Hugging Face's logo

Datasets:
open-world-agents
/
D2E-480p

Modalities:
Video
Size:
< 1K
ArXiv:
Tags:
vision-action
embodied-ai
game-dataset
imitation-learning
pretraining
OWA
Libraries:
Datasets
License:
Dataset card Data Studio Files
xet
Community
2

Update README.md

#2
by MilkClouds - opened
base: refs/heads/main
←
from: refs/pr/2
Discussion Files changed
+127
-91
Files changed (1) hide show
  1. README.md +127 -91
README.md CHANGED
@@ -7,128 +7,164 @@ tags:
7
  - imitation-learning
8
  - pretraining
9
  ---
10
- > 🚨 Urgent Update — December 1, 2025: We have **revised the dataset as of 12/1** due to critical sync issues found in the previous version. **Any data downloaded before this date is invalid and should not be used.** Please **re-download the dataset** to ensure you are working with the corrected version.
11
-
12
- # 🕹️ D2E: Scaling Vision-Action Pretraining on Desktop Data for Transfer to Embodied AI
13
- > 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.
14
- > If you need the original high-resolution dataset (HD/QHD) for **world-model** or **video-generation** training, please visit [open-world-agents/D2E-Original](https://huggingface.co/datasets/open-world-agents/D2E-Original).
15
-
16
- ## Dataset Description
17
- 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”**](https://arxiv.org/abs/2510.05684).
18
-
19
- 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.
20
-
21
- ## Motivation & Use Cases
22
- - 🎮 **Train your own game agent** using high-quality vision-action trajectories.
23
- - 🤖 **Pretrain vision-action or vision-language-action models** on diverse human gameplay to learn transferable sensorimotor primitives.
24
- - 🌍 **Use as world-model data** for predicting future states or generating coherent action-conditioned videos (recommend using the original HD dataset for this).
25
- - 🧠 **Generalist learning** — unify multiple game domains to train models capable of cross-environment reasoning.
26
-
27
- ## Dataset Structure
28
- - Each **game** entry includes:
29
- - 🖥️ Video — desktop screen capture stored as {filename}.mkv
30
- - 🧩 Action Metadata — synchronized desktop interactions stored as {filename}.mcap
31
- - **Format:** Each file is an OWAMcap sequence (a variant of MCAP) recorded using the **OWA Toolkit**, synchronizing:
32
- - Screen frames (up to 60 Hz)
33
- - Keyboard & mouse events
34
- - Window state changes
35
- - **Compatibility:** Easily convertible to RLDS-style datasets for training or evaluation.
36
-
37
- ## Dataset Details
38
- - **Recording Tool:** [ocap](https://github.com/open-world-agents/ocap) — captures screen, keyboard, and mouse events with precise timestamps, stored efficiently in OWAMcap.
39
- - **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.
40
- - **Data Collection:**
41
- - Human demonstrations collected across **31 games** (~335 h total).
42
- - Public release covers **29 games** (~**267.81 h**) after privacy filtering.
43
- - **Frame Resolution:** 480p (originals are HD/QHD in D2E-Original).
44
-
45
- ## Dataset Summary
46
- | Game Title | Files | Total Duration (hours / seconds) | Average Duration (seconds / minutes) |
47
- |-------------|--------|----------------------------------|--------------------------------------|
48
- | Apex_Legends | 36 | **25.58 h (92093.44 s)** | 2558.15 s (42.64 min) |
49
- | Euro_Truck_Simulator_2 | 14 | **19.62 h (70641.61 s)** | 5045.83 s (84.10 min) |
50
- | Eternal_Return | 31 | **17.13 h (61677.25 s)** | 1989.59 s (33.16 min) |
51
- | Cyberpunk_2077 | 7 | **14.22 h (51183.25 s)** | 7311.89 s (121.86 min) |
52
- | MapleStory_Worlds_Southperry | 8 | **14.09 h (50720.40 s)** | 6340.05 s (105.67 min) |
53
- | Stardew_Valley | 10 | **14.55 h (52381.45 s)** | 5238.14 s (87.30 min) |
54
- | Rainbow_Six | 11 | **13.74 h (49472.80 s)** | 4497.53 s (74.96 min) |
55
- | Grand_Theft_Auto_V | 11 | **11.81 h (42518.18 s)** | 3865.29 s (64.42 min) |
56
- | Slime_Rancher | 9 | **10.68 h (38463.32 s)** | 4273.70 s (71.23 min) |
57
- | Dinkum | 9 | **10.44 h (37600.32 s)** | 4177.81 s (69.63 min) |
58
- | Medieval_Dynasty | 3 | **10.32 h (37151.27 s)** | 12383.76 s (206.40 min) |
59
- | Counter-Strike_2 | 10 | **9.89 h (35614.96 s)** | 3561.50 s (59.36 min) |
60
- | Satisfactory | 4 | **9.79 h (35237.30 s)** | 8809.32 s (146.82 min) |
61
- | Grounded | 4 | **9.70 h (34912.31 s)** | 8728.08 s (145.47 min) |
62
- | Ready_Or_Not | 11 | **9.59 h (34521.40 s)** | 3138.31 s (52.31 min) |
63
- | Barony | 10 | **9.28 h (33406.96 s)** | 3340.70 s (55.68 min) |
64
- | Core_Keeper | 7 | **9.02 h (32460.05 s)** | 4637.15 s (77.29 min) |
65
- | Minecraft_1.21.8 | 8 | **8.64 h (31093.47 s)** | 3886.68 s (64.78 min) |
66
- | Monster_Hunter_Wilds | 5 | **8.32 h (29951.88 s)** | 5990.38 s (99.84 min) |
67
- | Raft | 5 | **9.95 h (35833.27 s)** | 7166.65 s (119.44 min) |
68
- | Brotato | 13 | **5.99 h (21574.78 s)** | 1659.60 s (27.66 min) |
69
- | PUBG | 7 | **4.88 h (17584.92 s)** | 2512.13 s (41.87 min) |
70
- | Vampire_Survivors | 2 | **2.81 h (10132.96 s)** | 5066.48 s (84.44 min) |
71
- | Battlefield_6_Open_Beta | 7 | **2.21 h (7965.42 s)** | 1137.92 s (18.97 min) |
72
- | Skul | 1 | **1.97 h (7078.00 s)** | 7078.00 s (117.97 min) |
73
- | PEAK | 2 | **1.75 h (6288.88 s)** | 3144.44 s (52.41 min) |
74
- | OguForest | 1 | **0.84 h (3040.94 s)** | 3040.94 s (50.68 min) |
75
- | Super_Bunny_Man | 2 | **0.72 h (2604.00 s)** | 1302.00 s (21.70 min) |
76
- | VALORANT | 1 | **0.25 h (911.94 s)** | 911.94 s (15.20 min) |
77
-
78
- ## Usage Example
79
 
