MineRL was a reinforcement-learning competition hosted at NeurIPS 2019, organized by researchers from Carnegie Mellon, Microsoft Research, and others. Its central task, ObtainDiamond, asked agents to mine a diamond in Minecraft from scratch - a long-horizon problem requiring hierarchical sub-goals (gather wood, craft tools, dig, smelt) under sparse rewards, where success could take thousands of correct actions in sequence. The defining twist was a strict sample budget: entrants had to learn efficiently rather than brute-force the problem with unlimited training.
To make that possible, the organizers released MineRL-v0, a dataset of over 60 million state-action pairs of human demonstrations, so teams could learn from how people play rather than purely from trial and error. Submitted algorithms were trained from scratch on held-out environment variations for a fixed time on standardized hardware, which kept the contest about method quality rather than compute access. No team obtained a diamond that year, so the same task was carried into later editions of the competition.
MineRL matters because it targeted two of reinforcement learning’s hardest open problems at once: sample efficiency and long-horizon, hierarchical tasks with sparse rewards. Most flashy RL results of the era relied on enormous amounts of simulated experience; MineRL deliberately removed that crutch. For a general reader, the fact that strong teams could not solve a task a child masters in an afternoon is a clear illustration of where machine learning was - and in many ways still is - far from human flexibility.