A simulated fighter, coached like a real one.
We use reinforcement learning to train a physically simulated humanoid to fight, starting with a Muay Thai stance, with a Muay Thai coach judging whether every result looks right.
Throw a cube at the fighter in a live simulation that runs entirely in your browser, using the same physics and the same trained neural network. Nothing is scripted.
kakuto-io/fighter is a neural network policy for a 1.75 m, 77 kg humanoid with 83 motors, simulated in MuJoCo. It reads 465 numbers about its body 50 times a second and sets the targets for every joint. It is the student that the techniques are distilled into:
The model is free to use with credit to Kakuto.
Techniques come from motion capture of a real martial artist. Teachers learn one technique each, their skills are distilled into a single student, and a human checks the result at every step. Successes, failures and costs are all published.