EpicPinkPenguin/visual_distracting_metaworld

Visual Distracting Meta-World with Agent Masks

Visual Distracting Meta-World with Agent Masks contains successful expert trajectories for all 50 Meta-World MT50 v3 manipulation tasks. Every visual observation includes agent masks: a ground-truth robot-arm mask obtained from the simulator, a SAM 2.1 mask predicted from the clean observation, and a separate SAM 2.1 mask predicted from the distracted observation. Each step pairs these masks with the same underlying simulator state, clean RGB render, and RGB render with a dynamic natural-video background. The distractors use the full DAVIS-2017 train/validation split (90 videos) at the easy setting with dynamic mode enabled.

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

The dataset is intended for offline reinforcement learning, imitation learning, visual representation learning, and research on robustness to task-irrelevant visual motion. This is an independently generated dataset and is not an official Meta-World, DAVIS, or Meta release.

Tasks	50 MT50 v3 configurations
Transitions	55,000,000 total (50,000,000 train + 5,000,000 test)
Per task	1,000,000 train + 100,000 test transitions
Images	128 × 128 pixels
Actions / states	4-D continuous actions / 39-D simulator states
Trajectory policy	Meta-World scripted expert policy
Visual distractors	Easy dynamic setting; all 90 DAVIS-2017 train/validation videos
Success rate	1.0 for every configuration and split

Episode previews

Each video shows one complete training episode from the corresponding task. The five panels are time-aligned; only the visual representation changes.

Task	Clean, distracted, and mask observations
assembly-v3
basketball-v3
bin-picking-v3
box-close-v3
button-press-topdown-v3
button-press-topdown-wall-v3
button-press-v3
button-press-wall-v3
coffee-button-v3
coffee-pull-v3
coffee-push-v3
dial-turn-v3
disassemble-v3
door-close-v3
door-lock-v3
door-open-v3
door-unlock-v3
drawer-close-v3
drawer-open-v3
faucet-close-v3
faucet-open-v3
hammer-v3
hand-insert-v3
handle-press-side-v3
handle-press-v3
handle-pull-side-v3
handle-pull-v3
lever-pull-v3
peg-insert-side-v3
peg-unplug-side-v3
pick-out-of-hole-v3
pick-place-v3
pick-place-wall-v3
plate-slide-back-side-v3
plate-slide-back-v3
plate-slide-side-v3
plate-slide-v3
push-back-v3
push-v3
push-wall-v3
reach-v3
reach-wall-v3
shelf-place-v3
soccer-v3
stick-pull-v3
stick-push-v3
sweep-into-v3
sweep-v3
window-close-v3
window-open-v3

Dataset usage

Because the complete dataset is approximately 2.46 TB, streaming is recommended for exploration and small-scale experiments.

from datasets import load_dataset

dataset = load_dataset(
    "EpicPinkPenguin/visual_distracting_metaworld",
    name="assembly-v3",
    split="train",
    streaming=True,
)

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

print(sample["observation"].size)  # (128, 128)
print(len(sample["state"]))        # 39
print(len(sample["action"]))       # 4

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

from datasets import load_dataset

train_dataset = load_dataset(
    "EpicPinkPenguin/visual_distracting_metaworld",
    name="pick-place-v3",
    split="train",
)
test_dataset = load_dataset(
    "EpicPinkPenguin/visual_distracting_metaworld",
    name="pick-place-v3",
    split="test",
)

Rows are stored in temporal order. Episodes can be reconstructed using terminated and truncated as boundaries:

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

Dataset structure

Each row is one time step from a trajectory. All image columns at a given row are aligned to the same simulator state.

Field	Hosted type	Description
observation	image	Clean 128 × 128 RGB observation.
observation_distracted	image	The same scene with a dynamic DAVIS-2017 video replacing background pixels.
mask	image	Binary ground-truth robot-arm mask obtained from MuJoCo simulation geometry.
pred_mask	image	SAM 2.1 robot-arm mask predicted from observation.
pred_mask_distracted	image	SAM 2.1 robot-arm mask predicted independently from observation_distracted.
state	list of float64	39-D Meta-World simulator observation used by the expert policy.
action	list of float64	4-D continuous expert action, clipped to [-1, 1].
reward	float64	Environment reward associated with the transition.
terminated	bool	Whether the transition reaches an environment terminal state.
truncated	bool	Whether the episode ends because of a time limit.

