// the find
UnityTechnologies/MachineLearningRoguelike
A small Roguelike game that uses Machine Learning to power its entities. Originally used in talks by Ciro & Alessia.
A Unity demo from 2017 showing ML-Agents v0.2.1 applied to a roguelike — both the player and enemies are trained RL agents navigating a tilemap dungeon. It was built for conference talks, not as a reusable framework, and it shows. This is a historical artifact, not a starting point for a new project.
The separation of training scene from game scene is a genuinely useful pattern — shows how to train agents in a controlled environment and deploy them in a different one. Pre-trained model binaries are included so you can see the result without running training yourself. The combination of Tilemap, Cinemachine 2D, and ML-Agents in one project covers three Unity systems that rarely appear together in examples.
Built against ML-Agents v0.2.1 from 2017 — the current SDK is v2.x and the API is completely different, so nothing here ports directly. The TFSharpPlugin it requires is a dead end; ML-Agents now uses Sentis/Barracuda. No training reward shaping is documented, so you can't tell why the agents behave the way they do or how to improve them. Last commit was 2022, which itself was just minor cleanup — the actual game logic has been frozen since 2017.