Google DeepMind’s Video-to-Sim-to-Real pipeline teaches a humanoid robot to play cornhole

Google DeepMind has taught a humanoid robot to play cornhole. The robot needed one video, some simulated practice, and zero coaching. The model behind it is Gemini 4 Argon, which Google DeepMind unveiled on October 13, 2026, during an internal robotics demonstration. The robot was Apptronik’s Apollo 2, and it sank seven consecutive shots. Backyard games are not the point. The demo is about how fast a machine can turn a human motion it has watched into a physical skill it can perform. From one video to seven straight shots The setup was deliberately sparse. Apollo 2 got a single video of a human throwing a cornhole bag. There was no library of demonstrations and no army of engineers hand-tuning arm angles. Gemini 4 Argon used its video-understanding and reasoning abilities to break down what the human was doing. It then mapped that throw onto the robot’s own joints. That step is harder than it sounds. A human arm and a humanoid robot arm are built differently, so copying a motion means translating it, not just mirroring it. Next, the system built a physics simulation using MuJoCo, a physics engine used to model how bodies move and collide. Apollo 2 practiced its throws inside that v...