WALKERSphysics creatures learning to run with evolution strategies, live in your browser

Each generation, 65 copies of a tiny neural network (about a thousand weights) are nudged with random noise and each runs one episode of a deterministic rigid-body sim. The nudges that scored best are averaged into the next brain. No gradients, no backprop. In the heist the quadruped sees with whiskers (rays for walls and for lasers), knows where the gem or the exit is and when the next beam flips, and a curriculum adds the lasers one group at a time.