leesuyun/DiVE-data

DiVE — cluttered shelf scenes

Data for DiVE: Learning Decomposed Visibility for Efficient Active Exploration of Cluttered Scenes (CoRL 2026).

• Project page: https://lee-su-yun.github.io/DiVE/
• Code: https://github.com/lee-su-yun/DiVE
• Checkpoints: https://huggingface.co/leesuyun/DiVE

What is here

Cluttered shelf scenes collected in NVIDIA Isaac Sim with YCB objects. Each scene holds 5 pushes; each push holds the pre-push state and 20 candidate post-push states, observed from 98 candidate viewpoints on a 7x14 grid. The voxel grid is (D, H, W) = (60, 120, 80) at 5 mm.

scenes/low/{scene}.tar.zst     320 low-occlusion scenes
scenes/high/{scene}.tar.zst    320 high-occlusion scenes
scenes/{split}/fallen_log.txt  objects that fell during a push
scenes/{split}/tipped_log.txt  objects that tipped over
test/low_occlusion_test_v1.tar.zst   100 held-out evaluation episodes

Each archive expands to one scene directory:

{scene}/
├── episode_log.txt
└── push_{1..5}/
    ├── camera_poses.npz          98 candidate camera poses
    ├── pre_occ.npz               per-view occupancy, stacked
    ├── pre_semantic_occ.npz      per-view semantic labels
    ├── pre_gt.npz                ground-truth occupancy
    ├── semantic_map.npz          ground-truth semantics
    └── post_{00..19}/            the same, after executing candidate push NN
        ├── occ.npz
        ├── post_semantic_occ.npz
        ├── gt.npz
        └── action.npz            push parameters and swept volume

What is not here, and why

These archives hold what the simulator wrote. Everything the training pipeline needs on top of it is derived, and the code regenerates it bit-identically:

Not included	Regenerated by
per-view pre_occ/*.npz, post_occ/*.npz	preprocessing/visibility/data_preprocessing.py
pre_ray_casting.npy, post_ray_casting.npy	preprocessing/visibility/making_ray_casting_per_view.py
push_visibility_all_marginal.npy (the push-resolvable target)	preprocessing/visibility/making_push_visibility_per_view_all_marginal.py
ua_reward_dataset, ub_reward_dataset	preprocessing/reward_dataset/UAB_reward_generate_gt.py

Regeneration takes about 80 seconds per scene on 32 cores, and shrinks the download from roughly 1.1 TB to 169 GB. RGB-D frames and 2D semantic masks are also left out: no training or preprocessing step reads them, and they do not compress.

Usage

# one scene
huggingface-cli download leesuyun/DiVE-data scenes/low/000000000.tar.zst \
    --repo-type dataset --local-dir .
mkdir -p data_root && tar -I zstd -xf scenes/low/000000000.tar.zst -C data_root

# everything
huggingface-cli download leesuyun/DiVE-data --repo-type dataset --local-dir dive_data

Then unpack every archive of a split into one directory, put that split's fallen_log.txt and tipped_log.txt at its root, and run the preprocessing:

bash preprocessing/visibility/preprocessing_run_all.sh data_root

Citation

@inproceedings{lee2026dive,
  title     = {Learning Decomposed Visibility for Efficient Active Exploration of Cluttered Scenes},
  author    = {Lee, Suyun and Choi, Minsoo and Gong, Jihwan and Nam, Unghui and Bae, Minji and Shim, Byonghyo},
  booktitle = {Conference on Robot Learning (CoRL)},
  year      = {2026},
}