80
  ```bash
81
- $ pip install mcap-owa-support owa-msgs huggingface_hub
82
  ```
83
 
 
 
84
  ```python
85
  from huggingface_hub import hf_hub_download
86
  from mcap_owa.highlevel import OWAMcapReader
87
 
88
- # Download MCAP and video files
89
  mcap_file = hf_hub_download(
90
  repo_id="open-world-agents/D2E-480p",
91
  filename="Apex_Legends/0805_01.mcap",
92
  repo_type="dataset"
93
  )
94
- video_file = hf_hub_download(
95
- repo_id="open-world-agents/D2E-480p",
96
- filename="Apex_Legends/0805_01.mkv",
97
- repo_type="dataset"
98
- )
99
 
100
- # Load and iterate through messages
101
  with OWAMcapReader(mcap_file) as reader:
102
- # Screen frames with video loading
103
  for msg in reader.iter_messages(topics=["screen"]):
104
  screen = msg.decoded
105
- screen.resolve_relative_path(mcap_file) # Resolve video path
106
- frame = screen.load_frame_array() # Load actual frame from video
107
- print(f"Frame: {frame.shape}")
108
  break
109
 
110
- # Keyboard events
111
  for msg in reader.iter_messages(topics=["keyboard"]):
112
- kbd = msg.decoded
113
- print(f"Keyboard: {kbd.event_type} VK={kbd.vk}")
114
  break
115
 
116
- # Raw mouse events (we use this instead of "mouse", see why: https://open-world-agents.github.io/open-world-agents/env/plugins/desktop/#raw-mouse-input)
117
  for msg in reader.iter_messages(topics=["mouse/raw"]):
118
- mouse = msg.decoded
119
- print(f"Mouse: dx={mouse.last_x} dy={mouse.last_y}")
120
  break
121
  ```
122
 
123
- For more details, see the [OWA Documentation](https://open-world-agents.github.io/open-world-agents/).
124
 
125
- ## Citation
126
- If you find this work useful, please cite our paper:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
127
  ```
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
128
  @article{choi2025d2e,
129
  title={D2E: Scaling Vision-Action Pretraining on Desktop Data for Transfer to Embodied AI},
130
  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},
131
  journal={arXiv preprint arXiv:2510.05684},
132
  year={2025}
133
  }
134
- ```
 
