Synchronized multiplayer Minecraft video with actions, released with PAWA: Panoramic Action-conditioned World Modeling for Multiple Agents. Every player in an episode has a first-person (POV) video, a world-stabilized equirectangular panorama rendered from the same camera origin, and a per-frame record of that player's actions and camera pose. All streams of an episode share the server clock.

At a glance

	episodes	player views	episode hours	player hours
Full release	8162	45716	278.5	1554.1
Recommended training subset	5387	26538	184.0	903.1

Episodes by number of players: N=1: 350 · N=2: 2555 · N=4: 2770 · N=8: 1327 · N=16: 1160. Episodes last about two minutes (median 122.1 s). A player view is one player's POV stream, panorama stream and action record.

POV video: 640×360, H.264, 20 fps.
Panorama video: 1024×512 equirectangular, H.264, 20 fps, world-stabilized (it does not turn with the player's head), with spherical-video metadata.
Actions: keyboard state, camera deltas, and discrete actions (attack, use, place, mine, ...), one row per video frame, plus absolute position and camera pose.

Interactions are scripted scenarios (building, following, meeting, pursuit, revisiting landmarks, crowds, ...) executed by headless players on a Fabric server and rendered offline with ReplayMod; see the pipeline repository for how episodes are produced.

Layout

dataset/<YYYYMMDDTHHMMSSZ>-<scenario>-<N>p-<hash8>/
  episode.json               episode metadata; povs[].videos.*.path is authoritative
  qc.json                    behavioral QC metrics and verdict
  events.jsonl               sparse semantic action events
  pNN-default.mp4            player NN's POV video
  pNN-equirectangular.mp4    player NN's world-stabilized panorama
  pNN-frames.jsonl           one row per video frame: action + camera pose
  pNN-episode_info.json      per-player summary
  diagnostics.tar.gz         (6642 episodes) raw capture: ReplayMod recordings, action/event streams
docs/                        field reference and timing model
manifest.jsonl               one row per episode (see below)
surface-audit.jsonl          one row per player view (see below)

Row i of every pNN-frames.jsonl in an episode corresponds to frame i of that player's two videos and to the same server tick across all players. The per-frame fields follow the Solaris (Savva et al., 2026) layout (the dataset is a superset of it); the full field reference is docs/dataset-spec.md and the timing model is docs/time-alignment.md (both also in the pipeline repository).

diagnostics.tar.gz holds the capture intermediates of an episode, including each player's ReplayMod recording (.mcpr). With the pipeline these recordings can be rendered again with other projections, relative depth, or free cameras, without replaying the gameplay. They are not needed for training on the released videos.

`manifest.jsonl`

field	meaning
`episode_id`	id from `episode.json`, unique in the release
`episode_dir`	the episode's directory
`run_id`, `scenario`, `seed`, `n_players`	collection provenance
`duration_ms`	measured episode length
`qc_verdict`	`pass` / `fail`: behavioral QC, a soft label (episodes are kept either way)
`pov_anchor_spread_ticks`	cross-player time alignment; 0 for every episode
`visual_review`	`GOOD` / `WEAK` / `BAD` from a human visual review, where available
`has_diagnostics`	whether `diagnostics.tar.gz` is present

Recommended training subset

The release is complete on purpose: nothing is dropped for quality, and consumers select. We recommend qc_verdict == "pass" and visual_review != "BAD": 5387 episodes, 26538 player views, 903.1 player hours (N=1: 338 · N=2: 1854 · N=4: 1775 · N=8: 916 · N=16: 504). BAD marks scenarios in which players got stuck against terrain for most of the episode.

surface-audit.jsonl adds one per-player-view signal, below_frac: the fraction of frames at least three blocks below the surface where the player started. Descending is legitimate in some scenarios (canyons, shorelines), so treat it as a filter input rather than a verdict.

Access and loading

Downloads are gated: request access on this page and log in (hf auth login) before downloading. Paths in manifest.jsonl (episode_dir) already start with dataset/.

from huggingface_hub import snapshot_download

# Metadata only
snapshot_download("YOUSIKI/pawa-dataset", repo_type="dataset",
                  allow_patterns=["manifest.jsonl", "dataset/*/episode.json"])

# One episode, without the raw capture archive
episode_dir = "dataset/20260821T212256Z-pillar_rise_and_collapse-2p-9f474e71"  # a manifest row's episode_dir
snapshot_download("YOUSIKI/pawa-dataset", repo_type="dataset",
                  allow_patterns=[f"{episode_dir}/*"],
                  ignore_patterns=["*/diagnostics.tar.gz"])

Collection scope

Episodes were collected in several rounds of the pipeline and de-duplicated by episode_id. Sixteen episodes of two scenarios (follow_leader at N=16, formation_morph at N=4/8/16) were captured twice; the release keeps the later capture of each.
For the earliest round included here, only episodes that passed QC were published, so it contributes no QC-fail episodes; later rounds are published complete with soft labels.
No evaluation split is carved out of this release. The held-out episodes used in the paper are selected from manifest.jsonl and excluded at training time.

License

The dataset is released under CC BY-NC 4.0. It consists of gameplay footage of Minecraft, which is a trademark of Mojang Synergies AB; this dataset is not affiliated with or endorsed by Mojang or Microsoft, and use of the footage is also subject to the Minecraft Usage Guidelines.

Citation

@misc{yang2026pawa,
  title        = {{PAWA}: Panoramic Action-conditioned World Modeling for Multiple Agents},
  author       = {Yang, Siqi and Wang, Yimu and Cai, Ziqi and Zhu, Chengxuan and Weng, Shuchen and
                  Wu, Erwin and Yu, Zhaofei and Zhang, Kaipeng and Shi, Boxin},
  year         = {2026},
  howpublished = {\url{https://pawa.siki.moe}}
}

Downloads last month: 443

Total file size:

3.17 TB