kaifz/pgnd-checkpoints

PGND Pretrained Checkpoints

Pretrained dynamics-model checkpoints for Particle-Grid Neural Dynamics for Learning Deformable Object Models from RGB-D Videos. This release contains the final 100,000-iteration model and resolved Hydra configuration for each of the six published object categories.

• Project website
• Paper
• Code
• Dataset

Categories

Category	Checkpoint	Size	SHA-256
Box	experiments/log/box/train/ckpt/100000.pt	3,375,690 bytes	916d64541431bbea9e4b3a1b6ee877ae9f7c83ab25266018d6532439136b28d4
Bread	experiments/log/bread/train/ckpt/100000.pt	3,383,370 bytes	ba4a293335716729663a56ae31f410a696129e77de035f1f0148e6e53d9c44bf
Cloth	experiments/log/cloth/train/ckpt/100000.pt	3,375,690 bytes	496477195f305b7adcb5dadd839afcd25b7cbf1033aa7b4be8542b5eff47a348
Paper bag	experiments/log/paperbag/train/ckpt/100000.pt	3,375,690 bytes	48bdb1fc30473dada60848589840008d3aaef333c7e3da947eb65e855c47fbb2
Rope	experiments/log/rope/train/ckpt/100000.pt	3,375,690 bytes	ec3c2141a4371856717b203bfa5aa9e71caf80716aa4c75f33edb05f09f68fe5
Sloth	experiments/log/sloth/train/ckpt/100000.pt	3,375,690 bytes	43c8cca1ce340738c932532fcb468edebe73ac6d5d5a9afde4b40a7e326968a3

Download One Category

Run from the root of a PGND checkout. Replace box with bread, cloth, paperbag, rope, or sloth as needed:

hf download kaifz/pgnd-checkpoints \
  --revision release-20260831 \
  --include "experiments/log/box/**" \
  --include "SHA256SUMS" \
  --local-dir .
sha256sum --check --ignore-missing SHA256SUMS

The project-relative repository layout places the downloaded files directly where the PGND evaluation code expects them:

experiments/log/box/train/hydra.yaml
experiments/log/box/train/ckpt/100000.pt

Omit the --include options to download all six categories. The corresponding dataset release uses the same immutable revision name, release-20260831.

Checkpoint Format

Each .pt file is a PyTorch ZIP checkpoint containing the material model state dictionary written by experiments/train/train_eval.py. The adjacent hydra.yaml is required because it records the model, simulation, dataset, and category-specific preprocessing configuration used by eval.py.

Use these checkpoints through the PGND codebase rather than a Transformers pipeline. As with any pickle-based PyTorch checkpoint, load files only from a trusted source and verify them against SHA256SUMS.

See CHECKPOINT_MANIFEST.json for machine-readable file metadata and split information.

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}
}