Humanoid robotics just got more real.
A research team led by NVIDIA Corp. has raised the curtain on DreamDojo, an ambitious artificial intelligence (AI) system that teaches humanoid machines how to navigate the physical world by watching tens of thousands of hours of human video.
The research, which could break a long-standing bottleneck of robotic development and usher in an era of major breakthroughs, is a collaboration between NVIDIA and institutions including Stanford University, the University of California-Berkeley, and University of Texas, introduces what the team calls the first “robot world model” capable of generalizing across diverse environments and objects.
DreamDojo bypasses the slow, expensive process of manual robot training by leveraging a staggering 44,000 hours of human-centric video, according to researchers.
Spearheading the system is DreamDojo-HV dataset, which offers 15 times longer duration than previous datasets, nearly 100-fold more distinct skills represented, and 2,000 times more unique scenes than the previous gold standard.
Robots typically require robot-specific data, and information gathered while the machine is physically moving. This makes training brittle and localized. DreamDojo operates in two stages: first, it acquires comprehensive physical knowledge by watching humans interact with the world; second, it fine-tunes that knowledge for specific hardware, such as the GR-1 or AgiBot humanoid platforms.
Beyond mere observation, DreamDojo achieves a technical milestone in speed. Through a distillation process, the system can simulate interactions at 10 frames per second (FPS) for over a minute. This allows for real-time “action-conditioned rollouts,” essentially letting a robot imagine the outcome of its movements before it makes them.
For enterprises, this could be a gamechanger. DreamDoJo allows for “reliable policy evaluation” — testing a robot’s logic in a simulator — without the risk or cost of a real-world crash.
The release comes as NVIDIA CEO Jensen Huang doubles down on robotics as the next pillar of AI, or what he calls physical AI.
At the World Economic Forum in Davos earlier this year, Huang described the current era as the “largest infrastructure buildout in human history,” noting that capital expenditures from tech giants could reach $660 billion this year.
“AI robotics represents a once-in-a-generation opportunity,” Huang said. He predicted the next decade will be a “critical period of accelerated development.”
The financial landscape supports Huang’s optimism. Robotics startups raised a record $26.5 billion in 2025. With companies like Tesla Inc. claiming that much of their future value lies in the Optimus humanoid, the race to build a general physical intuition for machines has become a global priority.
DreamDojo signals NVIDIA’s definitive shift from a gaming hardware company to a “robotics powerhouse.” While the code release timeline remains unspecified, the message is clear: NVIDIA believes the future of computing is physical.
