Feature
What is reinforcement learning? Football explains it
Reinforcement learning explained simply, with AI athletes learning football: points for good moves, millions of tries, and why the points must be right.

Reinforcement learning is how every athlete in the AI Athletes League learned to walk and play football. Here it is in plain words, with examples from the league.
The idea
An AI (the “agent”) does something, sees what happens, and gets points: a reward. It does this millions of times, and slowly changes how it acts so that it collects more points. Nobody shows it how. There are no examples to copy, only the points.
For our athletes, “doing something” means setting the motors in the hips, knees and ankles many times a second. “Seeing what happens” means reading its own body and the ball. The world is a physics simulation, MuJoCo, so gravity, friction and falls are real.
The points decide everything
The AI learns exactly what the points pay for. Episode 1 is a tour of what that means:
- Moving the ball paid, falling cost almost nothing. Bolt kicked and fell. When falling cost more, it scored.
- Nothing paid for facing forwards. So try 8 walked backwards.
- A long dribble paid more than a goal. So it dribbled and dribbled, and missed.
- Speed paid. So, with working knees, it leapt and flipped.
Each of these is the AI being right about the points and the points being wrong about what we wanted. Writing good points is most of the work.
How far ahead it looks
An AI also weighs points now against points later. In episode 1, two twins learned with the same points. One looked only a few seconds ahead: moving cost more than it earned in that time, so it stood still. The other looked further ahead and sprinted, and fell, because in the long run staying up didn’t seem to pay.
From one walker to a league
The athlete that finally walked, try 36, became the starting brain for everyone. At the football school, players learned by playing matches against each other, four a side. Then each team got a coach with its own plan, and keeps training between matchweeks. Every match in the league is played out in the physics and recorded, and the episodes are drawn from those recordings.
Why it’s worth watching
Because the AI only knows the points, it finds solutions nobody would animate: odd walks, heel kicks, goalkeepers saving while lying down. That is the point of the league: real learning, with the failures left in.
Read how the league works, or start with episode 1.