Google DeepMind uses Gemini 4 Argon to teach Apptronik’s Apollo humanoid new skills from video

Google DeepMind has shown a new way to train robots. Its Gemini 4 Argon AI model learns a physical task by watching a video. It then uses what it learned to control Apptronik’s Apollo humanoid robot.

This demonstration is part of the launch of Gemini 4 Argon, Google’s newest advanced AI model. Google says the model can understand what it sees, understand how things work in the physical world, and write computer code on its own for robot tasks.

In an internal test, Argon watched an unedited video of a person playing cornhole, a game in which players throw bags toward a target. The model studied the surroundings, worked out where things were in space, and calculated the joint movements the robot would need to carry out the task with its whole body.

Argon then wrote computer code to create a simulation on its own. The behavior it learned was then transferred to Apptronik’s Apollo humanoid robot.

Google DeepMind says Apollo then successfully made seven shots in a row during a real-world test.

The company calls this method a “Video-to-Sim-to-Real” process. It says the method can cut a robot simulation task that would normally take weeks of manual work down to just a few hours.

Google DeepMind says: “This ‘Video-to-Sim-to-Real’ pipeline compresses weeks of manual simulation modeling into a few hours, transforming ordinary videos into a scalable, infinite training ground for robotics.”

The demonstration aims to show that general-purpose AI models could reduce the amount of hands-on engineering needed to teach robots new physical skills.

Google DeepMind also tested Argon using its Embodied Reasoning benchmark, or ERQA. This test measures how well a model can solve problems about physical objects and space in three-dimensional environments. The company says Argon scored 87 percent.

Gemini 4 Argon is also designed to handle difficult tasks in engineering, science, coding, and business. Google says it has already used the model in projects such as improving large software codebases, quantum computing, nuclear fusion, and life-sciences research.

The model is now available through Google’s AI products and Gemini API. Google presents the robot demonstration as part of a wider effort to use advanced AI reasoning to work with physical systems.