liguang0115/overseer-minecraft

OverSeer Minecraft dataset

Project page | Paper: arXiv link to be added

Two-player Minecraft gameplay seen from eight synchronised views at once: each player's own view and six fixed auxiliary cameras placed around them. This is the Minecraft training data of OverSeer: Multi-Agent World Models Beyond What Agents See.

The episodes were recorded with SolarisEngine, extended with the auxiliary cameras: two Mineflayer bots play on a Minecraft 1.21 server and every view is rendered by a real Minecraft client, so the frames are the game's own graphics and every frame carries the bot's ground-truth action and pose. The files follow the layout of the Solaris training dataset: the Solaris loader reads the two players' videos and actions as they are, and the auxiliary views sit next to them under their own names.

Contents

Episodes	Views per episode	Frames per episode	Size	Shards
38,780	8 (2 players + 6 auxiliary)	166 to about 1,000 (8 to 50 s)	1365 GB	1387

• Every view is 1280 x 720 H.264 at 20 frames per second, copied byte for byte from the recording.
• Outdoor terrain worlds. 18,780 episodes (minecraft_000000 to minecraft_018779) record the full action set and are labelled with one of eleven interaction modes (build, revisit, fight, coverage, chase, pve, explore, mine, follow, circle, reunite) in episode_type; the other 20,000 are longer exploration walks that record movement and camera only (jump, dig and the other interaction channels read as false). The actions column of episodes.csv says which.
• Six auxiliary cameras per episode, static. They are placed in a second pass: the episode is played once to establish the terrain, then the cameras are sampled at random in the free space around the agents' starting area (median 4.9 blocks above the agents and 3.0 blocks from their path, tilted about 52 degrees down). Each camera's _episode_info.json holds its pose as the engine recorded it: x, y, z are the camera entity's feet, its eye is eye_height_blocks (1.62) higher, and it renders with a 100 degree vertical field of view (vertical_fov_deg; fitted against the rendered frames: the 70 degrees stored by the engine, kept as fov_deg_recorded, is not what the cameras use). Every auxiliary video covers the whole episode; spawn_frame is copied from the engine's metadata as recorded.

Files

The repository holds tar shards of about 1 GB, as Solaris does. Extract them with the included script (or tar -xf):

hf download liguang0115/overseer-minecraft --repo-type dataset --local-dir overseer-minecraft
python overseer-minecraft/unshard_dataset.py --shards overseer-minecraft/minecraft_sharded --out datasets/minecraft

The dataset becomes one flat folder, datasets/minecraft/train/, with these files per episode (<episode> is e.g. minecraft_000042):

File	Content
<episode>_Alpha_instance_000.mp4	player 1's view
<episode>_Alpha_instance_000.json	player 1's per-frame pose and action (format below)
<episode>_Alpha_instance_000_episode_info.json	episode and view description
<episode>_Bravo_instance_000.{mp4,json}, ..._episode_info.json	player 2
<episode>_Aux1_instance_000.mp4 ... Aux6	the six auxiliary views
<episode>_Aux<k>_instance_000_episode_info.json	the camera's fixed pose (x, y, z, yaw, pitch), eye height, field of view

episodes.csv lists every episode with its shard, frame count, interaction mode, action set and the name it had in our recordings (source_episode).

Per-frame JSON

A list with one entry per video frame, in the format of Solaris's training data:

{"frame_count": 120, "renderTime": 6000.0, "relativeTimeMs": 6000.0,
 "x": 191.62, "y": 49.62, "z": -97.54, "yaw": 214.85, "pitch": 0.0,
 "action": {"forward": true, "back": false, "left": false, "right": false, "jump": false, "sprint": false,
            "sneak": false, "camera": [0.07, 0.0], "attack": false, "use": false, "mount": false,
            "dismount": false, "place_block": false, "place_entity": false, "mine": false,
            "hotbar.1": false, "...": false, "hotbar.9": false},
 "extra": {"swing": false, "hotbar_switch": false, "held_item_changed": false, "held_item": null}}

• x, y, z in blocks, the player entity's position as the server records it; yaw, pitch in degrees in Minecraft's own convention (yaw grows when turning right, pitch grows looking down). The players see with Minecraft's default 70 degree field of view.
• action at index t is the input applied between frame t and frame t+1: Solaris's loader pairs action[t] with video frame t+1. (The raw recording stores it one index later, with the frame it produced; the release shifts it so both of our datasets and Solaris agree. Add one to the index to get the raw alignment back.) The last frame's action is empty.
• renderTime / relativeTimeMs: milliseconds since the first frame (frame index x 50), the timestamps Solaris's batch sampler aligns the two players on.

Key	Type	Meaning
forward, back, left, right	bool	movement keys (W / S / A / D)
jump, sprint, sneak	bool
camera	[float, float]	change of view direction (yaw, pitch) in radians; + yaw = turn left, + pitch = look up (mineflayer's convention, the same as Solaris)
attack	bool	left click on an entity
use	bool	right click / interact
mine	bool	digging a block
place_block	bool	placing a block
mount, dismount, place_entity	bool	always false (not recorded)
hotbar.1 ... hotbar.9	bool	always false (slot switches are recorded without the slot number, see extra)

extra keeps what has no Solaris key: arm swings, hotbar-switch and held-item-change events, and the name of the item in hand at frame t (null = empty hand). Frames past the end of a short action recording, if any, have an empty action and "extra": {"recorded": false}; their episode's _episode_info.json says "actions_complete": false.

Using it

With the Solaris code base, point a multiplayer dataset at the unsharded folder (bot1_name: Alpha, bot2_name: Bravo, train_dataset_name: minecraft/train): it pairs the two players' videos and actions and ignores the auxiliary views. To read every view of an episode:

from load_episode import load_episode          # included in this repository
episode = load_episode("datasets/minecraft/train", "minecraft_000042")
episode["Alpha"]["actions"][120], episode["Aux3"]["info"]["pose"], episode["Aux3"]["frames"][120]

Citation

The citation will be added here together with the arXiv version of the paper.

The data was collected with SolarisEngine; please also cite Solaris:

@article{savva2025solaris,
  title   = {Solaris: Building a Multiplayer Video World Model in Minecraft},
  author  = {Georgy Savva and Oscar Michel and Daohan Lu and Suppakit Waiwitlikhit and Timothy Meehan and Dhairya Mishra and Srivats Poddar and Jack Lu and Saining Xie},
  journal = {arXiv preprint arXiv:2602.22208},
  year    = {2025}
}