Ripple10K

Precomputed Inspire-hand training and evaluation inputs for the fixed Ripple10K protocol. This release contains the seven training-input packages (9.88 GB).

The protocol selects exactly 10,000 scenes from the source collection:

Split	Scenes
Training pool	7,000
Test: in distribution	1,000
Test: unseen objects	800
Test: unseen configurations	600
Test: dense scenes	300
Test: physics	300

The training pool is divided into 6,300 fitting and 700 validation scenes. The source collection and precomputed arrays retain surplus scenes from the original collection. Apply artifacts/ripple10k/split.json when selecting the fixed 10K protocol; do not treat every row of the source arrays as selected. The split file also records excluded scene IDs, input-index hashes, the physics subset and the imagination panels.

Packages

• training/metadata.tar.gz: fixed split, scene manifests, object pool and validation-selected CADGrasp-style collision-energy coefficient.
• training/<split>.tar.gz: contact dataset, temporal segments and perceived inputs for that split, including their metadata.

Archives restore paths relative to the project root. Dataset indices retain their original source-path provenance; index_path records the relative path of a raw episode. The archive contents do not require the original absolute filesystem path to be recreated.

The training inputs become ready when manifest-training.json appears. It is published only after every archive it lists has uploaded. Raw episode archives are not included in this release; the current training code reads the prepared contact and segment arrays, with their split and perception metadata.

Download and verify

Install huggingface_hub, choose a fixed dataset commit, and download the desired part. For example, for the training inputs:

hf download Linpeng502502/Ripple10K --repo-type dataset \
  --include 'training/*' manifest-training.json SHA256SUMS-training \
  --local-dir ripple10k-download
cd ripple10k-download
sha256sum -c SHA256SUMS-training
# Run only after all checks pass; replace /path/to/project with the target root.
for archive in training/*.tar.gz; do
  tar -xzf "$archive" -C /path/to/project
done

Use --revision <commit> to pin a snapshot. The project transfer helper also verifies each archive and restores files atomically:

python pipeline/ripple10k/hf_dataset.py download --root /path/to/project --part training

On networks that cannot reach the Hub directly, the public repository can be downloaded through a compatible mirror by setting HF_ENDPOINT. Always verify the published SHA-256 values.

The packages contain numeric data and JSON metadata. Learned checkpoints, third-party simulator/robot assets, environments and training logs are separate dependencies and are not included in this dataset repository.