crafter

Why we included this project

Crafter is a compact 2D open-world survival environment for reinforcement learning research, where agents forage for food and water, find shelter to sleep, fend off monsters, gather materials, and craft better tools, all inside a single Gym-compatible simulator. Because so much happens in one world, a single run exercises generalization, exploration, representation learning, and long-term credit assignment, which saves you from training across a whole suite of separate environments. Progress is measured through 22 semantically defined achievements and a geometric-mean score, so you can see what an agent actually learned instead of a single opaque reward number. The environment is light enough for fast iteration, ships with baseline scores and analysis scripts, and even has a human-playable GUI for checking behavior by eye. Teams benchmarking new algorithms or studying unsupervised agents will find it a practical testbed that is far cheaper to run than large 3D simulators.