ChenyuanC/3D-CovDiffusion-Train-Ready

3D-CovDiffusion Train-Ready Dataset

Processed artifacts for 3D-CovDiffusion: compact training tensors plus canonical preprocessing for the complete fixed test splits. The release removes machine-dependent rebuilding of model inputs.

Contents

Category	Training episodes	Training steps	Cached test cases
Windows	800	161,436	200
Cuboids	800	486,169	200
Shelves	800	459,439	200
Containers	70	37,175	18
Total	2,470	1,144,219	618

The 618 canonical test NPZ files total 119.46 MiB. They fix the exact preprocessed 5,120-point model input and corresponding trajectory/stroke arrays for each item in the archived test split.

Raw meshes, expert trajectories, and fixed split files remain attributed to Zenodo record 14967945. The public code's prepare command obtains them because rollout geometry, ground truth, coverage metrics, and historical top-k selection still require the raw source records. The canonical cache makes preprocessing deterministic; it does not replace the raw dataset.

Layout

3D-CovDiffusion-Train-Ready/
├── dataset_manifest.json
├── evaluation_cache_manifest.json
├── data/
│   ├── windows-v2/{manifest.json,train.zarr/}
│   ├── cuboids-v2/{manifest.json,train.zarr/}
│   ├── shelves-v2/{manifest.json,train.zarr/}
│   └── containers-v2/{manifest.json,train.zarr/}
└── evaluation-cache/
    ├── windows-v2/*.npz
    ├── cuboids-v2/*.npz
    ├── shelves-v2/*.npz
    └── containers-v2/*.npz

Each train.zarr follows schema 3dcov-train-v1:

train.zarr/
├── meta/episode_ends       int64   [episodes]
├── data/action             float32 [steps, 24]
├── data/stroke_ids         float32 [steps]
└── obs/point_cloud         float32 [episodes, 5120, 3]

Every 24-D action token concatenates four ordered 6-DoF poses. The model trains on horizon-16 sequences of these tokens. Point clouds are stored once per episode instead of once per temporal step.

Each evaluation NPZ is addressed by sample ID and contains the canonical preprocessed point cloud, trajectory, and stroke IDs. The 3dcov-evaluation-cache-v1 manifest binds all files to their category test_split.json and SHA-256 digests.

Processing configuration

• 5,120 XYZ points per observation
• four-pose action tokens (lambda_points=4)
• one-pose overlap between neighboring tokens
• position plus orientation-normal representation (action_dim=24)
• orientation weight 0.25
• per-dataset limits normalization
• float32 training storage
• episode-level training point-cloud deduplication

Download and use

The code pins the immutable dataset revision. The full public workflow is:

git clone https://github.com/crystalccy1/3D-CovDiffusion.git
cd 3D-CovDiffusion
pip install -r requirements.txt

python reproduce.py prepare all
python reproduce.py train windows
python reproduce.py evaluate windows \
  --run-dir artifacts/runs/windows/seed42

For released checkpoints instead of retraining:

python reproduce.py evaluate windows
python reproduce.py infer windows

Use cuboids, shelves, containers, or all where appropriate. One --artifact-root controls the dataset, raw evaluation records, models, and training runs.

Direct snapshot download is also possible:

from huggingface_hub import snapshot_download

dataset_root = snapshot_download(
    repo_id="ChenyuanC/3D-CovDiffusion-Train-Ready",
    repo_type="dataset",
)
print(dataset_root)

Rebuild and validation

The compact training Zarr is built from legacy replay buffers with:

python scripts/build_train_ready_dataset.py \
  --input-root /path/to/existing/replay-buffers \
  --output-root /path/to/release

The converter checks every repeated episode observation before deduplication, writes the float32 tensors consumed by training, and records logical checksums. python reproduce.py prepare <category> validates both this training schema and the separately versioned canonical evaluation cache.

Citation

@misc{chen2026_3dcovdiffusion,
  title  = {{3D-CovDiffusion}: 3D-Aware Diffusion Policy
            for Coverage Path Planning},
  author = {Chen, Chenyuan and Ding, Haoran and Ding, Ran
            and Liu, Tianyu and He, Zewen and Duan, Anqing
            and Nakamura, Yoshihiko},
  year   = {2026},
  note   = {Accepted at IROS 2026}
}

Attribution and license

This release contains modified material derived from data created by Gabriele Tiboni and published at https://zenodo.org/records/14967945 under the Creative Commons Attribution 4.0 International license: https://creativecommons.org/licenses/by/4.0/. The processed training representation, canonical preprocessing cache, and fixed category splits were produced by the 3D-CovDiffusion authors. No endorsement by the original creator is implied.