Google DeepMind trains Apollo humanoid using Gemini 4 Argon
English translation from Unibot. Source published 2026-10-23.
Google DeepMind showed a new way to train robots. The Gemini 4 Argon model masters a physical task simply by watching a video, then transfers the skill to Apptronik's Apollo humanoid.
What Gemini 4 Argon can do
The demonstration was part of the launch of Gemini 4 Argon, Google's latest flagship model. According to the company, it combines visual semantic reasoning, physics understanding, and autonomous code generation for robotics.
How the test went
A test was conducted within the company. Argon was shown an unprocessed video of a man playing cornhole. The model analyzed the situation, estimated the spatial coordinates and calculated how all the joints should move in order to complete the throw.
Then Argon itself wrote the simulation code - and the trained behavior was transferred to the humanoid Apollo. In the real test, Apollo hit seven times in a row, according to Google DeepMind.
Video-to-Sim-to-Real pipeline
The company calls the approach itself a “Video-to-Sim-to-Real” pipeline. It claims to compress robot simulations that would typically take weeks of manual work into just a few hours.
“This Video-to-Sim-to-Real pipeline condenses weeks of manual simulation modeling into a couple of hours and turns ordinary videos into a scalable, infinite training ground for robots,” says Google DeepMind.
Why all this? The demonstration aims to show how general-purpose models can reduce manual engineering when teaching robots new physical skills.
Result on the ERQA benchmark
By the way, Argon was also tested on its own Embodied Reasoning, or ERQA, benchmark. It measures how well a model reasons about physical and spatial problems in a 3D environment. The result is 87 percent.
Where else is the model used?
Gemini 4 Argon is positioned more broadly: complex engineering, science, programming, business tasks. Google has already used the model in projects ranging from optimizing large codebases to quantum computing, nuclear fusion, and life sciences research.
The model is now available through Google's AI products and Gemini API. And the company is pitching its robotics work as part of a broader effort to apply advanced AI thinking to physical systems.