Datasetpublic

Shiki42/piperx-sortletter-0905-53ep-mixed

PiperX Sort Letters — Mixed This dataset mixes three execution orders, with 36 episodes per order and 108 episodes total (88,552 frames at 30 FPS): Output episode indices (zero-based) Execution order Source episodes 0–35 Left arm, then right arm bi-sequential 0–35 36–71 Right arm, then left arm bi-sequential 54–89 72–107 Concurrent: both arms start together retimed 0–35 Source datasets PiperX Sort Letters — Bi-sequential supplies the original… See the full description on the dataset page: https://huggingface.co/datasets/Shiki42/piperx-sortletter-0905-53ep-mixed.

sourceHugging Faceupdated 17d agoView on Hugging Face

0likes475downloads

Dataset Card

PiperX Sort Letters — Mixed

This dataset mixes three execution orders, with 36 episodes per order and 108 episodes total (88,552 frames at 30 FPS):

Output episode indices (zero-based)	Execution order	Source episodes
0–35	Left arm, then right arm	bi-sequential 0–35
36–71	Right arm, then left arm	bi-sequential 54–89
72–107	Concurrent: both arms start together	retimed 0–35

Source datasets

—PiperX Sort Letters — Bi-sequential supplies the original left-then-right and generated right-then-left groups. Pinned revision: e75e16917f36f483870efaee23715afc7466b431.
—PiperX Sort Letters — Retimed (Concurrent) supplies the concurrent group. Pinned revision: b7d7061e4bd07917ea9b2de2e0ae1e0aedac4a7f.

Each group selects the first 36 demonstrations of its execution order. The groups use the same underlying first 36 source demonstrations, with three execution schedules. Rows are stored in the group order shown above; they are not shuffled. All episodes use the task text sort letters.

Action, state, per-episode frame indices and timestamps are preserved exactly from the selected source episodes. Episode IDs and global row indices are rebuilt. RGB videos are trimmed to the selected episodes and re-encoded, with unchanged frame order and FPS. No additional retiming is applied. The generated right-first and concurrent top views inherit their sources' left/right split compositing.

The data uses LeRobot v3 metadata, three RGB cameras and per-frame idle/overlap labels. Concurrent labels are derived from its original detected arm durations. Per-episode mixed/source_repo, mixed/source_revision, mixed/source_episode and mixed/mode fields record provenance. Original retime mappings are copied where available. mixed_manifest.json and VALIDATION_RECEIPT.json record selection, numeric validation, decoded video counts and pixel checks.