EpicPinkPenguin/visual_distracting_control_suite

Visual Distracting Control Suite with Agent Masks

Visual Distracting Control Suite with Agent Masks contains expert trajectories for four continuous-control tasks from DeepMind Control Suite. It adds the visual variations of the Distracting Control Suite and provides two agent masks for every RGB observation: a ground-truth mask from the simulator and a mask predicted with SAM 2.1.

The dataset is designed for offline reinforcement learning, imitation learning, latent-action learning, segmentation-conditioned representation learning, and controlled studies of robustness to task-irrelevant visual variation. It is an independently generated dataset and is not an official DeepMind, Google, DAVIS, or Meta release.

The dataset accompanies the paper Segment to Focus: Guiding Latent Action Models in the Presence of Distractors, where Distracting Control Suite is used to study mask-guided latent action learning under dynamic visual distractions.

Tasks	cheetah-run, hopper-hop, humanoid-walk, walker-run
Visual variants	Vanilla, distractor low, distractor hard
Transitions	120,000,000 total (108,000,000 train + 12,000,000 test)
Per configuration	9,000,000 train + 1,000,000 test transitions
Trajectories	9,000 train + 1,000 test per configuration; 1,000 steps each
Images	64 × 64 pixels
Dataset size	Approximately 875 GB of hosted Parquet files
Policies	PPO for cheetah, hopper, and walker; SAC for humanoid

Visual variants

The low and hard suffixes are project-specific names; they do not map directly to the identically named presets in the original Distracting Control Suite. Both distracted variants use the complete DAVIS-2017 source split: all 60 training videos for dataset training and all 30 validation videos for dataset testing.

Visual variant	Hosted suffix	Dynamic DAVIS background	Dynamic color augmentation	Moving camera
Vanilla	none	No	No	No
Distractor low	_distractor_low	Yes	Low	No
Distractor hard	_distractor_hard	Yes	High	Yes

Trajectory previews

Each preview shows one complete 1,000-step training trajectory. The columns are time-aligned variants of the same task; the rows compare the RGB observation, simulation ground truth, and SAM 2.1 prediction.

Task	Vanilla, distractor low, and distractor hard
cheetah-run
hopper-hop
humanoid-walk
walker-run

Dataset usage

The complete dataset is large, so streaming is recommended for exploration and for pipelines that do not require a local copy of every configuration.

from datasets import load_dataset

dataset = load_dataset(
    "EpicPinkPenguin/visual_distracting_control_suite",
    name="cheetah_run_distractor_hard",
    split="train",
    streaming=True,
)

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

print(sample["observation"].size)  # (64, 64)
print(sample["mask"].size)         # (64, 64)
print(sample["pred_mask"].size)    # (64, 64)

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

from datasets import load_dataset

test_dataset = load_dataset(
    "EpicPinkPenguin/visual_distracting_control_suite",
    name="walker_run_distractor_low",
    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 = []

Every collected trajectory contains 1,000 transitions and ends through the time limit (truncated=True).

Dataset structure

Each row represents one transition. Its observation, masks, state, action, and reward refer to the same environment step.

Field	Hosted type	Description
observation	image	64 × 64 RGB observation for the selected visual variant.
mask	image	Binary ground-truth agent mask obtained from a synchronized simulation render.
pred_mask	image	Agent mask predicted from observation with SAM 2.1.
state	list of float64	Proprioceptive simulator observation used by the expert policy.
action	list of float64	Continuous action selected by the expert policy.
reward	float64	Environment reward associated with the transition.
terminated	bool	Whether the environment reported task termination.
truncated	bool	Whether the episode ended at its time limit.

