NVIDIA's Project GR00T, Running on the Jetson Thor, Aims to Deliver Embodied AI for Humanoid Robots
Built on new capabilities coming to the Isaac platform, GR00T looks to deliver natural-language understanding and rapid skill transfer.
NVIDIA has announced Project GR00T, designed to provide a general-purpose foundation model for humanoid robots — and primarily built with the company's new Jetson Thor computer in mind.
"Building foundation models for general humanoid robots is one of the most exciting problems to solve in AI [Artificial Intelligence] today," claims Jensen Huang, NVIDIA founder and chief executive officer of the company's latest robotics offering. "The enabling technologies are coming together for leading roboticists around the world to take giant leaps towards artificial general robotics."
Project GR00T, a backronym from Generalist Robot 00 Technology and in no way related to the Marvel character of limited vocabulary, aims to offer natural language understanding and the ability to emulate movements through observation of humans — both key to delivering humanoid-style robots that are at once more human-like and more skilled than current offerings. In demonstrations during Huang's keynote at the 2024 GPU Technology Conference (GTC) GR00T was showcased driving robots to carry out a range of tasks.
GR00T, unsurprisingly, requires computational grunt, and it's here that NVIDIA's hardware offering comes in: Jetson Thor, a new entry in the Jetson family of systems-on-modules (SOMs) that uses the company's latest Blackwell graphics architecture and a transformer engine to deliver 800 teraflops (TFLOPS) of eight-bit floating-point compute. Combined with a high-performance CPU cluster, 100GB of Ethernet bandwidth, and a functional safety coprocessor Jetson Thor is being positioned by the company has the ideal platform for robotics with embodied AI capabilities.
GR00T itself is built atop the company's existing Isaac robotics platform, which has been upgraded in turn to include the Isaac Lab for GPU-accelerated reinforcement learning, the OSMO compute orchestration service, and the Isaac Manipulator and Isaac Perceptor pre-trained models an supporting reference hardware, designed for faster robot arm performance and improved multi-camera 3D computer vision respectively.
While NVIDIA is showcasing Project GR00T, and has named companies including Boston Dynamics, Sactuary AI, Unitree Robotics, and XPENG Robotics as partners in the effort, it has not yet put a firm date on its availability beyond promising the new Isaac features "in the next quarter."
More information is available on the Project GR00T landing page.
Freelance journalist, technical author, hacker, tinkerer, erstwhile sysadmin. For hire: freelance@halfacree.co.uk.