Yunncheng/gamewam-vizdoom

GameWAM ViZDoom APPO Dataset

This repository contains 50,000 policy-generated gameplay trajectories from four ViZDoom combat scenarios. The trajectories were collected for GameWAM: A World Action Model for Video Games by running pretrained Sample Factory APPO policies in the native ViZDoom simulator. They are agent rollouts rather than human demonstrations.

Dataset summary

Scenario directory	Episodes	Frames	Prompt paraphrases
battle1	12,500	26,249,217	5
battle2	12,500	26,232,815	5
defend_the_center	12,500	12,614,412	5
defend_the_line	12,500	14,606,311	5
Total	50,000	79,702,755	20

The five prompts in each scenario paraphrase the same scenario objective; they do not represent five different environment tasks. Each episode is associated with one of these prompts.

Each scenario directory is a self-contained LeRobot v2.1 dataset containing meta/, data/, and videos/. Every dataset currently provides a train split covering all 12,500 episodes; no separate validation or test split is included.

Property	Value
Sampling rate	35 FPS
Video	640×480 H.264, yuv420p, no audio
Total duration	Approximately 633 hours
Approximate size	1.88 TB

Data fields

Field	Shape	Type	Description
observation.images	(3, 480, 640)	video	First-person RGB gameplay frames
observation.state	(4,)	float32	POSITION_X, POSITION_Y, ANGLE, HEALTH
observation.state_dim_is_pad	(4,)	bool	Marks unavailable state dimensions
action	(9,)	float32	Canonical action vector described below
timestamp	(1,)	float32	Episode-relative frame time
frame_index	(1,)	int64	Frame index within the episode
episode_index	(1,)	int64	Episode index within the scenario dataset
task_index	(1,)	int64	Index of the assigned prompt paraphrase

The canonical action order is:

[shoot, move_forward, move_backward, strafe_left, strafe_right,
 speed, turn_left, turn_right, turn_delta]

The first eight dimensions are binary values stored as float32. turn_delta is continuous in [-1, 1] and maps to at most ±10 degrees. battle1 and battle2 use movement, shooting, speed, and turn_delta. The two defend scenarios use shooting plus the binary turn_left and turn_right dimensions; action dimensions unavailable in a scenario are zero.

The defend scenarios do not expose all pose variables. Missing observation.state dimensions are stored as zero and identified by observation.state_dim_is_pad; consumers should use the mask rather than interpreting those zeros as observed values.

Collection and provenance

The collector runs the following public Sample Factory policies on CPU while capturing ViZDoom at its native 35 Hz simulation rate:

Scenario	Policy checkpoint
battle1	andrewzhang505/sample-factory-2-doom-battle
battle2	andrewzhang505/sample-factory-2-doom-battle2
defend_the_center	edbeeching/doom_defend_the_center_2222
defend_the_line	execbat/rl_course_vizdoom_doom_defend_the_line

Collection mixes deterministic and stochastic policy rollouts. Candidate trajectories pass scenario-aware action and motion checks plus minimum-score thresholds, and higher-scoring candidates are retained preferentially. The episode metadata in meta/episodes.jsonl records the scenario, rollout seed, policy mode, score, reward, collection settings, and quality metrics.

Limitations

• The data covers four combat scenarios and should not be treated as a broad sample of ViZDoom gameplay.
• The trajectories reflect the behavior and failure modes of the source APPO policies and the collection-time quality filters.
• Prompt variation is linguistic augmentation over four underlying objectives.
• Only a training split is provided. Users requiring held-out evaluation data should split by episode and record the selected indices.

Download and use with GameWAM

Download the complete release:

hf download Yunncheng/gamewam-vizdoom --repo-type dataset \
    --local-dir /path/to/VizDoom_Lerobot

The full repository is approximately 1.88 TB. To download one scenario, use an include pattern:

hf download Yunncheng/gamewam-vizdoom --repo-type dataset \
    --include "battle1/*" \
    --local-dir /path/to/VizDoom_Lerobot

Point the GameWAM data configuration at the four downloaded scenario directories. See configs/data/vizdoom.yaml for the complete schema.

dataset_dirs:
  - /path/to/VizDoom_Lerobot/battle1
  - /path/to/VizDoom_Lerobot/battle2
  - /path/to/VizDoom_Lerobot/defend_the_center
  - /path/to/VizDoom_Lerobot/defend_the_line

Citation

@misc{guo2026gamewamworldactionmodel,
      title={GameWAM: A World Action Model for Video Games},
      author={Yuncheng Guo and Zhanqiu Zhang and Yiwen Guo and Weijia Li},
      year={2026},
      eprint={2608.26200},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2608.26200},
}