In an internal test at Google DeepMind, Gemini 4 Argon watched a raw video of a person playing cornhole and turned it into a working robot skill without anyone hand-building the physics in between.
The model mapped the spatial coordinates of the scene from the footage, calculated the whole-body joint movements the throw required, and then wrote the simulation code on its own. When that work was transferred onto the Apollo humanoid from Apptronik, the robot sank seven consecutive shots in a real-world trial.
What matters most here is the pipeline that Google DeepMind calls Video-to-Sim-to-Real, because it clears the slowest step in robotics out of the way. Building a faithful simulation of a new task has always meant weeks of manual modeling by people who know the physics well, and Google says this compresses that work into a few hours.
Once a recording of someone doing something is enough to start from, the supply of robot training material grows to include the enormous amount of ordinary video of people throwing, lifting, folding and carrying that already exists.
Robotics has spent years short on data, and the most common kind of data in the world is now becoming usable for teaching robots how to move.