DeepMind’s Gemini 4 Argon taught Apptronik’s Apollo cornhole from a video

Teaching a humanoid a new move usually means weeks of people building a physics simulation by hand before the robot ever tries it for real. Google DeepMind says its new Gemini 4 Argon model can skip most of that. In an internal test, Argon was given raw video of a person playing cornhole, worked out the layout of the scene and the whole-body joint motions the throw needed, and then wrote its own simulation code.

The behavior learned in that simulation was then moved onto Apptronik’s Apollo humanoid. In a real-world trial, DeepMind says Apollo landed seven shots in a row. The company calls the approach a “Video-to-Sim-to-Real” pipeline and says it shrinks simulation work that normally takes weeks down to a few hours.

The bigger idea is that everyday video could become training material for robots, instead of every skill needing a custom setup. DeepMind also reports that Argon scored 87 percent on ERQA, its benchmark for reasoning about physical and spatial problems in 3D.

Some caution is due. This is a single task shown by the company itself, and the report does not say how many attempts came before the streak or how the pipeline handles messier jobs. Argon is a general-purpose frontier model, and Google is offering it through its AI products and the Gemini API, with the robotics work presented as one part of that launch.