open-world-agents/D2E-480p

🕹️ D2E: Scaling Vision-Action Pretraining on Desktop Data for Transfer to Embodied AI

This repository hosts the Vision-Action subset of the D2E dataset, preprocessed at 480p for training G-IDM, Vision-Action Pretraning or other game agents.
If you need the original high-resolution dataset (HD/QHD) for world-model or video-generation training, please visit open-world-agents/D2E-Original.

Dataset Description

This dataset is a curated subset of the desktop gameplay data introduced in the paper “D2E: Scaling Vision-Action Pretraining on Desktop Data for Transfer to Embodied AI”.

The dataset enables vision-action pretraining on large-scale human gameplay data, facilitating transfer to real-world embodied AI tasks such as robotic manipulation and navigation.

Motivation & Use Cases

• 🎮 Train your own game agent using high-quality vision-action trajectories.
• 🤖 Pretrain vision-action or vision-language-action models on diverse human gameplay to learn transferable sensorimotor primitives.
• 🌍 Use as world-model data for predicting future states or generating coherent action-conditioned videos (recommend using the original HD dataset for this).
• 🧠 Generalist learning — unify multiple game domains to train models capable of cross-environment reasoning.

Dataset Structure

• Each game entry includes:
  • 🖥️ Video — desktop screen capture stored as {filename}.mkv
  • 🧩 Action Metadata — synchronized desktop interactions stored as {filename}.mcap
• Format: Each file is an OWAMcap sequence (a variant of MCAP) recorded using the OWA Toolkit, synchronizing:
  • Screen frames (up to 60 Hz)
  • Keyboard & mouse events
  • Window state changes
• Compatibility: Easily convertible to RLDS-style datasets for training or evaluation.

Dataset Details

• Recording Tool: ocap — captures screen, keyboard, and mouse events with precise timestamps, stored efficiently in OWAMcap.
• Game Genres: Includes FPS (Apex Legends), open-world (Cyberpunk 2077, GTA V), simulation (Euro Truck Simulator 2), strategy (Stardew Valley, Eternal Return), sandbox (Minecraft), and more.
• Data Collection:
  • Human demonstrations collected across 31 games (~335 h total).
  • Public release covers 29 games (~267.81 h) after privacy filtering.
• Frame Resolution: 480p (originals are HD/QHD in D2E-Original).

Dataset Summary

Game Title	Files	Total Duration (hours / seconds)	Average Duration (seconds / minutes)
Apex_Legends	36	25.58 h (92093.44 s)	2558.15 s (42.64 min)
Euro_Truck_Simulator_2	14	19.62 h (70641.61 s)	5045.83 s (84.10 min)
Eternal_Return	31	17.13 h (61677.25 s)	1989.59 s (33.16 min)
Cyberpunk_2077	7	14.22 h (51183.25 s)	7311.89 s (121.86 min)
MapleStory_Worlds_Southperry	8	14.09 h (50720.40 s)	6340.05 s (105.67 min)
Stardew_Valley	10	14.55 h (52381.45 s)	5238.14 s (87.30 min)
Rainbow_Six	11	13.74 h (49472.80 s)	4497.53 s (74.96 min)
Grand_Theft_Auto_V	11	11.81 h (42518.18 s)	3865.29 s (64.42 min)
Slime_Rancher	9	10.68 h (38463.32 s)	4273.70 s (71.23 min)
Dinkum	9	10.44 h (37600.32 s)	4177.81 s (69.63 min)
Medieval_Dynasty	3	10.32 h (37151.27 s)	12383.76 s (206.40 min)
Counter-Strike_2	10	9.89 h (35614.96 s)	3561.50 s (59.36 min)
Satisfactory	4	9.79 h (35237.30 s)	8809.32 s (146.82 min)
Grounded	4	9.70 h (34912.31 s)	8728.08 s (145.47 min)
Ready_Or_Not	11	9.59 h (34521.40 s)	3138.31 s (52.31 min)
Barony	10	9.28 h (33406.96 s)	3340.70 s (55.68 min)
Core_Keeper	7	9.02 h (32460.05 s)	4637.15 s (77.29 min)
Minecraft_1.21.8	8	8.64 h (31093.47 s)	3886.68 s (64.78 min)
Monster_Hunter_Wilds	5	8.32 h (29951.88 s)	5990.38 s (99.84 min)
Raft	5	9.95 h (35833.27 s)	7166.65 s (119.44 min)
Brotato	13	5.99 h (21574.78 s)	1659.60 s (27.66 min)
PUBG	7	4.88 h (17584.92 s)	2512.13 s (41.87 min)
Vampire_Survivors	2	2.81 h (10132.96 s)	5066.48 s (84.44 min)
Battlefield_6_Open_Beta	7	2.21 h (7965.42 s)	1137.92 s (18.97 min)
Skul	1	1.97 h (7078.00 s)	7078.00 s (117.97 min)
PEAK	2	1.75 h (6288.88 s)	3144.44 s (52.41 min)
OguForest	1	0.84 h (3040.94 s)	3040.94 s (50.68 min)
Super_Bunny_Man	2	0.72 h (2604.00 s)	1302.00 s (21.70 min)
VALORANT	1	0.25 h (911.94 s)	911.94 s (15.20 min)

Usage Example

from datasets import load_dataset

dataset = load_dataset("open-world-agents/D2E", split="train")

Citation

If you find this work useful, please cite our paper:

@article{choi2025d2e,
    title={D2E: Scaling Vision-Action Pretraining on Desktop Data for Transfer to Embodied AI},
    author={Choi, Suwhan and Jung, Jaeyoon and Seong, Haebin and Kim, Minchan and Kim, Minyeong and Cho, Yongjun and Kim, Yoonshik and Park, Yubeen and Yu, Youngjae and Lee, Yunsung},
    journal={arXiv preprint arXiv:2510.05684},
    year={2025}
}