Google DeepMind releases Gemini 4 Argon, which can teach the Apollo humanoid robot new skills from video
Google DeepMind releases Gemini 4 Argon and demonstrates the model learning physical tasks by watching videos, then transferring them to Apptronik's Apollo humanoid robot. In internal tests, after watching a video of beanbag throwing, Argon analyzed the environment, mapped spatial coordinates, and computed full-body joint motions, autonomously generating simulation code; Apollo hit the target seven times in a row in real trials. Google DeepMind says this Video-to-Sim-to-Real pipeline can compress simulation modeling that originally took weeks into hours, Argon scored 87% on the embodied reasoning benchmark ERQA, and the model has been opened through Google's AI products and the Gemini API.
Why read this
SelectedAI score 82
The demo shows the complete pipeline from raw video to simulation code to real-machine execution, from which one can judge the degree of involvement of general models in robot training.
Topics
Body
Google DeepMind has demonstrated a new robotics training technique in which its Gemini 4 Argon AI model learns a physical task by watching video and then transfers the resulting behavior to Apptronik’s Apollo humanoid robot.
The demonstration forms part of the launch of Gemini 4 Argon, Google’s latest frontier AI model, which the company says combines visual-semantic reasoning, physical understanding and autonomous code generation for robotics applications.
In an internal test, Argon was shown raw video of a person playing cornhole. The model analyzed the environment, mapped its spatial coordinates and calculated the whole-body joint movements required to perform the task.
Argon then autonomously generated simulation code before the learned behavior was transferred to Apptronik’s Apollo humanoid.
According to Google DeepMind, Apollo subsequently landed seven consecutive shots during a real-world trial.
The company describes the approach as a “Video-to-Sim-to-Real” pipeline and says it can reduce a robotics simulation process that would normally require weeks of manual work to 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 is intended to show how general-purpose AI models could reduce the amount of manual engineering required to teach robots new physical skills.
Google DeepMind also tested Argon on its Embodied Reasoning benchmark, or ERQA, which measures a model’s ability to reason about physical and spatial problems in 3D environments. The company says Argon achieved a score of 87 percent.
Gemini 4 Argon is being positioned more broadly for complex engineering, scientific, coding and business tasks. Google says it has already used the model for projects ranging from optimizing large software codebases to quantum computing, nuclear fusion and life-sciences research.
The model is now being made available through Google’s AI products and Gemini API, with Google presenting the robotics work as part of a wider effort to apply advanced AI reasoning to physical systems.