AI plays Mini Metro - 2172 passengers 71 days x3, live decisions

Mimo, our AI, plays the real Mini Metro (London) using only screenshots and mouse drags. This is its best game so far: 2,172 passengers in 71 days. The video runs at about 3x, with idle stretches cut. On the right you see each decision as it happens: the question (a new station, a weekly upgrade, a spare train), Mimo's answer, and whether the drag worked. The graph below tracks passengers by day. What happened in this game • Only 3 lines the whole game, with 48 stations by the end • Weekly upgrades: 5 carriages, 3 tunnels, 2 interchanges • How it ended: Mimo misread the screen and thought station S15 had dropped off the red line. It dragged the line over S15 to reconnect it, and in Mini Metro that removes the station. S15 was off the line for about 3 minutes and overflowed. This is now fixed. How it works • Screen in: it reads stations, passengers, trains and lines from the screen. • Decision: a small transformer (Mimo 5) scores every legal move and picks the best one. • Mouse out: it carries out the move with a mouse drag, then checks the screen again. • Training: reinforcement learning (PPO) in a Mini Metro simulator we rebuilt and calibrated to the real game. Full explanation (inputs, network, training): [link to the explainer video] Made by Sumin Han (suminhan@ksa.kaist.ac.kr) Mini Metro is a game by Dinosaur Polo Club. This is an independent research project, not affiliated with the developer. #minimetro #ai #reinforcementlearning