Datasetpublic

THULab/GoldenPeel_so101-block-stacking-recovery

SO101 Block Stacking Recovery TsFile Apache TsFile representation of the LeRobot v3.0 SO101 follower-arm recovery demonstrations from GoldenPeel/so101-block-stacking-recovery. The task is block stacking recovery with one task label, block_stacking. Source and Attribution Original dataset: GoldenPeel/so101-block-stacking-recovery Author, repository owner, and uploader: GoldenPeel License: Apache-2.0 Robot: so_follower; LeRobot codebase version: v3.0 Citation and… See the full description on the dataset page: https://huggingface.co/datasets/THULab/GoldenPeel_so101-block-stacking-recovery.

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes30downloads

Dataset Card

SO101 Block Stacking Recovery TsFile

Apache TsFile representation of the LeRobot v3.0 SO101 follower-arm recovery demonstrations from GoldenPeel/so101-block-stacking-recovery. The task is block stacking recovery with one task label, block_stacking.

Source and Attribution

—Original dataset: GoldenPeel/so101-block-stacking-recovery
—Author, repository owner, and uploader: GoldenPeel
—License: Apache-2.0
—Robot: so_follower; LeRobot codebase version: v3.0
—Citation and paper: the source card does not provide a citation or paper.

Data Layout

—Split: train, 50 episodes, 37,450 frames, 1 task, sampled at 30 Hz.
—One table-model TsFile: data/goldenpeel_so101_block_stacking_recovery.tsfile.
—Table name: goldenpeel_so101_block_stacking_recovery.
—Original frame source: 15 Parquet shards under data/chunk-000/.
—The source has 30 AV1 MP4 files in two streams: observation.images.wrist and observation.images.top, located at https://huggingface.co/datasets/GoldenPeel/so101-block-stacking-recovery/tree/main/videos. Videos are not included here; frame indices and timestamps retain their alignment with the original videos.

Schema

Role	Columns
Time	`Time` (`round(timestamp * 1000)`, milliseconds)
TAG	`episode_index`, `task_index`
FIELD	`frame_index`, `sample_index`
FLOAT FIELD	`action_0` through `action_5`; `observation_state_0` through `observation_state_5`

The source index field is named sample_index. Vector fields are flattened in source order, with . replaced by _. Source timestamp is omitted after deriving Time; it is recoverable as Time / 1000. No other numeric trajectory field is omitted.

FLOAT fields use GORILLA plus LZ4. INT64 fields and Time use TS_2DIFF plus LZ4. TAG values are stored through TsFile table-model TAG/device semantics. There are no BOOLEAN fields in this source; BOOLEAN is configured as RLE plus LZ4.

Reading

python

from tsfile import TsFileReader

reader = TsFileReader("data/goldenpeel_so101_block_stacking_recovery.tsfile")
schema = reader.get_all_table_schemas()["goldenpeel_so101_block_stacking_recovery"]
fields = [column.get_column_name() for column in schema.get_columns()
          if column.get_column_name() != "Time"]
with reader.query_table("goldenpeel_so101_block_stacking_recovery", fields, batch_size=65_536) as result:
    batch = result.read_arrow_batch()