Apollo humanoid learns cornhole from a video: seven throws in a row, according to DeepMind

In an internal test, the Argon model watched raw video of a person playing cornhole. The model then analysed the scene, identified positions in space and calculated the movements of each joint in the body needed to reproduce the action.

Argon wrote the code for a simulation on its own, and the learned behaviour was then transferred to the Apollo robot. According to Google DeepMind, Apollo then made seven throws in a row in a real-world trial.

The company says its method, which goes from video to simulation and then to the real robot, cuts to a few hours a modelling job that usually takes weeks by hand.

Google DeepMind also put Argon through its own three-dimensional physical and spatial reasoning test. According to the company, the model scores 87% on it.

The model is now available in Google's artificial intelligence products and through its Gemini programming interface.