What is shadow-rogue-self-play?

Shadow-Rogue Self-Play is a 2D game built in Unity where two players dodge projectiles and shoot at each other. The twist is that the AI opponents are trained through reinforcement learning using a technique called self-play. The AI learns by playing against older versions of itself, gradually improving its ability to dodge, jump, crouch, dash, and shoot across a variety of strategies. The game is playable directly in the browser through itch.io. In human versus AI mode, you control one character using keyboard inputs, A and D to move, space to jump, S to crouch, and K to shoot. Both sides can also run fully as AI. The training uses ML-Agents, Unity's reinforcement learning toolkit, with the PPO algorithm. The AI observes its own position, velocity, crouch and dash state, and surrounding sensors that detect bullets, walls, the ground, and the opponent. From that input it picks one of five discrete actions each step. The trained model is stored as an ONNX file and can be dropped back into Unity to run inference. Training scripts let you run multiple simultaneous game environments for faster learning. This is an educational project aimed at people learning reinforcement learning or game AI in Unity. The tech stack is C# in Unity with ML-Agents. Training requires building the project first, then running the ML-Agents command line tool against the built executable. The creator notes the project predates their other polished work and may contain some leftover experimental files.