real_data/clickbell/README.md

ClickBell selected 16,000 trajectories

This release contains exactly 16,000 selected simulated ClickBell trajectories: 8,000 using the left arm and 8,000 using the right arm. It was collected in simulation using the Piper real-robot URDF from piper_ros@ac41fcbcdda598f01b51cf6175ed9a24d0dacadc. The repository directory is named real_data as requested; these are not real-world robot recordings.

Each episode is the original, byte-identical .hdf5 file, containing three RGB views, joint states, control targets, the raw 250 Hz control stream, camera calibration, and scene metadata. The source uses 250 Hz physics with save_freq=15 (nominal saved observations at 16.6667 Hz). No action realignment, normalization, image conversion, or data resampling was applied during publication.

The seven clickbell-*-of-00007.tar.gz files contain only the selected HDF5 episodes. They exclude the 391 candidates not selected for the formal dataset and exclude collection caches and failed trajectories. Within each archive, the original shard_*/data/episode*.hdf5 paths are retained. Extract all seven archives into the same directory.

from pathlib import Path
import tarfile
from huggingface_hub import snapshot_download

root = Path(snapshot_download("INTACT-JEPA/INTACT", allow_patterns="real_data/clickbell/*")) / "real_data/clickbell"
for archive in sorted(root.glob("clickbell-*-of-00007.tar.gz")):
    with tarfile.open(archive, "r:gz") as f:
        f.extractall("clickbell_episodes", filter="data")

selection_manifest.jsonl has one record per selected episode with its published archive, member, byte count, and SHA-256, together with the original selection fields. source_selection_manifest.jsonl and source_summary.json preserve the original selector output. The original selection manifest SHA-256 is 215e85dcba42d81a8f52ad646254721f681a1e29de724bb3f2e3cda97165aeaf.

DATASET_INFO.json records the collection and release identities. SHA256SUMS covers every other published file. Every source episode archive was checked against the source release SHA-256 manifest while reading it; every published HDF5 has its own SHA-256 in the selection manifest.

The inherited task, camera, action, and success contracts are retained, with the robot URDF replaced by the stated real-robot model. The real_data folder name does not imply a measured sim-to-real result.