NVIDIA Unveils Project GR00T & Jetson Thor: Generalist Foundation Models and Blackwell Computing for Humanoid Physical AI

SAN JOSE, CA — August 21, 2026 — In a sweeping move to establish the foundational software and hardware standard for embodied artificial intelligence, NVIDIA has unveiled Project GR00T (Generalist Robot 00 Technology) alongside Jetson Thor, a dedicated system-on-a-chip (SoC) engineered specifically to serve as the onboard brain for next-generation humanoid robots.

1. The "ChatGPT Moment" for Physical AI and Robotics

While large language models (LLMs) transformed digital text and code, physical robotics has historically suffered from the fragmentation of specialized, single-purpose robotic arms and rigid deterministic control routines. Project GR00T bridges this chasm by delivering a multimodal generalist foundation model capable of understanding natural language instructions and translating them directly into dexterous, cross-embodiment motor coordination.

Trained across massive multi-modal datasets incorporating video observations, human motion capture demonstrations, and synthetic simulation data, GR00T enables robots to learn new skills rapidly through imitation and natural interaction with physical objects.

"Building foundation models for general-purpose humanoid robots is one of the most exciting problems to solve in AI today. The convergence technologies are coming together, enabling the world's leading roboticists to take giant leaps toward artificial general robotics."

2. Jetson Thor: 800 TFLOPS of Blackwell Edge Computing

Powering real-time physical AI requires extreme compute performance within strict power, thermal, and weight constraints suitable for bipedal robotic chassis. To meet this demand, NVIDIA created Jetson Thor, built on the cutting-edge NVIDIA Blackwell architecture:

• 800 Teraflops of FP8 AI Compute: Delivers unconstrained local inference capacity to run complex transformer-based vision-language-action (VLA) models without reliance on high-latency cloud uplinks.
• Functional Safety Co-Processor: Integrated ASIL-D functional safety clusters monitor kinematic limits, collision boundaries, and sensor redundancy in real-time.
• 100Gb Ethernet & Ultra-High Sensor Bandwidth: Ingests synchronous multi-camera 360-degree vision streams, high-frequency LiDAR scans, and tactile fingertip arrays with sub-millisecond bus latency.

3. Industry-Wide Adoption Across Leading Robotics OEMs

Rather than building closed hardware silos, NVIDIA is partnering with the world's premier humanoid robotics developers. Global innovators including Boston Dynamics (powering the electric Atlas), Unitree, Agility Robotics, Apptronik, and 1X Technologies are integrating Project GR00T and Jetson Thor into their commercial production pipelines.

By providing open reference humanoid architectures and synthetic data generation workflows through Isaac GR00T Reference Designs, NVIDIA is accelerating the transition from bespoke laboratory prototypes to ubiquitous industrial, warehouse, and healthcare assistants.