Gameplay demonstrations for the mind-games project.
200 episodes of LunarLander-v3 (Gymnasium) collected with a heuristic PD-controller expert.
Format: H.264 video (84×84) + NumPy observation arrays + JSON metadata per episode
Size: ~35MB
Mean reward: ~170 (expert heuristic averages ~163)
Structure: lunar_lander/episode_NNNN/ with frames.mp4, observations.npy, metadata.json
Index: lunar_lander/index.json with per-episode reward and step… See the full description on the dataset page:
https://huggingface.co/datasets/mad-bot/mind-games-demos.