State and action dimensions depend on the domain:

Task	State dimension	Action dimension	Expert policy	Reference expert return
cheetah-run	17	6	PPO	837.67
hopper-hop	15	4	PPO	307.33
humanoid-walk	67	21	SAC	616.52
walker-run	24	6	PPO	738.37

Configurations and splits

Domain	Vanilla	Distractor low	Distractor hard
cheetah-run	cheetah_run	cheetah_run_distractor_low	cheetah_run_distractor_hard
hopper-hop	hopper_hop	hopper_hop_distractor_low	hopper_hop_distractor_hard
humanoid-walk	humanoid_walk	humanoid_walk_distractor_low	humanoid_walk_distractor_hard
walker-run	walker_run	walker_run_distractor_low	walker_run_distractor_hard

Every configuration has the same split sizes:

Split	Trajectories per configuration	Transitions per configuration	DAVIS source pool
train	9,000	9,000,000	All 60 DAVIS-2017 training videos
test	1,000	1,000,000	All 30 DAVIS-2017 validation videos

Dataset creation

Expert policies receive proprioceptive state observations. PPO experts generate the cheetah, hopper, and walker trajectories; a SAC expert generates humanoid trajectories. Images are rendered only during collection.

For every transition, synchronized simulator instances receive the same seed and action. The main instance produces the selected visual variant. A second instance uses a black background and transparent ground plane, allowing a binary agent mask to be extracted as simulation ground truth. SAM 2.1 is then applied to the stored RGB observations to produce pred_mask.

The train and test datasets use disjoint DAVIS source splits and different trajectory seeds. The action and state dynamics are unchanged by visual augmentation.

Intended uses and limitations

Suitable applications include offline visual control, behavior cloning, latent-action learning, foreground-aware representation learning, segmentation evaluation, and robustness studies under dynamic nuisance variation.

Important limitations:

• The trajectories come from high-performing state-based experts and do not cover the state distribution of arbitrary learned policies.
• All episodes have a fixed 1,000-step horizon and the rows do not contain an explicit episode identifier.
• DAVIS backgrounds are synthetic replacements for simulator backgrounds and do not capture every real-world visual shift.

License and attribution

No repository-wide license is declared for this derived dataset. Users should review the licenses and usage terms of dm_control, Distracting Control Suite, DAVIS-2017, and SAM 2.1 before using the dataset.

Citation

Please cite the accompanying paper and the upstream resources that made this dataset possible:

@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}
}

@article{stone2021distracting,
  title   = {The Distracting Control Suite: A Challenging Benchmark for Reinforcement Learning from Pixels},
  author  = {Stone, Austin and Ramirez, Oscar and Konolige, Kurt and Jonschkowski, Rico},
  journal = {arXiv preprint arXiv:2101.02722},
  year    = {2021}
}

@article{tunyasuvunakool2020dmcontrol,
  title   = {dm\_control: Software and Tasks for Continuous Control},
  author  = {Tunyasuvunakool, Saran and Muldal, Alistair and Doron, Yotam and Liu, Siqi and Bohez, Steven and Merel, Josh and Erez, Tom and Lillicrap, Timothy and Heess, Nicolas and Tassa, Yuval},
  journal = {Software Impacts},
  volume  = {6},
  pages   = {100022},
  year    = {2020}
}

@article{ponttuset2017davis,
  title   = {The 2017 DAVIS Challenge on Video Object Segmentation},
  author  = {Pont-Tuset, Jordi and Perazzi, Federico and Caelles, Sergi and Arbelaez, Pablo and Sorkine-Hornung, Alexander and Van Gool, Luc},
  journal = {arXiv preprint arXiv:1704.00675},
  year    = {2017}
}

@inproceedings{ravi2025sam2,
  title     = {SAM 2: Segment Anything in Images and Videos},
  author    = {Ravi, Nikhila and Gabeur, Valentin and Hu, Yuan-Ting and Hu, Ronghang and Ryali, Chaitanya and Ma, Tengyu and Khedr, Haitham and Radle, Roman and Rolland, Chloe and Gustafson, Laura and Mintun, Eric and Pan, Junting and Alwala, Kalyan Vasudev and Carion, Nicolas and Wu, Chao-Yuan and Girshick, Ross and Dollar, Piotr and Feichtenhofer, Christoph},
  booktitle = {International Conference on Learning Representations},
  year      = {2025}
}