Nvidia's next big AI bet: Cloud-connected humanoid robots

The AI sector has seen tremendous growth in just 4 years, mostly thanks to Nvidia and its powerful GPUs specifically designed for AI workloads. Team Green actually started betting on AI as early as the mid-2010s, but the results only became apparent a few years ago. Now that the hardware side of AI is vigorously covered, even bordering on monopoly, Nvidia is looking for future bets that could further boost its value in the future. This year at Computex, Team Green is signaling that one such bet could be physical AI and humanoid robotics.

As CEO Jensen Huang explained in his keynote earlier today, “physical AI and robotics will bring about the next industrial revolution. From AI brains for robots to simulated worlds to practice in or AI supercomputers for training foundation models, Nvidia provides building blocks for every stage of the robotics development journey.” 
 

Nvidia Isaac Gr00T

With the focus on physical AI, Nvidia is also expanding its forays into the software side. The core component in this regard is the company’s first open humanoid robot OS called Isaac. On top of it, developers can add trained models like the Gr00t N1.0 introduced in March this year. This has been update to version 1.5 that adds the Dreams component. Version 1.0 with the Mimic component constituted the basis to train robot reasoning and behavior, and with the Dreams component, Nvidia is introducing a blueprint that helps generate vast amounts of synthetic motion data (neural trajectories) that physical AI developers can use to teach robots new motor skills, including how to adapt to changing environments.

The training process resembles the text-to-image / video models. First, the robots are post-trained with Cosmos Predict world foundation models (WFMs). Using a single image as the input, GR00T-Dreams then generates videos of the robot performing new tasks in new environments. The blueprint extracts action tokens (compressed pieces of data easily processed by the robot’s neural network), which teach a robot how to perform new actions.

Nvidia’s Isaac GR00T is closely linked with the Omniverse and Cosmos platforms in order to generate training data. The N1.5 update as well as other models like Isaac Sim 5.0 and Isaac Lab 2.2 will soon be available for download on the Hugging Face repository. Pioneering companies that are already using Isaac N models for humanoid robots include AeiRobot, Foxlink, Foxconn, Lightwheel, NEURA Robotics, Boston Dynamics, Agility Robotics, and XPENG Robotics.

Furthermore, Nvidia is offering its Universal Blackwell Systems powered by RTX PRO 6000 GPUs like the DGX Cloud infrastructure, so that developers can easily tap large amounts of processing power for robot training.