Google says Gemini 4 Argon can control a full-body humanoid
Keerthana Gopalakrishnan says the newly released model picked up a skill it had no data on after a few hours of real2sim2real.
Google released Gemini 4 Argon, and one of the first things a Google researcher said about it was a robotics claim. Keerthana Gopalakrishnan (@keerthanpg) says the model can control a full body humanoid with dexterous hands, and that it improved through real2sim2real (rebuild the real task in simulation, train there, run it back on the hardware) in just a few hours, picking up a skill the team had no data on.
I was personally very mind blown to see this in the lab.
"I was personally very mind blown to see this in the lab," she wrote.
What else she says it does
In a follow-up, Gopalakrishnan listed more. Direct control of armed quadrupeds. Physical understanding good enough to be used for hindsight labeling (going back over collected robot data and writing in what actually happened) and data quality work. Generating simulation environments from real video zero shot, including egocentric video, both human egocentric to robot sim and robot egocentric to robot sim. And reproducing several robotics papers from a single prompt.
Her summary is that Argon is a step change and will have massive impact on robotics. Google published a developer blog laying out the robotics use cases.
Shown versus claimed
This is a Google researcher describing a Google model, and the humanoid result is something that happened inside the lab. Nobody outside has run it. There is no success rate in the sources, no robot platform named, no description of which skill the model had no data on, and no video breakdown of how much of the control loop Argon is actually closing versus what a lower-level policy is doing. The claim that it improved in a few hours of real2sim2real is a claim about speed with no baseline attached to it.
What is checkable is the availability. Sundar Pichai says Gemini 4 Argon is available globally today for developers and enterprises on the paid tier of the Gemini API and the Gemini platform, for Ultra subscribers in Google Antigravity, and for Ultra and Pro subscribers in the Gemini app and Search. So the model itself is out. The robot results sit behind lab walls.
That gap is the interesting part for anyone in the field. Gopalakrishnan's framing is that every roboticist now has this model at their fingertips, which means the sim generation from video, the hindsight labeling and the paper reproduction claims are all things people can start testing this week on their own setups. The full body humanoid control is the one piece they cannot.
Every roboticist now has this model at their finger tips.
Sound like you were at the demo
One email a day. Under three minutes. The demos, the deployments and the part the press release left out.