NVIDIA Announces Project GR00T Humanoid Foundation Model and Jetson Thor

NVIDIA announced Project GR00T, a general-purpose foundation model for humanoids, and Jetson Thor, a computing platform based on the company’s NVIDIA Thor SoC. The company says Jetson Thor’s Blackwell-generation GPU delivers 800 teraflops of 8-bit floating-point AI performance. NVIDIA also expanded its Isaac robotics platform, including its simulation and AI workflow infrastructure.

NVIDIA’s announcement brings together computing for humanoids, thousands of parallel simulations, training workflow management, and manipulation and vision capabilities in a single platform family. For developers, a key question will be how this compares with assembling each stage separately. The figures of 800 teraflops and up to 80 times faster, however, do not by themselves indicate a robot’s task performance or safety.

The focus of the announcement is not a standalone model for humanoids, but the integration on Isaac of the computing platform for GR00T, its training environment, workflow management, and manipulation and perception capabilities. The 800-teraflop performance figure for Jetson Thor is NVIDIA’s claim; the company did not specify the processing conditions on a physical robot or how the figure relates to power consumption. The announced computing performance therefore cannot be equated directly with a robot’s task or operational performance. By stage, Isaac Lab provides a simulation-based training environment, while OSMO orchestrates data generation, model training, and software-and-hardware-in-the-loop validation. GR00T then serves as the robot’s action model, and Jetson Thor handles computing on the robot. Isaac Manipulator and Isaac Perceptor appear positioned to supplement general-purpose models with arm manipulation and perception of the surroundings. This integration indicates that NVIDIA’s offerings extend beyond robot development to the computing infrastructure needed for repeated training and validation. At the same time, thousands of parallel simulations and up to 80 times faster path planning are descriptions of platform capabilities, not necessarily demonstrated improvements in actual deployment environments. GR00T is designed to acquire skills by observing human movement and understanding natural language, but the announcement does not specify which tasks it can learn or under what conditions. The announcement also does not make clear how much models and simulations can be shared across different robot platforms or how synthetic data will be connected to real-world data. Questions to watch include Jetson Thor’s power consumption, latency and safety on physical robots; training time and task success rates when using Isaac’s features; and the range of robots and skills supported by GR00T. Even if the announced components are implemented as a single development environment, that would not automatically guarantee that results from simulation-based training can be reproduced on physical robots or transferred to different platforms. Evaluating them will require disclosure of supported conditions and physical-robot validation.

According to NVIDIA, robots powered by GR00T are designed to understand natural language and imitate movements by observing people. Jetson Thor includes a functional safety processor, a high-performance CPU cluster and 100Gb of Ethernet bandwidth. NVIDIA cited running multimodal generative AI models as a use case. Isaac Lab is a lightweight, GPU-accelerated application that runs thousands of simulations in parallel for robot training. OSMO, a computing orchestration service, coordinates data generation, model training, and software- and hardware-in-the-loop workflows in distributed environments. Isaac Manipulator provides features for robotic arms, and NVIDIA says it makes path planning up to 80 times faster. Isaac Perceptor provides 3D, 360-degree vision using multiple cameras. The new Isaac platform features are expected to become available in the next quarter.