kaifz/pgnd-dataset

PGND Dataset

This is the training and evaluation dataset for Particle-Grid Neural Dynamics for Learning Deformable Object Models from RGB-D Videos. It contains 2,180 trajectory windows across six deformable-object categories, including 160 evaluation windows with matching Gaussian splats.

• Project website
• Paper
• Code

Categories

Category	Extracted size	Unique payload	Trajectories	Eval trajectories	Raw episodes
Box	43.654 GiB	32.472 GiB	323	20	16
Bread	23.297 GiB	16.700 GiB	163	20	32
Cloth	93.732 GiB	67.875 GiB	650	40	19
Paper bag	35.019 GiB	24.441 GiB	220	20	8
Rope	109.156 GiB	75.685 GiB	691	40	15
Sloth	21.215 GiB	14.891 GiB	133	20	3
Total	326.075 GiB	232.068 GiB	2,180	160	93

Each category is independent and stored as one archive under data/<category>/<category>.tar. The shared Gaussian assets used by rendering and gripper point sampling are in assets/pgnd_gs_assets.tar.

Download One Category

The included helper downloads one category, verifies archive checksums, and extracts the files into the paths expected by the PGND code:

hf download kaifz/pgnd-dataset download_category.py \
  --repo-type dataset \
  --revision release-20260831 \
  --local-dir .
python download_category.py box \
  --revision release-20260831 \
  --output /path/to/pgnd

Equivalent manual download:

hf download kaifz/pgnd-dataset \
  --repo-type dataset \
  --revision release-20260831 \
  --include "data/box/*" \
  --include "assets/*" \
  --include "SHA256SUMS" \
  --local-dir pgnd-download

cd /path/to/pgnd
find pgnd-download/data/box pgnd-download/assets -name '*.tar' \
  -exec tar -xf '{}' \;

Every tar archive stores project-relative paths beginning with experiments/log/. After extraction, the category data and shared assets appear at:

experiments/log/data/<source_dataset>/...
experiments/log/data/<category>_merged/...
experiments/log/gs/ckpts/gripper.splat
experiments/log/gs/ckpts/gripper_new.splat
experiments/log/gs/ckpts/table.splat

Deduplicated Packaging

Merged trajectory files are byte-identical to files in their processed source subepisodes. Each category archive stores the processed file once and represents its merged path as a POSIX hardlink. This removes 94.008 GiB of duplicate payload while preserving the exact directory structure expected by PGND.

After extraction, hardlinked paths behave like ordinary files and require no loader changes. The provided downloader recreates all 13,080 hardlinks and falls back to independent copies when the destination filesystem does not support hardlinks. All link targets are internal to the same category archive; categories remain independently downloadable.

Dataset Contents

Raw processed episodes contain four synchronized RGB-D camera streams, current object masks, numeric camera calibration, timestamps, robot state, and the Gaussian splats required by the evaluation set. The corresponding <category>_merged/sub_episodes_v/ directory contains the self-contained trajectory windows consumed by training and evaluation.

Every processed and merged trajectory directory has the same six files: traj.npz, meta.txt, eef_traj.txt, eef_rot.txt, eef_gripper.txt, and cam_indices.txt. Every traj.npz contains exactly xyz and v; 220 unused legacy color arrays from the paper-bag trajectories were removed. The 2,180 processed trajectories map one-to-one onto the merged train/eval metadata, with no unreferenced source trajectories.

Every merged trajectory includes eef_rot.txt and eef_gripper.txt. The 413 position-only trajectories from 1130_rope_short_processed use the processing defaults: identity end-effector rotation and zero gripper state. Every trajectory also includes cam_indices.txt. The 413 previously missing rope sidecars were recovered by exact correspondence between existing anchor trajectory points and the four raw masked RGB-D projections. All 4,016,485 labels matched within 1.93e-16 with no cross-camera ambiguity; trajectory payloads were not regenerated or modified.

The published release intentionally excludes incomplete or generated intermediate artifacts that are not consumed by the current pipeline:

• reconstructed point clouds (pcd_clean/)
• depth masks (depth_mask/)
• optical-flow velocity trees (vel/)
• historical masks and legacy Gaussian trees
• dated traj_orig_*.npz trajectory backups
• dated sub_episodes_v_orig* trajectory trees
• complete processed tail trajectories not referenced by a merged split
• empty, unreferenced subepisode directories
• empty Gaussian directories
• generated calibration, tracking, segmentation, and debug visualizations

See DATASET_MANIFEST.json for exact category membership, archive statistics, and checksums for the shared assets. SHA256SUMS contains checksums for every downloadable tar archive.

Validation

Before packaging, the release was validated against the current PGND training and evaluation paths:

• 2,020 training and 160 evaluation trajectories are present.
• All 93 raw episodes have matching four-camera RGB/depth/mask, robot, and timestamp frame sets (140,329 synchronized frames per camera).
• The exact 160 evaluation-anchor splats are present and structurally valid.
• All 372 active camera intrinsic matrices match the stored 848x480 images.
• The three shared Gaussian assets are present and checksum-verified.
• All 13,080 processed/merged file pairs are byte-identical and packaged as intra-category hardlinks.
• All trajectory directories and NPZ members follow the uniform schema, and all 2,180 processed source trajectories are referenced by a merged split.
• No excluded point-cloud, depth-mask, historical-mask, velocity, legacy-splat, or visualization directories remain.

Limitations

The data was collected in a controlled robot-lab setup with six object categories and four fixed RGB-D viewpoints. It should not be treated as a representative sample of unconstrained environments, camera layouts, robots, or deformable objects. RGB-D observations and masks may retain sensor noise and segmentation errors.

Citation

@inproceedings{zhang2025particle,
  title={Particle-Grid Neural Dynamics for Learning Deformable Object Models from RGB-D Videos},
  author={Zhang, Kaifeng and Li, Baoyu and Hauser, Kris and Li, Yunzhu},
  booktitle={Proceedings of Robotics: Science and Systems (RSS)},
  year={2025}
}