| # Download archives | |
| All 1,651 episodes are stored in TAR shards under data/<task>/<batch>/part-*.tar. | |
| An episode is never split between shards. Original MP4, timestamp JSONL, pose and gripper Parquet, and episode.json bytes are preserved. No loose episode files are published alongside the shards. | |
| README, metadata, tools and all versioned calibration files are available separately. metadata/episodes.jsonl maps each episode to its archive_path; metadata/shards.jsonl gives archive sizes, checksums and episode membership. | |
| Use tools/download_dataset.py with --task, --episode-id or --all and a pinned --revision. The tool downloads complete required shards, verifies archive SHA256, extracts only selected episodes and verifies their payloads. A single-episode selection still requires its containing shard, typically up to 1.5 GiB and at most 2 GiB. TARs are retained in the output data/ directory; allow disk space for both downloaded archives and extracted episodes. | |
| The optional downloads/reframe_handheld_starter_17.zip is a self-contained 17-episode sample covering all 17 batches and four tasks. It is a subset of the main dataset, not additional recordings. See QUICKSTART.md for examples. | |
| Manual extraction: download the shared calibration and metadata, then run `tar -xf path/to/part-00000.tar -C your_dataset_root`. The original episodes/<task>/<batch>/<episode>/ paths are restored. checksums/EPISODE_FILES_SHA256SUMS verifies all extracted episode files; checksums/SHA256SUMS verifies repository files and archives. | |