EpicPinkPenguin/atari_frostbite_with_masks

Atari Frostbite Expert Trajectories with Player Masks

This dataset contains expert-policy transitions from Atari Frostbite together with a programmatically derived binary mask of the player in every stored RGB observation. The expert is a Rainbow agent trained with CleanRL, and player localization is provided by OCAtari.

The dataset accompanies the paper Segment to Focus: Guiding Latent Action Models in the Presence of Distractors, where Frostbite is used to study mask-guided latent action learning when moving ice floes also drive the player's motion.

It is an independently generated research dataset and is not an official Atari, Arcade Learning Environment, OCAtari, or CleanRL release.

Environment	FrostbiteNoFrameskip-v4
Policy	Rainbow
Reference expert return	3076.0
Transitions	2,200,000 total (2,000,000 train + 200,000 test)
Observations	128 × 128 RGB images
Masks	128 × 128 binary player masks
Action space	18 discrete actions
Hosted data size	Approximately 8.85 GB

Preview

The video shows the RGB observation, binary player mask, and overlay side by side.

Dataset usage

Streaming is recommended for exploration because it avoids downloading the complete split:

from datasets import load_dataset

dataset = load_dataset(
    "EpicPinkPenguin/atari_frostbite_with_masks",
    name="FrostbiteNoFrameskip-v4",
    split="train",
    streaming=True,
)

sample = next(iter(dataset))
print(sample.keys())
# dict_keys(['observation', 'mask', 'action', 'reward',
#            'terminated', 'truncated'])

print(sample["observation"].size)  # (128, 128)
print(sample["mask"].size)         # (128, 128)

Remove streaming=True to download and cache a complete split locally:

from datasets import load_dataset

test_dataset = load_dataset(
    "EpicPinkPenguin/atari_frostbite_with_masks",
    name="FrostbiteNoFrameskip-v4",
    split="test",
)

Rows are stored in temporal order. Episodes can be reconstructed from the terminal flags:

def episodes(dataset):
    episode = []
    for step in dataset:
        episode.append(step)
        if step["terminated"] or step["truncated"]:
            yield episode
            episode = []

Dataset structure

Each row contains one observation, its player mask, the action selected from that observation, and arrival metadata for that observation.

Field	Hosted type	Description
observation	image	128 × 128 RGB frame rendered from the Atari environment.
mask	image	128 × 128 single-channel player mask with pixel values 0 or 255.
action	int32	Discrete action selected by the expert from the current observation; values are in [0, 17].
reward	float32	Raw game-score delta on the transition into the current observation.
terminated	bool	Whether that incoming transition caused life loss or true game over.
truncated	bool	Whether that incoming transition was truncated by the environment.

The stored 128 × 128 RGB image is not the direct policy input. The policy acts on a stack of the four most recent 84 × 84 grayscale frames. Stored masks and RGB observations are spatially aligned at 128 × 128.

Splits

Split	Transitions	Uncompressed dataset bytes
train	2,000,000	8,059,217,323
test	200,000	805,274,963

Dataset creation

The dataset was collected from a CleanRL Rainbow expert in FrostbiteNoFrameskip-v4, without action repetition or frame skipping. The policy consumes four stacked 84 × 84 grayscale frames, while the dataset stores 128 × 128 RGB observations. Binary player masks are programmatically derived from OCAtari RAM-based object localization and player sprite color matching, then aligned with the stored observations.

Intended uses and limitations

Suitable applications include offline visual control, behavior cloning, latent-action learning, player-aware representation learning, segmentation experiments, and controlled comparisons of input masking and loss masking.

Important limitations:

• The data represent the visitation distribution of one high-performing expert, not the full state-action distribution or the behavior of arbitrary policies.

License and attribution

No repository-wide license is declared for this derived dataset. Users are responsible for reviewing the licenses and usage terms of CleanRL, OCAtari, Gymnasium, and the Arcade Learning Environment, as well as the terms governing Atari ROMs and game audiovisual content. The presence of data on the Hugging Face Hub does not grant rights to third-party game content.

Citation

Please cite the accompanying paper and the relevant upstream work:

@misc{fechner2026segmentfocusguidinglatent,
  title         = {Segment to Focus: Guiding Latent Action Models in the Presence of Distractors},
  author        = {Marcus Fechner and Hamza Adnan and Constantin C. L{\"u}th and Matthew T. Jackson and Alexey Zakharov and J. Marius Z{\"o}llner},
  year          = {2026},
  eprint        = {2602.02259},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2602.02259}
}

@inproceedings{hessel2018rainbow,
  title     = {Rainbow: Combining Improvements in Deep Reinforcement Learning},
  author    = {Hessel, Matteo and Modayil, Joseph and van Hasselt, Hado and Schaul, Tom and Ostrovski, Georg and Dabney, Will and Horgan, Dan and Piot, Bilal and Azar, Mohammad and Silver, David},
  booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence},
  volume    = {32},
  number    = {1},
  year      = {2018},
  doi       = {10.1609/aaai.v32i1.11796}
}

@article{bellemare2013arcade,
  title   = {The Arcade Learning Environment: An Evaluation Platform for General Agents},
  author  = {Bellemare, Marc G. and Naddaf, Yavar and Veness, Joel and Bowling, Michael},
  journal = {Journal of Artificial Intelligence Research},
  volume  = {47},
  pages   = {253--279},
  year    = {2013}
}

@article{delfosse2024ocatari,
  title   = {{OCAtari}: Object-Centric {Atari} 2600 Reinforcement Learning Environments},
  author  = {Delfosse, Quentin and Bl{\"u}ml, Jannis and Gregori, Bjarne and Sztwiertnia, Sebastian and Kersting, Kristian},
  journal = {Reinforcement Learning Journal},
  volume  = {1},
  pages   = {400--449},
  year    = {2024}
}

@article{huang2022cleanrl,
  title   = {{CleanRL}: High-quality Single-file Implementations of Deep Reinforcement Learning Algorithms},
  author  = {Huang, Shengyi and Dossa, Rousslan Fernand Julien and Ye, Chang and Braga, Jeff and Chakraborty, Dipam and Mehta, Kinal and Ara{\'u}jo, Jo{\~a}o G. M.},
  journal = {Journal of Machine Learning Research},
  volume  = {23},
  number  = {274},
  pages   = {1--18},
  year    = {2022},
  url     = {https://jmlr.org/papers/v23/21-1342.html}
}