NVIDIA Launches Project GROOT for Humanoid Robotics at GTC 2026: The Embodied AI and Thor Revolution

On March 17, 2026, during the grand opening keynote of GTC 2026 at the SAP Center in San Jose, California, NVIDIA CEO Jensen Huang walked onto the stage flanked by a diverse cohort of bipedal humanoid robots demonstrating fluid physical locomotion. Amid sustained applause, Huang announced the commercial mass-production milestone of Project GROOT (Generalist Robot 00 Technology)—a multimodal foundation model architecture engineered exclusively for humanoid robotics—coupled with the volume production of the DRIVE Thor centralized superchip. This breakthrough signified the historic emergence of Physical AI, liberating artificial intelligence from screens and text buffers into the kinetic physical universe.

Project GROOT: The Multimodal Foundation Brain for Kinetic Machines

If large language models serve as encyclopedic minds for digital text, Project GROOT operates as the generalist cognitive brain for the physical world:

1. Multimodal Environmental Reasoning and Teleological Planning: Pre-trained on tens of millions of simulated operational hours within NVIDIA Omniverse, GROOT translates natural language imperatives (e.g., “Inspect the assembly line, identify cosmetically blemished housings, and place them gently into the designated bin”) into real-time spatial trajectories, inverse kinematics, and micro-torque force actuation.
2. Whole-Body Imitation Learning: By ingesting human spatial telemetry from spatial headsets and motion-capture rigs, GROOT rapidly assimilates intricate manual dexterity tasks—such as wiring delicate harness looms or torquing micro-fasteners—automatically generalizing the acquired motor skills across diverse robotic morphologies.
3. Isaac Lab Synthetic Reinforcement Learning: Operating inside NVIDIA Isaac Lab, robotics teams simulate decades of physical interaction in parallel across virtual physics environments, allowing physical robots to deploy onto factory floors with pre-trained balance recovery and dynamic collision mitigation.

DRIVE Thor: The 2,000 TOPS Edge Robotics Powerhouse

To run these massive cognitive models on-robot without relying on latency-sensitive cloud connectivity, NVIDIA commenced volume shipments of the DRIVE Thor centralized edge computing processor:

• Integrating next-generation Blackwell-class GPU tensor cores with high-performance Arm enterprise CPU clusters.
• Delivering a staggering 2,000 TFLOPS (2 PFLOPS) of FP4/FP8 neural execution within a ruggedized vehicular envelope.
• A single Thor processor concurrently orchestrates autonomous mobility, obstacle avoidance, manipulator dexterity, and conversational speech models simultaneously.

Worldwide Industrial Manufacturing Deployment

Representatives from leading robotics and automotive pioneers—including Boston Dynamics, Figure AI, Unitree, BMW, and Mercedes-Benz—joined Huang onstage to confirm adoption of GROOT foundation models and DRIVE Thor controllers across 2026 production lines. From automotive powertrain sub-assembly to hazardous materials sorting and automated micro-electronics fabrication, GROOT-enabled workforces began active deployment worldwide.

Historical Perspective: The Dawn of the Physical AI Epoch

GTC 2026 marked the definitive inflection point for physical robotics. By synthesizing full-stack strengths across accelerated silicon, photorealistic Omniverse physics simulation, and foundational multi-modal neural networks, NVIDIA catalyzed the commercial transition of embodied humanoid robotics from laboratory experiments into pervasive industrial realities.