MindCraftV2

Cognitive probes injected into VLN-CE navigation episodes, regenerated with the LASAR data-generation pipeline (arXiv:2605.16899).

10,819 trajectories · 61 Matterport3D scenes · 56,755 queries · 279 GB

Split

80/20 by episode, seed 42. The split is at trajectory level on purpose: the ~5 probes of one episode all target the same rooms and objects, so a query-level split would leak test scenes into training.

split	episodes	queries
train	8,655	45,373
test	2,164	11,382

Layout

train/000517/
  data.npz           observations, semantic_observations, actions, instruction,
                     instruction_text, scene_id, memory_log,
                     mindcraft_queries, primary_query
  meta.json          instruction + queries as plain JSON
  query_summary.png  one-page render of every query for the episode
split.json           the episode ids in each split

data.npz holds object arrays, so it needs allow_pickle:

import numpy as np
d = np.load("train/000517/data.npz", allow_pickle=True)
memory_log = d["memory_log"].item()      # per-timestep pose, room, visible objects
queries    = d["mindcraft_queries"].tolist()
frames     = d["observations"]           # (T, 256, 256, 3) uint8
semantics  = d["semantic_observations"]  # (T, 256, 256, 1) int32 instance ids

Query types

type	count	coverage	answer balance
L1.1 Object Attribute Recall	9,852	91.1%	right 51 / left 49
L1.2 Temporal Relation Recall	9,447	87.3%	after 50 / before 50
L2.1 Self-Localization	10,816	100.0%	A 33 / B 33 / C 34
L2.2 Local Spatial Relation	8,805	81.4%	right 48 / left 46 / behind 3 / in front of 3
L3.1 Topological Adjacency	9,190	84.9%	A 33 / B 34 / C 33
L3.2 Landmark Path Validation	8,645	79.9%	yes 50 / no 50

Coverage below 100% is deliberate: a generator that cannot find an unambiguous probe emits nothing rather than a guessable question.

Differences from MindCraft v1

Regenerated with the label leaks and unanswerable questions in the original generators fixed:

• L1.2 answered "before" 100% of the time (the earlier object was always named first). The pair is now swapped half the time.
• L1.1 read left/right off one arbitrarily sampled frame, so an object the agent walked past had no single correct answer; the relation must now hold across every frame where the object was clearly visible. Instance uniqueness and room naming now use the object's own region rather than the observer's.
• L2.1 drew distractors from a hand-written list of tidy room names, so against a Matterport label like familyroom/lounge the odd one out was the answer. Distractors now come from the scene's own region vocabulary.
• L2.2 always asked "left or right?" but could answer front/back. The question now matches the relation.
• L3.1 only excluded the chosen answer's name from the distractors, so a "wrong" option could name a room type that really is adjacent through another instance. All adjacent room names are excluded now.
• L3.2 was 64% "yes" and its negatives were room types absent from the scene, answerable without reasoning about the path. Both labels are now enumerated before one is chosen, and negatives are rooms the scene really has.

No episode-level video is shipped; observations in data.npz holds every frame.

License

The RGB and semantic frames are renders of Matterport3D scenes and are governed by the Matterport3D Terms of Use. You must sign that agreement before using this data. Generation code follows LASAR / habitat-lab (MIT).

Citation

@article{tang2026lasar,
  title={LASAR: Towards Spatio-temporal Reasoning with Latent Cognitive Map},
  author={Tang, Jinzhou and Liu, Sidi and Xiu, Waikit and Chen, Weixing and Wang, Keze},
  journal={arXiv preprint arXiv:2605.16899},
  year={2026}
}