Skip to content

Dataset · Yunncheng

ViZDoom gameplay video recordings by RL agents (50k episodes)

Yunncheng/gamewam-vizdoom

A large corpus of ViZDoom gameplay videos for training world models and gaming AI.

What's inside

The collection contains 50,000 trajectory recordings of four combat scenarios in the ViZDoom simulator (10 to 100 million frames/steps). The recordings were generated not by humans, but by pre-trained agents (Sample Factory APPO) specifically for the GameWAM project. The exact format of the video files and accompanying action logs is not specified at the beginning of the card.

What you can do with it

  • Training generative world action models for video games
  • Offline reinforcement learning and imitation learning of agents
  • Predicting next video frames based on player actions
Tasks
Reinforcement learning
Size
10M–100M rows
Downloads this month
23.3K
Likes
1

How to load it

from datasets import load_dataset

ds = load_dataset("Yunncheng/gamewam-vizdoom")