7
  - imitation-learning
8
  - pretraining
9
  ---
10
+
11
+ # D2E-480p
12
+
13
+ This is the dataset for [**D2E: Scaling Vision-Action Pretraining on Desktop Data for Transfer to Embodied AI**](https://worv-ai.github.io/d2e/).
14
+
15
+ 267 hours of synchronized video, audio, and input events from 29 PC games, for training vision-action models and game agents.
16
+
17
+ **What's included:**
18
+
19
+ - **Video + Audio**: H.264 encoded at 480p 60fps with game audio. Fixed 0.5s keyframe intervals for efficient random seek without sequential decoding.
20
+ - **Input events**: Keyboard press/release, raw mouse deltas (bypasses pointer acceleration), mouse clicks, and active window info—all with nanosecond timestamps synchronized to video frames.
21
+ - **[OWAMcap](https://open-world-agents.github.io/open-world-agents/data/getting-started/why-owamcap/) format**: Built on [MCAP](https://mcap.dev/) (widely adopted in robotics). Indexed for fast random access, crash-safe writes, and standardized message schemas that work across different datasets without custom parsing.
22
+
23
+ **Recommended for:**
24
+
25
+ - Training game agents with vision-action trajectories
26
+ - Pretraining vision-action models for transfer to embodied AI (robotic manipulation, navigation)
27
+ - World model / video generation training (use [D2E-Original](https://huggingface.co/datasets/open-world-agents/D2E-Original) for HD/QHD)
28
+
29
+ > ⚠️ **December 1, 2025**: Dataset revised due to sync issues. Re-download if you obtained data before this date.
30
+
31
+ ## Visualize
32
+
33
+ Explore recordings directly in your browser with synchronized keyboard/mouse overlay:
34
+
35
+ **👉 [Open in Dataset Visualizer](https://huggingface.co/spaces/open-world-agents/visualize_dataset?repo_id=open-world-agents/D2E-480p)**
36
+
37
+ ## Load the data
38
+
39
+ Install dependencies:
40
+
41
+ | Package | Description |
42
+ | ------------------ | -------------------------------------------------------------- |
43
+ | `mcap-owa-support` | Reader for OWAMcap files |
44
+ | `owa-msgs` | Message type definitions (`KeyboardEvent`, `MouseEvent`, etc.) |
45
+ | `huggingface_hub` | Download files from this dataset |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
46
 
47
  ```bash
48
+ pip install mcap-owa-support owa-msgs huggingface_hub
49
  ```
50
 
51
+ Then load and iterate through the data:
52
+
53
  ```python
54
  from huggingface_hub import hf_hub_download
55
  from mcap_owa.highlevel import OWAMcapReader
56
 
 
57
  mcap_file = hf_hub_download(
58
  repo_id="open-world-agents/D2E-480p",
59
  filename="Apex_Legends/0805_01.mcap",
60
  repo_type="dataset"
61
  )
 
 
 
 
 
62
 
 
63
  with OWAMcapReader(mcap_file) as reader:
 
64
  for msg in reader.iter_messages(topics=["screen"]):
65
  screen = msg.decoded
66
+ screen.resolve_relative_path(mcap_file) # Resolve video path relative to mcap
67
+ frame = screen.load_frame_array() # numpy array (H, W, 3)
 
68
  break
69
 
 
70
  for msg in reader.iter_messages(topics=["keyboard"]):
71
+ print(msg.decoded) # KeyboardEvent(event_type='press', vk=87)
 
72
  break
73
 
 
74
  for msg in reader.iter_messages(topics=["mouse/raw"]):
75
+ print(msg.decoded) # RawMouseEvent(last_x=12, last_y=-3, button_flags=0)
 
76
  break
77
  ```
78
 