Splits

Every task configuration has the same split sizes:

Split	Transitions per task	Transitions across 50 tasks	Distractor videos
train	1,000,000	50,000,000	DAVIS-2017 train (60 videos)
test	100,000	5,000,000	DAVIS-2017 validation (30 videos)

All 50 configurations

assembly-v3	basketball-v3	bin-picking-v3	box-close-v3	button-press-topdown-v3
button-press-topdown-wall-v3	button-press-v3	button-press-wall-v3	coffee-button-v3	coffee-pull-v3
coffee-push-v3	dial-turn-v3	disassemble-v3	door-close-v3	door-lock-v3
door-open-v3	door-unlock-v3	drawer-close-v3	drawer-open-v3	faucet-close-v3
faucet-open-v3	hammer-v3	hand-insert-v3	handle-press-side-v3	handle-press-v3
handle-pull-side-v3	handle-pull-v3	lever-pull-v3	peg-insert-side-v3	peg-unplug-side-v3
pick-out-of-hole-v3	pick-place-v3	pick-place-wall-v3	plate-slide-back-side-v3	plate-slide-back-v3
plate-slide-side-v3	plate-slide-v3	push-back-v3	push-v3	push-wall-v3
reach-v3	reach-wall-v3	shelf-place-v3	soccer-v3	stick-pull-v3
stick-push-v3	sweep-into-v3	sweep-v3	window-close-v3	window-open-v3

Dataset creation

For each MT50 v3 task, three synchronized environments received identical seeds and expert actions:

1. A clean environment produced observation.
2. A visually distracting environment composited dynamic natural-video frames from the full DAVIS-2017 train/validation split (90 videos) into background regions to produce observation_distracted. This is the easy dynamic setting, and background video is the sole distractor type.
3. A segmentation environment used MuJoCo geometry IDs to isolate the robot arm and produce the binary mask.

The action at each step came from Meta-World's task-specific scripted expert and was clipped to the environment action range. Episodes terminate on task success. SAM 2.1 was subsequently run separately on the clean and distracted observations to produce pred_mask and pred_mask_distracted.

Intended uses and limitations

This dataset is suitable for offline visual control, behavior cloning, latent action learning, segmentation-conditioned representation learning, and controlled studies of robustness to dynamic backgrounds.

Important limitations:

• The trajectories cover successful scripted-expert behavior, not the full state distribution encountered by arbitrary learned policies.
• The visual shift is synthetic: DAVIS video is composited into simulator background regions and does not reproduce every real-world nuisance factor.
• The masks cover the robot arm, not every task-relevant object.
• SAM 2.1 masks are predictions and may contain omissions, leakage, or temporal inconsistency; mask should be used when simulator ground truth is required.
• Samples do not include an explicit episode ID. Use the terminal flags to recover episode boundaries, and preserve row order when processing trajectories.

License and attribution

No repository-wide license is declared for this derived dataset. Users should review the terms and citation requirements of the upstream Meta-World, DAVIS-2017, and SAM 2.1 resources and determine whether their intended use is permitted.

Citation

If this dataset is useful in your work, cite the upstream resources that made it possible:

@article{fechner2026segment,
  title   = {Segment to Focus: Guiding Latent Action Models in the Presence of Distractors},
  author  = {Fechner, Marcus and Adnan, Hamza and L{\"u}th, Constantin C. and Jackson, Matthew T. and Zakharov, Alexey and Z{\"o}llner, J. Marius},
  journal = {arXiv preprint arXiv:2602.02259},
  year    = {2026},
  url     = {https://arxiv.org/abs/2602.02259}
}

@inproceedings{mclean2025metaworld,
  title     = {Meta-World+: An Improved, Standardized, RL Benchmark},
  author    = {McLean, Reginald and Chatzaroulas, Evangelos and McCutcheon, Luc and R{\"o}der, Frank and Yu, Tianhe and He, Zhanpeng and Zentner, K. R. and Julian, Ryan and Terry, J. K. and Woungang, Isaac and Farsad, Nariman and Castro, Pablo Samuel},
  booktitle = {Advances in Neural Information Processing Systems, Datasets and Benchmarks Track},
  year      = {2025}
}

@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}
}

@article{ravi2024sam2,
  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 R{\"a}dle, 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 Doll{\'a}r, Piotr and Feichtenhofer, Christoph},
  journal = {arXiv preprint arXiv:2408.00714},
  year    = {2024}
}