Scale and delivery
60M+ annotated state-action pairs reported in the paper
- Formats
- Compressed MineRL archives, MineRL Python data pipeline
- Timebase
- Sequential simulator-aligned trajectories
Modalities
- RGB video
- Keyboard and mouse actions
- Rewards
- Simulator state
Observations
- Player-view pixels
- Inventory and environment observations in supported tasks
Actions
- Keyboard controls
- Mouse movement
- Inventory interaction
State
- Task-specific simulator state
- Inventory
- Position and game variables
Outcomes
- Reward
- Task progress
- Terminal state
Rights note
The Zenodo record labels the archive MIT. That label does not grant trademark rights or replace review of Minecraft terms, third-party rights, and the intended commercial use.
Preferred citation
Guss et al. MineRL: A Large-Scale Dataset of Minecraft Demonstrations. IJCAI 2019.