Google says Gemini 4 Argon taught humanoid robot to play cornhole

Google said its Gemini 4 Argon model helped an Apptronik Apollo humanoid robot make seven consecutive cornhole shots in a real-world trial after analyzing a video of a human playing the game.

The demonstration appears in Google’s October 13 announcement of Argon. The company described it as an internal test of the model’s ability to learn a physical skill outside its training distribution by watching video.

According to Google, Argon mapped the environment’s spatial coordinates, calculated whole-body joint movements and generated simulation code. The resulting work was transferred to the Apollo robot for the physical trial.

Google calls the process Video-to-Sim-to-Real. It says the workflow reduced simulation modeling that would normally take weeks to a few hours. The announcement does not establish how the result generalizes beyond the reported test.

The company also reported an 87% score on the Embodied Reasoning benchmark, which evaluates understanding of physical and spatial problems in three-dimensional environments. Google said Argon outperformed other frontier models on that test.

Argon is being made available through Google’s developer, enterprise and consumer AI products. Google’s announcement places the robotics experiment alongside examples in coding, scientific research and cybersecurity.