras for close-range tasks, and an inertial measurement unit, backed by Ethernet, Wi-Fi 6, Bluetooth 5.2, USB, microphones, and speakers.

Software integration and rollout timing

The integrated system targets NVIDIA's Isaac GR00T software ecosystem for physical simulation, policy training, and runtime deployment. NVIDIA noted that participating researchers retain ownership and control over their robot data, training sets, telemetry, and logs. Initial participating institutions include Ai2, ETH Zurich, the Stanford Robotics Center, UC San Diego's Advanced Robotics and Controls Laboratory, and NVIDIA Research.

For the complete assembled system, Unitree plans to make the reference humanoid available in late 2026, with commercial pricing unannounced. In the interim, NVIDIA plans to publish Isaac GR00T reference workflows on GitHub and Hugging Face supporting the smaller Unitree G1 humanoid, letting research teams begin developing on the software stack using existing hardware.