Gemini 4 Argon is driving humanoids, Google says
A world model trained on 30,000 hours of first-person video learned exactly where hands go and still could not say what happens to the things they touch. Going to 30,000 hours did not close the gap, and the field spent the day arguing about why.
Shown this week
Google says Gemini 4 Argon controls a full-body humanoid
Keerthana Gopalakrishnan says Google's newly released Gemini 4 Argon can control a full-body humanoid with dexterous hands, and picked up a skill it had no data on after a few hours of real2sim2real (rebuild the task in simulation, train, run it back on hardware).
I was personally very mind blown to see this in the lab.
A model SSH'd into a robot and learned cornhole in 6 tries
Takahiro Miki says he prompted a model to connect to a robot, and it figured out SSH on its own, then read the codebase to work out how to command each joint.
I was impressed how good it can use the classical robotics method as a tool.
Teleop and other arguments
30,000 hours of video taught a world model hands, not objects
Jiahua Dong says his team trained a world model on 30,000 hours of first-person video and it learned exactly where hands go, but not what happens to the things they touch. Going from 300 to 30,000 hours did not narrow that gap, and a model four times larger barely did either.
Sorry man, perception is deception.