NVIDIA has built a humanoid robot nearly two meters tall to dominate that sector as well
NVIDIA has announced during the "Isaac GR00T Reference Humanoid Robot" event its first open-source reference humanoid robot based on the NVIDIA Jetson Thor ecosystem and the Isaac GR00T development platform.
The stated goal is to lower the entry barriers for research on humanoid robots by providing an integrated architecture that covers the entire development chain: from data acquisition to simulation, and deployment on real hardware.
The project arises from the need to overcome the fragmentation that currently characterizes humanoid robot development. Research teams often have to manage hardware integration, data collection, simulation, model training, evaluation, and deployment separately.
The reference design proposed by NVIDIA attempts to unify these workflows into a single coherent platform. From a mechanical perspective, the system is based on the Unitree H2 Plus, a human-scale robot approximately 1.83 meters tall and weighing about 68 kg, equipped with 31 degrees of freedom distributed throughout its body. Additionally, it includes Sharpa Wave tactile hands, with five fingers, each having 22 additional degrees of freedom, for a total of 75 degrees of freedom between the body and hands. The hands provide tactile feedback, an important feature for fine manipulation tasks.
The sensory system includes a front stereo camera with a field of view of 140 degrees horizontally and 102 degrees vertically, wrist-mounted cameras for close manipulation, and an inertial measurement unit for motion tracking. On the actuator front, the arms can achieve a maximum torque of 120 Newton-meters, while the legs can reach 360 Nm, with a nominal load for the arms of 7 kg and a peak of 15 kg.
The computational brain of the system is the NVIDIA Jetson AGX Thor T5000 module, based on Blackwell architecture. The board integrates a GPU with 2,070 teraflops in FP4, a 14-core Arm processor, 128 GB of unified memory, and a configurable power range between 40 and 130 watts. This ensures real-time inference capabilities directly on the robot without the need for remote processing for critical control and perception functions.
Connectivity is ensured by Ethernet, Wi-Fi 6, Bluetooth 5.2, and USB interfaces, along with microphones and speakers for voice interaction. The battery has a capacity of 972 Wh (15 Ah) and provides about three hours of operational autonomy. There is also a remote emergency stop system.
In parallel with the hardware, NVIDIA provides a complete software platform. The package includes Isaac Teleop for demonstration data collection via teleoperation, open-source foundational models Isaac GR00T for reasoning and multitask learning, Isaac Sim and Isaac Lab simulators for training and validating policies, Isaac ROS middleware for transferring trained models to the physical robot, and finally Jetson Thor for on-board inference.
The modular architecture allows teams to adopt the entire platform or integrate individual components into existing pipelines. Training data, telemetry, and logs remain under the control of the researchers.
NVIDIA has also confirmed that the Isaac GR00T platform will support the Unitree G1 robot, expanding the user base to one of the currently most widely used models in academic environments. The reference workflow for the G1 will soon be available on GitHub and Hugging Face.
The reference design has already been adopted by some of the leading research institutions in the field. The Allen Institute for AI (Ai2), ETH Zurich, Stanford Robotics Center, and the Advanced Robotics and Controls lab at UC San Diego will utilize the platform to develop research in locomotion-manipulation, policy learning, and generalized physical intelligence. NVIDIA Research will likewise employ the design to advance the GR00T models and frameworks.
The robot will be available through Unitree in the second half of 2026.