What happened?
On October 13, 2026, Google released Gemini 4 Argon and showed that it could teach a humanoid robot a new skill by watching a video Google blog ↗ Google Developers ↗. Argon watched a person play cornhole, built a physics simulation of the scene, and then coached Apptronik's Apollo humanoid, which sank seven shots in a row on real hardware Google blog ↗.
Teaching a robot a new physical skill usually takes weeks. Engineers build a simulation by hand, tune it to match the real world and collect many demonstrations. Google says Argon compresses weeks of manual simulation work into a few hours Google blog ↗.
In Google's internal test, Argon watched a raw video of a human throwing bean bags, mapped the positions of everything in the scene and worked out the whole body joint movements needed for the throw Google blog ↗. It then wrote the simulation code itself so the robot could practice. When the result was moved onto Apollo, the robot quickly made seven consecutive shots Google blog ↗.
Google calls this a Video to Sim to Real pipeline Google blog ↗. The same model also scored 87% on ERQA, a test of physical and spatial reasoning, which Google says is higher than every other frontier model Google blog ↗ Google Developers ↗.
Argon is a general model, not a robot product. It launched the same day for developers and paying Gemini subscribers Google blog ↗.
How does it work?
Argon combines three skills: understanding video, writing code and reasoning about physics. Given a video, it identifies the objects that matter, estimates where they are in 3D and builds a matching scene in MuJoCo, a widely used robot physics simulator Google Developers ↗.
All three figures are Google's own reporting Google blog ↗ Google Developers ↗.
The robot then practices in that simulated copy of the real world. Because the simulation is generated from footage, any ordinary video can in principle become training material Google blog ↗. Google's developer post shows the same approach starting from a robot camera feed, a person's head mounted video or even a hand drawn sketch Google Developers ↗.
Google also used Argon to create extra training data. Starting from one photo of a real tabletop, it built a calibrated simulation, generated more than 4,500 practice runs and used them to improve a robot control model. Real robot success on Google's test tasks rose from 56% to 75%, and average completion time fell from 35 to 24 seconds Google Developers ↗.
Which fields could this affect?
This is a research and developer step, not a robot you can buy. These connections are our assessment.
Robot developers
Teams can call Argon through the Gemini API today to build simulations and label robot data, which Google says saves weeks of manual work Google blog ↗ Google Developers ↗.
Factories and warehouses
If robots can learn a task from a short video of a worker, new jobs could be taught far faster than by programming each motion.
Home robots
Learning chores by watching people is the kind of ability home robots need, but this demo was one sports task in a lab Google blog ↗.
What has been checked?
The evidence is Google's launch post, its developer post with prompts and videos, and news coverage. Leapscope reviewed these sources; we did not run the robot or the model ourselves.
Shown so far
- Google released Gemini 4 Argon on October 13, 2026, and made it available through its API and Gemini app plans Google blog ↗.
- Google reported that Apollo made seven consecutive cornhole shots after Argon built a simulation from one human video Google blog ↗.
- Google reported an 87% ERQA score and a rise in real robot success from 56% to 75% using simulated data Google blog ↗ Google Developers ↗.
Still unknown
- How many throws Apollo attempted in total and how far away the board was; Google has not said Humanoids Daily ↗ Interesting Engineering ↗.
- Whether the approach works as well on harder, longer tasks such as household chores.
- Whether outside teams can reproduce the robot results; this was an internal Google test Google blog ↗.
Evidence status: Company announcement. Stage: Usable. Gemini 4 Argon is public through the Gemini API; the robot result is Google's own internal test.
From one throw to everyday skills
This is our suggested way to follow the story, not a promised timetable.
Can I use it today?
Developers can use Gemini 4 Argon through the Gemini API, and Google AI Pro and Ultra subscribers can try it in the Gemini app Google blog ↗. The Apollo robot demo is not something the public can use.
A few things you might be wondering
What is Gemini 4 Argon?
Gemini 4 Argon is Google's frontier AI model released on October 13, 2026. It is built for long, complex tasks in coding, science and robotics Google blog ↗.
How did the robot learn cornhole?
Argon watched a video of a person playing, built a physics simulation of the scene, worked out the movements and transferred them to Apptronik's Apollo humanoid, which made seven shots in a row Google blog ↗.
Is the robot for sale?
No. This was an internal Google test. Apptronik's Apollo is an industrial humanoid, and the demo did not come with any product for homes Google blog ↗.
What is ERQA?
ERQA is an embodied reasoning benchmark that tests whether a model can solve physical and spatial problems in 3D scenes. Google reports Argon scored 87% Google Developers ↗.
Can I use Gemini 4 Argon?
Yes. It is available to developers through the paid Gemini API and to Google AI Pro and Ultra subscribers in the Gemini app Google blog ↗.
Go straight to the sources
Checked Oct 8, 2026. The first source is the original announcement or research. Later sources add independent context; background pages do not validate the result on their own.
01Google's launch post, including the cornhole test with Apptronik's Apollo and the 87% ERQA score.
Robotics examples and prompts, including the 4,500 simulated episodes and the 56% to 75% success gain.
News coverage noting that Google has not published a total attempt count.
Video coverage of the demo and its open questions.