pvskand/3dpwm-datasets

3DPWM Datasets

Prepared data for 3D Point World Models. The current release contains StackCube demonstrations for world-model training and evaluation. The code release reconstructs completed point-cloud sequences from canonical geometry and frame transforms.

StackCube

Split	Demonstrations	HDF5 frames	Canonical-geometry files
Train	963	108,463	963
Test	11	1,238	11

The archive contents preserve the original prepared files byte for byte. The unpacked dataset contains 110,675 files totaling 1,230,868,680 bytes, excluding filesystem overhead. Each split is stored in one gzip-compressed tar archive:

stackcube/
  train.tar.gz
  test.tar.gz
  SHA256SUMS
  manifest.json

manifest.json records split counts, unpacked sizes, archive sizes, and SHA-256 checksums. The archives contain relative train/ and test/ paths, respectively.

Download

From the 3DPWM code repository, with its Python environment activated:

hf download pvskand/3dpwm-datasets --repo-type dataset \
  --include "stackcube/*" --local-dir data/hf_download
(cd data/hf_download/stackcube && sha256sum -c SHA256SUMS)
mkdir -p data/StackCube-v1-demos
tar -xzf data/hf_download/stackcube/train.tar.gz -C data/StackCube-v1-demos
tar -xzf data/hf_download/stackcube/test.tar.gz -C data/StackCube-v1-demos

The hf command is provided by huggingface_hub. Public downloads do not require authentication. Add --revision <dataset-commit> to download a specific version.

The extracted layout matches configs/stackcube.yaml:

data/StackCube-v1-demos/
  train/stack_cube/demo_*/
    phases_dict.pkl
    0.h5, 1.h5, ...
  test/stack_cube/demo_*/
    phases_dict.pkl
    0.h5, 1.h5, ...

Format

Each phases_dict.pkl contains the original cannonical_points mapping from object/link names to canonical XYZ points. The spelling is part of the supplied format. The pickle arrays may reference trimesh; use the code release's environment to load them.

Each HDF5 frame supplies matching 4x4 transforms, an action or actions array, and tcp_pose or proprioception/tcp_pose. The released model uses seven-dimensional actions. Additional original fields are preserved. Frames are zero-based and contiguous within each demonstration. Canonical point order is fixed through a sequence, allowing pointwise velocity calculation.

The released training configuration uses consecutive frames (training_fpt: 1) and dataset.dt: 0.01. Keep the provided train/test split when reporting results. These archives contain prepared geometry and motion data; the supplied loader reconstructs the point clouds locally after extraction.

License

The supplied StackCube data are released under the MIT license. See LICENSE.