MineRL

MineRL is a large-scale dataset of human demonstrations collected in the Minecraft video game, designed to advance research in sample-efficient imitation learning, hierarchical reinforcement learning, and long-horizon planning. The dataset contains over 60 million human-player action frames captured at full 20Hz action frequency, representing thousands of hours of gameplay. The primary benchmark task, 'ObtainDiamond', requires players to navigate a complex tech tree spanning multiple hours of gameplay — making it one of the most challenging long-horizon planning benchmarks in embodied AI. The dataset also includes subtask demonstrations for navigation, crafting, and combat. MineRL has driven significant advances in hierarchical RL, imitation learning, and curriculum learning, with annual competitions (MineRL Diamond Challenge) attracting hundreds of research teams worldwide. While not a robotics dataset per se, MineRL serves as a critical proxy environment for embodied AI research due to its rich, open-ended interaction space.