This repo contains the dataset and benchmark code used in

WildWorld: A Large-Scale Dataset for Dynamic World Modeling with Actions and Explicit State toward Generative ARPG

Zhen Li, Zian Meng, Shuwei Shi, Wenshuo Peng, Yuwei Wu, Bo Zheng, Chuanhao Li, Kaipeng Zhang

Alaya Studio, Shanda AI Research Tokyo; Beijing Institute of Technology; Shanghai Innovation Institute

🔥 Update

• [2026.07.06] WildWorld is now available on Hugging Face, starting with part 1 (574 hours)!
• [2026.03.25] We have released our paper — discussions and feedback are warmly welcome!

🧠 Introduction

TL;DR We present WildWorld, a large-scale action-conditioned world modeling dataset with explicit state annotations, automatically collected from a photorealistic AAA action role-playing game. It features:

• 🎬 108M+ frames with per-frame annotations: character skeletons, actions & states (HP, animation, etc.), camera poses, and depth maps
• ⚔️ 450+ semantically meaningful actions including movement, attacks, and skill casting
• 🐉 Diverse content: 29 monster species, 4 player characters, 4 weapon types, 5 distinct stages
• 🕒 Long-horizon sequences: clips spanning up to 30+ minutes of continuous gameplay
• 📝 Hierarchical captions: both action-level and sample-level natural language descriptions

🚀 Quick Start

The WildWorld dataset is hosted on Hugging Face and organized as follows:

data_part<x>/
├── <id>/
│   ├── rgb.mp4        # RGB frames at 720p@30 fps, no HUD/UI overlay
│   ├── depth.mkv      # 8-bit depth map encoded via lossless HEVC
│   ├── state.csv      # per-frame camera pose, player/monster state
│   └── skeleton.npz   # world-space 3-D joint positions per entity
└── skeleton_edges.json  # bone connectivity per skeleton type

It is planned to be released in three parts:

Part	Content	Duration	# Samples	rgb.mp4	depth.mkv	state.csv	skeleton.npz
part 1	1v1	574h	3434	~1.8 TB	~5.4 TB	~26 GB	~70 GB
part 2	TBD	-	-	-	-	-	-
part 3	TBD	-	-	-	-	-	-

📊 Data Format

World State Records (state.csv)

State records fall into three categories: camera, player/NPC, and monster.

Camera

Column	Format	Meaning
camera.K	(fx, fy, cx, cy)	Camera intrinsics
camera.pos	(x, y, z)	Camera position in world coordinates
camera.rot	(x, y, z, w)	Camera-to-world rotation in the OpenGL camera frame

Player/NPC

A frame may contain multiple players, indexed by <index>.

Column	Meaning
npc.count	Number of player characters
npc.tracking	Which NPC the camera focuses on
npc.list.<index>.type_id	NPC character type id
npc.list.<index>.member_id	Party-member id (used by npc.tracking)
npc.list.<index>.weapon_id	Weapon type id
npc.list.<index>.pos	(x, y, z) player position in world
npc.list.<index>.rot	(x, y, z, w) player world rotation
npc.list.<index>.motion_bank_id	Animation bank id
npc.list.<index>.motion_id	Animation clip id
npc.list.<index>.motion_frame	Current playback frame within the clip
npc.list.<index>.hp / max_hp / red_hp	Current / maximum / recoverable health
npc.list.<index>.sp / max_sp	Current / maximum stamina
npc.list.<index>.atk	Total physical attack (gear + buffs)
npc.list.<index>.wp_atk	Weapon's own physical attack
npc.list.<index>.attr	Total elemental attack value
npc.list.<index>.wp_attr	Weapon's own elemental attack value
npc.list.<index>.crit_rate	Critical-hit rate, in [-1, 1] (negative = weakened hits)
npc.list.<index>.crit_atk_rate	Critical damage multiplier (e.g. 1.25 = +25 %)
npc.list.<index>.def	Physical defense (can change with temporary buffs)
npc.list.<index>.resist_{fire,water,ice,elec,dragon}	Elemental resistances

Monster

A frame may contain multiple monsters, indexed by <index>.

Column	Meaning
monster.count	Number of monsters currently present
monster.list.<index>.type_id	Monster species type id; nan on frames where the monster is absent
monster.list.<index>.pos	(x, y, z) monster position in world
monster.list.<index>.rot	(x, y, z, w) monster world rotation (unit quaternion)
monster.list.<index>.motion_bank_id	Animation bank id
monster.list.<index>.motion_id	Animation clip id
monster.list.<index>.motion_frame	Current playback frame within the clip
monster.list.<index>.hp / max_hp	Current / maximum health
monster.list.<index>.cond	Active status conditions, angry / stun / poison / sleep / paralyse / parry / tired / block_npc, or none

Skeleton Joints (skeleton.npz)

Each sample's skeleton.npz holds one array per entity instance, keyed <kind><index>_<type_id>:

• kind: m = monster, n = player character, w = the player's weapon
• index: entity index number
• type_id: skeleton type id, i.e. monster species type id / player character type id / weapon type id (matches type_id / weapon_id)
• value: (N, J, 3) float32, world-space joint positions per frame

Example: m1_21 (monster index 1, species 21).

Bone Connectivity (skeleton_edges.json)

The joint axis J has a fixed, deterministic order per type_id. Bone connectivity is shipped in skeleton_edges.json, one entry per skeleton type:

{
  "em_21":  [[0, 3], [1, 5], ...],   // monsters:  key = "em_<type_id>"
  "npc":    [[0, 2], [0, 31], ...],  // player:    single shared skeleton
  "wp_4":   [[0, 1], ...],           // weapons:   key = "wp_<type_id>"
  ...
}

Visualization

The scripts/ directory provides code for rendering skeletons into videos, as shown in the intro video:

python scripts/render_skeleton_video.py data_part<x>/<id> --edges skeleton_edges.json --overlay-rgb -o skeleton.mp4

📦 Checklist

• Release part 1 of the WildWorld dataset.
• Release WildWorld dataset part 2 and part 3.
• Add detailed README and example code.
• Release code of WildBench benchmark.

📄 License

See LICENSE.

📖 Citation

If you find this project helpful, please consider citing:

@article{li2026wildworld,
  title={Wildworld: A large-scale dataset for dynamic world modeling with actions and explicit state toward generative arpg},
  author={Li, Zhen and Meng, Zian and Shi, Shuwei and Peng, Wenshuo and Wu, Yuwei and Zheng, Bo and Li, Chuanhao and Zhang, Kaipeng},
  journal={arXiv preprint arXiv:2603.23497},
  year={2026}
}