liguang0115/overseer-doom

OverSeer Doom dataset

Project page | Paper: arXiv link to be added

Two-player GZDoom episodes seen from ten synchronised views at once: each marine's own view and eight fixed auxiliary cameras. This is the Doom training data of OverSeer: Multi-Agent World Models Beyond What Agents See. The files use the layout of the Solaris training dataset (Solaris's Minecraft action dictionary, filled from the Doom controls), so the same tools read both of our datasets.

Contents

Episodes	Views per episode	Frames per episode	Size	Shards
19,775	10 (2 marines + 8 auxiliary)	250 (21.4 s of game time)	348 GB	350

• Played with GZDoom on the maps of Freedoom Phase 1 and Phase 2 (map and IWAD of each episode in its _episode_info.json and in episodes.csv). Two marines traverse the map, driven by a potential-field exploration controller.
• Every view is 1024 x 768 H.264 with a 120 degree horizontal field of view, copied byte for byte from the recording. One frame is captured every 3 game tics; Doom runs 35 tics per second, so the videos play at 11.67 frames per second of game time (250 frames = 21.4 s). The models in the paper use the first 241 frames.
• Eight auxiliary cameras per episode, static, about 100 map units above the marines' eye height; seven of them look 28 degrees down and the first one at a varying angle (pose of each in its _episode_info.json; Doom's pitch is positive looking down).

Files

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

One flat folder, datasets/doom/train/, with per episode (<episode> is e.g. doom_000042):

File	Content
<episode>_Alpha_instance_000.mp4 / .json / _episode_info.json	marine 1: view, per-frame pose and action, description
<episode>_Bravo_instance_000.mp4 / .json / _episode_info.json	marine 2
<episode>_Aux1_instance_000.mp4 ... Aux8, each with _episode_info.json	the eight auxiliary views and their fixed poses
<episode>_tracks.json	the engine's per-tic record: both marines' pose and controls at every tic, frame-to-tic map, camera poses, monsters alive, shots

Per-frame JSON

{"frame_count": 100, "tic": 350, "renderTime": 10000.0, "relativeTimeMs": 10000.0,
 "x": -2322.3, "y": -2797.23, "z": -16.0, "yaw": 51.91, "pitch": 0.0, "eye_z": 25.0,
 "action": {"forward": true, "back": false, "left": false, "right": false, "camera": [0.0208, -0.0],
            "attack": false, "jump": false, "...": false},
 "extra": {"forward_max": 2.82, "forward_min": 2.82, "strafe_max": 0.0, "strafe_min": 0.0, "turn_deg": 1.19,
           "attack_tics": 0, "n_tics": 3}}

• x, y, z, eye_z in map units; yaw in degrees, Doom's convention (counter-clockwise from east, grows when turning left); pitch in degrees.
• action at index t is the input applied between frame t and frame t+1 (the tics after frame t up to and including frame t+1's tic), exactly as our training loader decoded it and as Solaris's loader expects (it pairs action[t] with video frame t+1); the last frame's action is empty.
• tic is the game tic of the frame; renderTime / relativeTimeMs = tic x 1000 / 35, the timestamps Solaris's batch sampler aligns the two marines on.
• forward / back: the forward control exceeds +0.5 / falls below -0.5 map units per tic in that window; left / right: the same for the strafe control (positive strafe moves left).
• camera = (yaw change, -pitch change) over the window in radians, + = turn left / look up, the sign Solaris uses for Minecraft.
• attack: the marine fired in the window. The other Solaris keys (jump, sprint, sneak, use, mine, place, hotbar) do not exist in Doom and are always false.
• extra keeps the raw controls the keys were derived from.

Citation

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