Menu

How it works

Little two-legged AI athletes learn football from zero, and the league follows them season by season. Here is how it works, from the first step to a matchweek episode.

Learning by trial and error

Every athlete is controlled by an AI that learned by reinforcement learning. It tries something, gets points for what we want (moving forwards, staying up, scoring) and loses points for what we don’t (falling, walking the wrong way), and repeats that millions of times. Nobody shows it how to walk or kick. There is nothing to copy, only points. What reinforcement learning is, in more detail.

Real physics

The athletes live in MuJoCo, a physics engine used in robotics research. Gravity, friction, momentum and collisions are simulated; the AI only chooses how hard each joint pushes, many times a second. If a player falls, it fell. If the ball goes in, it was kicked in.

From one walker to eight teams

It took 36 tries to get one AI that walks. That AI went to a football school, where players with the same brain learned by playing four-a-side matches against each other. A separate brain learned to get a fallen player back up. Then each of the eight teams got a coach with its own plan, and every team keeps training between matchweeks.

The league’s rules

  • Four a side, with two halves of three minutes.
  • Any touch on the floor by anything but a foot is a fall.
  • The ball counts only off legs or boots.
  • A win is worth 3 points and a draw 1. The table is sorted by points, then goal difference, goals scored, head-to-head points and fewer falls.

How an episode is made

Every match is played out in the physics and recorded. The episode is drawn in Blender from those recordings, frame by frame, with a camera that follows the play. Nothing is animated by hand and nothing is posed: if you see it, it happened. There is no AI-generated video: the AI is the athletes’ brains.

The narrator’s voice is AI-generated; the footage is not. The sounds are CC0.