Google DeepMind demonstrated a video-based training method for Apptronik’s Apollo humanoid using Gemini 4 Argon, first reported by Roboticsandautomationnews. Google DeepMind says the pipeline can cut simulation work that normally takes weeks of manual effort to a few hours, reducing the engineering time needed to teach robots new physical skills.
During an internal test, Argon watched raw video of a person playing cornhole. It analyzed the environment, mapped spatial coordinates and calculated the whole-body joint movements needed to perform the task.
Argon then autonomously generated simulation code, and the learned behavior was transferred to Apollo. Google DeepMind says the humanoid landed seven consecutive shots during a real-world trial.
Google DeepMind calls the method a “Video-to-Sim-to-Real” pipeline. Apollo’s seven-shot result comes from the real-world trial, while the claimed time reduction applies to simulation work.
The demonstration is intended to show how general-purpose AI models could reduce the manual engineering needed to teach robots new physical skills. Google also positions Argon for complex engineering, scientific, coding and business tasks.
Google DeepMind also tested Argon on its Embodied Reasoning benchmark, or ERQA, which measures reasoning about physical and spatial problems in 3D environments. Google says Argon scored 87 percent.
Google says the model has been used on projects ranging from optimizing large software codebases to quantum computing, nuclear fusion and life-sciences research. Gemini 4 Argon is now being made available through Google’s AI products and Gemini API.