azithromycin/CSI-Agent_foldtowel_merged

CSI-Agent Fold Towel Merged TsFile

This repository is an Apache TsFile conversion of CSI-Agent/foldtowel_merged, a LeRobot v3.0 robot-manipulation dataset. The source task is Fold the clothes. with a bi-so_follower (dual SO-101) setup. The conversion contains numeric and phase-label time series; camera videos remain in the original dataset.

Source and attribution

• Source: CSI-Agent/foldtowel_merged
• Pinned source revision: 1d82051ad4307df6ace2c9e67737768b85299fcd
• Original owner/author: CSI-Agent, the Hugging Face organization and repository owner
• Author profile: https://huggingface.co/CSI-Agent
• License: Apache-2.0
• Paper/citation: the source card provides no paper or BibTeX citation
• Task: fold clothes; task index 0 maps to Fold the clothes.
• Robot type: bi_so_follower; LeRobot codebase: v3.0

Dataset and videos

• Split: train
• Episodes: 50
• Rows/frames: 51,203
• Tasks: 1
• Sampling rate: 30 fps (source meta/info.json)
• Frame Parquet shards: 50
• Source episode metadata shards: 1
• Video files: 150 total, 50 per stream
• Phase labels: step1 15,962, step2 13,895, step3 21,346

The original videos are located at:

• videos/observation.images.left_cam/
• videos/observation.images.top/
• videos/observation.images.right_cam/

Videos are not included in this TsFile repository. episode_index and frame_index preserve alignment with the source episode metadata and the three per-episode MP4 streams.

Converted artifact

• TsFile: data/csi_agent_foldtowel_merged.tsfile
• Table: csi_agent_foldtowel_merged
• Rows: 51,203
• Size: 1,813,681 bytes
• Source frame Parquet bytes: 3,308,349
• TsFile/source Parquet size ratio: 0.548
• Granularity: one merged table; each (episode_index, task_index) pair is a TsFile TAG device

Schema

Column(s)	TsFile type	Role	Encoding + compression	Description
Time	INT64	TIME	TS_2DIFF + LZ4	round(timestamp * 1000) milliseconds; restarts at 0 per episode
episode_index	STRING	TAG	TsFile table device/TAG	Source episode index; source dtype INT64
task_index	STRING	TAG	TsFile table device/TAG	Source task index; source dtype INT64
frame_index, sample_index	INT64	FIELD	TS_2DIFF + LZ4	sample_index is source index renamed
phase_label	STRING	FIELD	PLAIN + LZ4	Source phase label (step1, step2, step3)
action_0..action_11	FLOAT	FIELD	GORILLA + LZ4	12-dimensional action vector, flattened from action
observation_state_0..observation_state_11	FLOAT	FIELD	GORILLA + LZ4	12-dimensional state vector, flattened from observation.state

The action/state element order follows the source feature names: left_shoulder_pan.pos, left_shoulder_lift.pos, left_elbow_flex.pos, left_wrist_flex.pos, left_wrist_roll.pos, left_gripper.pos, followed by the corresponding six right_* names. Dots in source names are replaced with underscores only for flattened field names.

Conversion details

• Time = round(timestamp * 1000) with millisecond precision. The source timestamp column is dropped because it is exactly Time / 1000 seconds; frame_index is preserved.
• Source episode_index and task_index remain the TAG columns. No synthetic episode_id or task_id aliases are created.
• Vector columns are flattened to scalar FLOAT fields without dropping numeric values. phase_label is retained as a STRING field.
• All rows are sorted by TAG columns and then Time; duplicate TAG/Time keys are rejected before writing.
• Numeric codecs are explicit: FLOAT/DOUBLE GORILLA + LZ4, INT32/INT64 and Time TS_2DIFF + LZ4, BOOLEAN RLE + LZ4; TAGs use the TsFile table device mechanism. The source contains no BOOLEAN field.
• Only timestamp is dropped. Video columns are source metadata references and are intentionally not copied into TsFile.

Read example

from tsfile import TsFileReader

reader = TsFileReader("data/csi_agent_foldtowel_merged.tsfile")
table = reader.get_all_table_schemas()["csi_agent_foldtowel_merged"]
columns = [c.get_column_name() for c in table.get_columns()
           if c.get_column_name() != "Time"]
with reader.query_table("csi_agent_foldtowel_merged", columns, batch_size=65536) as result:
    batch = result.read_arrow_batch()
    print(batch)

Validation

The local validation report confirms 51,203 Java readback rows, 50 TAG devices, 2 TAG columns, 27 FIELD columns, strict per-device Time monotonicity, and the codec policy above. TIME was independently checked from aligned chunk metadata as TS_2DIFF + LZ4. Conversion scripts and JSON/Markdown validation reports are kept locally and are not part of the upload set.