Six ducks, one call a second.
Three Microducks a side, simulated in MuJoCo on the robot's own walking and kicking policies. Nothing is scripted. Once a second, System 1 is asked what every free duck should do next — walk, turn, step or kick — and where every duck should look, in a single call for both teams. Each duck knows the ball and the other ducks only from what its head camera has seen. The bodies act on the answers while the match plays on.
Simulated in MuJoCo (Google DeepMind, Apache-2.0) — not a sim-to-real result · the robot is Pollen Robotics' Microduck, used here for research: its CAD meshes are CC BY-NC-SA and its walking, kicking and stand-up policies are Apache-2.0 · System 1 chooses every movement and every look once a second; the body's only reflexes are getting up after a fall and settling before a kick · each duck knows its own place on the pitch; the ball and the other ducks it knows only as its head camera has seen them, judged from the camera's view geometrically in place of an image detector · the System 1 layer runs on the decision model chosen above, through OpenRouterHumanoid playground ↗