MineEngine / README.md
ZiyangYe's picture
Dataset card, metadata index and preview images
f663425 verified
|
Raw History Blame Contribute Delete
5.59 kB
metadata
license: cc-by-nc-4.0
pretty_name: MineEngine
language:
  - en
tags:
  - minecraft
  - multiplayer
  - world-model
  - video
  - first-person
  - depth
  - segmentation
  - embodied-ai
  - game
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files: metadata.parquet

MineEngine dataset

Two-player Minecraft episodes recorded with MineEngine, an automated multiplayer data engine. In every episode two bot players act together in one shared world: building, breaking, PvP, fighting mobs, riding, farming, redstone, water and lava. For each player the dataset has the first-person video, the action of every frame and the camera pose that rendered it; for each world, a state log that replays the world exactly; and for every frame a per-pixel white box rendered from that log: block type, block state, face, entity id, depth and dig stage.

The shared world, both players' first-person clips, their white boxes and the per-pixel channels

Episodes 2,952 (two players each)
First-person video 82.5 h (41.3 h per player), 1280 × 720, 20 fps
Episode types 38 (of the 45 in MineEngine)
Biomes 44, plus flat worlds; dawn, day, dusk and night; clear, rain, snow and thunderstorms
White box per frame, 192 × 336 channels + a 384 × 672 white-box video
Size 267 GB, one tar file per episode

Sixteen episode types: real frame over its white box

Files

metadata.parquet                      one row per episode (see below)
episodes/<wave>/<batch>/<episode>.tar one file per episode

Inside <episode>.tar:

<episode>/capture/Alpha.mp4                 first-person video, 1280 × 720, 20 fps
<episode>/capture/Alpha.camera_meta.json    per-frame camera meta of that video (render time, pose, FOV)
<episode>/capture/Alpha.actions.json.gz     per-frame actions, position, yaw / pitch, player state
<episode>/capture/Alpha.meta.json           recorder meta
<episode>/capture/Alpha.episode_info.json   episode type, environment (biome, time, weather), skins, roles
<episode>/capture/Bravo.*                   the same for the second player
<episode>/capture/world.events.jsonl.gz     world-state log: block changes, entities, damage, digging ... per tick
<episode>/capture/world.snapshot_start.json.gz, world.snapshot_end.json.gz
<episode>/whitebox/episode.json, states.json, types.json, world.npz
<episode>/whitebox/Alpha/gbuf.npz           white-box channels [F, 192, 336]
<episode>/whitebox/Alpha/cams.npz           camera per frame (c2w, eye, yaw, pitch, vfov, t_ms, pair_ok)
<episode>/whitebox/Alpha/proxy.mp4          the white box as a video, 384 × 672
<episode>/whitebox/Alpha/state.jsonl.gz     per-frame state: the player, the other player, every entity, events
<episode>/whitebox/Bravo/...

The white-box channels (logdepth, shade, edge, pid, btype, bstate, face, fx, block) are documented in docs/whitebox_channels.md; the capture files in capture/OUTPUT_SCHEMA.md. Frame i of gbuf.npz, cams.npz, proxy.mp4 and state.jsonl is the same moment.

metadata.parquet columns: episode, path, wave, batch, episode_type, category (build / destroy / fight / move / interaction), world (normal / flat), biome, time_of_day, weather, seconds, valid, size_bytes, sha256. valid is false for the 44 episodes in which a bot reported an error during the episode; their files are complete as recorded.

Usage

import io, json, gzip, tarfile
import numpy as np
from huggingface_hub import hf_hub_download

path = hf_hub_download("ZiyangYe/MineEngine", "episodes/w1/w1_j0_b0/w1_j0_b0_000013_000.tar", repo_type="dataset")
with tarfile.open(path) as tf:
    ep = "w1_j0_b0_000013_000"
    gbuf = np.load(io.BytesIO(tf.extractfile(f"{ep}/whitebox/Alpha/gbuf.npz").read()))
    actions = json.loads(gzip.decompress(tf.extractfile(f"{ep}/capture/Alpha.actions.json.gz").read()))
    tf.extract(f"{ep}/capture/Alpha.mp4", "out")          # the first-person video
print(gbuf["btype"].shape, len(actions))

Select episodes with metadata.parquet, e.g. all PvP episodes at night:

import pandas as pd
m = pd.read_parquet("hf://datasets/ZiyangYe/MineEngine/metadata.parquet")
m[(m.category == "fight") & (m.time_of_day == "night")]

The white box can also be re-rendered from the logs with the MineEngine code (python -m mineengine.cap2proxy), for example at another resolution.

How it was made

Captured with MineEngine on headless Linux GPU servers, four worlds per RTX 6000D GPU, in two waves (October 2026). Each world runs a Paper 1.21 server with the MirrorBot plugin, two Mineflayer bots and two Minecraft clients with the capture mod. The white boxes were rendered with mineengine.cap2proxy from the world-state logs. Player skins are MineEngine's procedural skins and the Minecraft default skins.

Licence

The dataset is released under CC BY-NC 4.0. The Minecraft content visible in the videos and images is © Mojang / Microsoft and is shown under the Minecraft Usage Guidelines. NOT AN OFFICIAL MINECRAFT PRODUCT. NOT APPROVED BY OR ASSOCIATED WITH MOJANG OR MICROSOFT.