MineRL Human Demonstrations

MineRL is a simulator-paired dataset of human demonstrations in Minecraft, introduced in the paper 'MineRL: A Large-Scale Dataset of Minecraft Demonstrations' (arXiv:1907.13440). It contains over 60 million automatically annotated state-action pairs of human players completing hierarchical tasks such as chopping trees, navigating to a goal, and obtaining an iron pickaxe or diamond. Each recorded trajectory pairs the player's first-person POV video (64x64x3 RGB in v0) with structured game state (inventory counts, equipped item, compass angle) and the low-level player actions (movement, camera/mouse, attack, craft, equip, smelt), plus the per-timestep reward from the task's reward function. The data is exposed through the minerl Python package as OpenAI Gym Dict observation/action tuples of the form (current_state, action, reward, next_state, done) via minerl.data.make(...) and the BufferedBatchIter. The original 2019 MineRL-v0 data mirrors went offline in June 2024, and backup copies of the four primary task datasets were re-published on Zenodo (record 12659939) in July 2024.