I Am GR00T: NVIDIA's Humanoid Robot Development Platform

For decades, sci-fi has promised us truly capable, general-purpose humanoid robots. So, what’s taking so long? There are many unsolved problems in robotics that could take the blame for the delay, but the biggest problem of all may be accessibility. Robotics labs frequently spend years and millions of dollars troubleshooting proprietary actuators, balancing algorithms, and bespoke tactile hands to build a usable platform before they can even start working on higher level functions.

Relatively few people have the time and resources to take on an effort like this. In the world of robotics, there just aren’t enough cooks in the kitchen. People need the freedom to experiment with a technology to improve it, and we urgently need more ideas and experimentation in this area. Fortunately, there may be some hope on the horizon. More accessible development kits are being released than ever before, such as NVIDIA’s Isaac GR00T Reference Humanoid Robot.

A unified reference platform

Announced at NVIDIA GTC Taipei, the new platform is designed to give academic researchers and robotics developers a complete humanoid robotics stack out of the box. Instead of forcing teams to independently assemble hardware, simulation software, AI models, data pipelines, and onboard compute power, NVIDIA is packaging the entire ecosystem into a unified reference design.

The system combines a Unitree H2 Plus humanoid chassis with dual Sharpa Wave tactile five-fingered hands, creating a machine with 75 total degrees of freedom. Standing nearly six feet tall and weighing around 150 pounds, the system is built for human-scale research tasks ranging from locomotion to dexterous manipulation.

Hardware, power, and performance

To support real-world experimentation, the platform includes a wide array of sensors, including stereo-vision cameras mounted in the head, wrist-mounted cameras for close-range manipulation tasks, and an inertial measurement unit for motion tracking. NVIDIA says the robot can deliver up to 360 Newton-meters of torque in the legs and carry payloads of up to 15 kilograms at peak capacity.

NVIDIA’s own Jetson AGX Thor T5000 computer powers the system. Built around a Blackwell GPU architecture, the module delivers up to 2,070 FP4 teraflops of AI performance while operating within a configurable 40- to 130-watt power envelope. The system also includes a 14-core Arm CPU and 128GB of unified memory for running robot perception, planning, and inference workloads directly on the machine.

The software ecosystem and future availability

NVIDIA’s Isaac GR00T software ecosystem includes tools for teleoperation data capture, robot simulation, policy training, testing, deployment, and open foundation models for humanoid reasoning and multitask behavior. Researchers can train policies in Isaac Sim and Isaac Lab before deploying them onto physical robots using Isaac ROS middleware.

The NVIDIA Isaac GR00T Reference Humanoid Robot is expected to become available from Unitree in late 2026, with GR00T workflows for the smaller Unitree G1 robot arriving sooner through GitHub and Hugging Face.