PEAK: authoring and training in one loop
An open-source 2D platformer engine where you can design a level and train an agent on it without leaving the tool.
Game engines are built for authoring content; reinforcement learning benchmarks are built for training agents. Wanting both means gluing them together yourself. PEAK is a lightweight, configuration-first engine that keeps level design, agent training, debugging, and playtesting in a single loop. Levels, mechanics, rewards, and observations live in plain text, so changing the game does not mean touching the training pipeline. Runs locally and offline.