TL;DR: DelveRL is a new open-source roguelike with a structured API for training RL agents, including deterministic simulation, partial observability, and a PPO baseline.
Summary: A developer released DelveRL, an open-source endless turn-based roguelike designed specifically for training game-playing agents. It provides a structured API, deterministic simulation, procedural levels, partial observability, and renderer-free batched environments; a recurrent PPO baseline reaches a median floor of 18, with extended runs reaching floor 33. The project includes training code, a checkpoint, and raw benchmarks.
Why it matters: For AI builders, DelveRL offers an accessible, reproducible environment to evaluate agent exploration, risk management, and improvement over a strong baseline. Try pitting your own RL agents against the included PPO checkpoint and see if you can beat floor 33.
Source: reddit