79
+ **Ready to train?**
80
 
81
+ We provide [owa-data](https://github.com/open-world-agents/open-world-agents/tree/main/projects/owa-data), a data pipeline that converts this dataset into HuggingFace Datasets ready for PyTorch DataLoader. It handles tokenization and sequence packing out of the box—so you can start training immediately without writing custom data loading code.
82
+
83
+ **Learn more:**
84
+
85
+ - [OWAMcap format guide](https://open-world-agents.github.io/open-world-agents/data/technical-reference/format-guide/)
86
+
87
+ ## Structure
88
+
89
+ Each game folder contains paired `.mcap` + `.mkv` files:
90
+
91
+ ```
92
+ Apex_Legends/
93
+ ├── 0805_01.mcap # Timestamped events + frame references
94
+ ├── 0805_01.mkv # Video + audio (480p 60fps, H.264)
95
+ ├── 0805_02.mcap
96
+ ├── 0805_02.mkv
97
+ └── ...
98
  ```
99
+
100
+ The `.mcap` file stores lightweight [MediaRef](https://github.com/open-world-agents/MediaRef) pointers to video frames instead of raw pixels—frames are decoded on-demand from the `.mkv` when you call `load_frame_array()`.
101
+
102
+ MCAP files contain timestamped messages on these topics:
103
+
104
+ | Topic | Message Type | Description |
105
+ | ---------------- | ------------------------ | --------------------------------------------------------------------------------------------------------------------------- |
106
+ | `screen` | `desktop/ScreenCaptured` | Frame timestamp + MediaRef pointer to video |
107
+ | `keyboard` | `desktop/KeyboardEvent` | Key press/release with [virtual key code](https://learn.microsoft.com/en-us/windows/win32/inputdev/virtual-key-codes) |
108
+ | `keyboard/state` | `desktop/KeyboardState` | Currently pressed keys |
109
+ | `mouse` | `desktop/MouseEvent` | Mouse clicks and screen coordinates |
110
+ | `mouse/raw` | `desktop/RawMouseEvent` | Raw HID movement ([→ why raw?](https://open-world-agents.github.io/open-world-agents/env/plugins/desktop/#raw-mouse-input)) |
111
+ | `mouse/state` | `desktop/MouseState` | Current position and button state |
112
+ | `window` | `desktop/WindowInfo` | Active window title, rect, and handle |
113
+
114
+ ## Games
115
+
116
+ Genres: FPS (Apex Legends, PUBG), open-world (Cyberpunk 2077, GTA V), simulation (Euro Truck Simulator 2), sandbox (Minecraft), roguelike (Brotato, Vampire Survivors), and more.
117
+
118
+ 29 games released (~267h) from 31 games collected (~335h) after privacy filtering.
119
+
120
+ | Game | Hours | Sessions |
121
+ | ---------------------- | ----: | -------: |
122
+ | Apex Legends | 25.6 | 36 |
123
+ | Euro Truck Simulator 2 | 19.6 | 14 |
124
+ | Eternal Return | 17.1 | 31 |
125
+ | Stardew Valley | 14.6 | 10 |
126
+ | Cyberpunk 2077 | 14.2 | 7 |
127
+ | MapleStory Worlds | 14.1 | 8 |
128
+ | Rainbow Six | 13.7 | 11 |
129
+ | Grand Theft Auto V | 11.8 | 11 |
130
+ | Slime Rancher | 10.7 | 9 |
131
+ | Dinkum | 10.4 | 9 |
132
+ | Medieval Dynasty | 10.3 | 3 |
133
+ | Raft | 10.0 | 5 |
134
+ | Counter-Strike 2 | 9.9 | 10 |
135
+ | Satisfactory | 9.8 | 4 |
136
+ | Grounded | 9.7 | 4 |
137
+ | Ready Or Not | 9.6 | 11 |
138
+ | Barony | 9.3 | 10 |
139
+ | Core Keeper | 9.0 | 7 |
140
+ | Minecraft | 8.6 | 8 |
141
+ | Monster Hunter Wilds | 8.3 | 5 |
142
+ | Brotato | 6.0 | 13 |
143
+ | PUBG | 4.9 | 7 |
144
+ | Vampire Survivors | 2.8 | 2 |
145
+ | Battlefield 6 | 2.2 | 7 |
146
+ | Skul | 2.0 | 1 |
147
+ | PEAK | 1.8 | 2 |
148
+ | OguForest | 0.8 | 1 |
149
+ | Super Bunny Man | 0.7 | 2 |
150
+ | VALORANT | 0.3 | 1 |
151
+
152
+ For HD/QHD resolution, see [D2E-Original](https://huggingface.co/datasets/open-world-agents/D2E-Original).
153
+
154
+ ## Links
155
+
156
+ - [Project Page](https://worv-ai.github.io/d2e/)
157
+ - [Paper (arXiv)](https://arxiv.org/abs/2510.05684)
158
+ - [GitHub](https://github.com/worv-ai/D2E)
159
+ - [OWA Toolkit Documentation](https://open-world-agents.github.io/open-world-agents/)
160
+
161
+ ## Citation
162
+
163
+ ```bibtex
164
  @article{choi2025d2e,
165
  title={D2E: Scaling Vision-Action Pretraining on Desktop Data for Transfer to Embodied AI},
166
  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},
167
  journal={arXiv preprint arXiv:2510.05684},
168
  year={2025}
169
  }
170
+ ```