What the Isaac GR00T Humanoid Reference Design Is
The Isaac GR00T humanoid reference design is an open humanoid robot platform that combines a human-scale body, powerful onboard Jetson Thor compute, and the Isaac GR00T software stack into a single, standardized robotics research platform. NVIDIA positions this system as a shared starting point for academic and frontier labs that need reliable hardware and mature software without being locked into proprietary ecosystems. Instead of assembling separate limbs, sensors, and controllers from many vendors, teams receive a coherent body–brain package tuned for physical AI. This focus on openness matters: the mechanical chassis comes from Unitree’s H2 Plus, tactile manipulation from Sharpa Wave five-finger hands, and AI compute from Jetson AGX Thor, all tied together by NVIDIA’s Isaac GR00T development platform. Together, they define a reference humanoid that any lab can study, modify, and build upon.
Standardizing the Humanoid Stack from Body to Brain
Isaac GR00T turns humanoid robotics into a reference system rather than an ad hoc integration project. The Unitree H2 Plus chassis stands nearly six feet tall, weighs about 150 pounds, and offers 31 degrees of freedom for human-scale motion experiments. Dual Sharpa Wave tactile hands add 22 more, bringing the total to as many as 75 degrees of freedom across body and hands, enabling research into fine grasping, tool use, and bimanual tasks. Multi-view sensing—head-mounted stereo camera, wrist cameras, and an inertia measurement unit—gives the open humanoid robot a detailed view of its environment and its own balance. Torque-rich joints (up to 120 Newton-metres in the arms and 360 in the legs) support lifting and reaching with a rated arm payload of 7 kilograms and peaks of 15 kilograms, so labs can test realistic workloads instead of toy-scale demos.
Jetson Thor Compute and the Isaac GR00T Software Brain
At the core of the NVIDIA robotics reference design is Jetson AGX Thor T5000, which acts as the robot’s brain. It combines an NVIDIA Blackwell GPU delivering 2,070 FP4 teraflops of AI performance, a 14-core Arm CPU, and 128GB of unified memory within a configurable 40–130W envelope, all powered by an onboard battery sized for several hours of operation. According to NVIDIA, “the NVIDIA Isaac GR00T Reference Humanoid Robot gives researchers a single, open platform to make breakthrough discoveries toward general-purpose physical intelligence.” Networking and I/O—Ethernet, Wi-Fi 6, Bluetooth 5.2, USB, microphones, and speakers—support voice interaction and high-bandwidth sensing. On top of this hardware, the Isaac GR00T platform supplies open models and workflows for simulation, training, evaluation, and deployment, so teams can focus on robot skills instead of low-level bring-up.
Unitree H2 Plus: From Reference Design to Practical Robots
The Isaac GR00T humanoid is not just a lab curiosity; third-party manufacturers are already using it as a template. Unitree’s H2 Plus, built on the NVIDIA Isaac GR00T development platform, is one of the first commercial embodiments of this design. It packages the H2 humanoid chassis, Sharpa Wave tactile hands, and Jetson Thor compute with Isaac GR00T software and workflows into a reference product that research labs can buy and deploy. This arrangement speeds up the path from receiving hardware to testing skills like locomotion, object manipulation, and whole-body control. By aligning on a common configuration, labs can share datasets, control strategies, and evaluation benchmarks more easily. The H2 Plus evolves the open humanoid robot concept into a ready-to-use robotics research platform that still allows deep customization at the software and control layers.
From Proprietary Robots to a Shared Robotics Research Platform
Open reference designs like Isaac GR00T mark a shift in how humanoid robots are built and studied, echoing how ROS standardized robotics software. Instead of each lab maintaining a proprietary stack, the NVIDIA robotics reference design promotes a shared baseline for hardware, compute, and software. This lowers the barrier for newcomers who can adopt a known configuration, then layer on their own control policies, perception pipelines, or large-model-based planners. It also encourages interoperability: skills developed on one Isaac GR00T humanoid should transfer to others with minimal adjustment. With a human-scale body, tactile hands, and Jetson Thor compute all tied together by an open software stack, the platform aims to accelerate progress toward general-purpose physical intelligence and move humanoids from controlled demos toward repeated, real-world deployment in research and industry.
