Gemini 4 Argon builds a simulation from a cornhole video, helping a humanoid land seven throws in a row
WesRoth says the highlight is not the seven straight throws but that Argon can build its own practice simulation, which could make teaching robots new tasks much cheaper if it holds up across different objects, distances and scenes.
Google’s Gemini 4 Argon built a simulation from a cornhole video and helped a humanoid land seven throws in a row. The simulation step is what interests me. A model that can construct somewhere to practice could make teaching robots new tasks much cheaper. I’d like to see the same process handle different objects, distances and environments. Seven successful throws make a compelling demo. Repeatable learning across those changes would make this much more consequential.
This post quotes another post: “Gemini 4 Argon learns from video to teach Apollo robot 7 straight cornhole throws”