THULab/1g0rrr_cables_v6

cables_v6 TsFile

This repository is a TsFile conversion of 1g0rrr/cables_v6, a LeRobot dataset published by Igor (1g0rrr). The task is Insert cable into the connector using an eyou_ft7_follower robot. The original dataset is licensed under Apache-2.0. No paper or complete BibTeX citation was provided by the original dataset card.

Dataset summary

• Split: train
• Source episode files / series: 46
• Rows / frames: 20,063
• Tasks: 1
• Sampling frequency: 30 fps
• Source Parquet shards: 46 files under data/chunk-000/
• Converted output: one merged table-model TsFile, data/cables_v6.tsfile
• Table: cables_v6
• TsFile/source-Parquet size ratio: 0.5426

Schema

episode_index and task_index are native TsFile TAG columns. All other measurements are FIELD columns, except Time, which is the table time column.

Column	TsFile type	Role
time	TIMESTAMP	TIME
episode_index	STRING	TAG
task_index	STRING	TAG
frame_index	INT64	FIELD
sample_index	INT64	FIELD
action_0	FLOAT	FIELD
action_1	FLOAT	FIELD
action_2	FLOAT	FIELD
action_3	FLOAT	FIELD
action_4	FLOAT	FIELD
action_5	FLOAT	FIELD
action_6	FLOAT	FIELD
observation_state_0	FLOAT	FIELD
observation_state_1	FLOAT	FIELD
observation_state_2	FLOAT	FIELD
observation_state_3	FLOAT	FIELD
observation_state_4	FLOAT	FIELD
observation_state_5	FLOAT	FIELD
observation_state_6	FLOAT	FIELD

action and observation.state are flattened from 7-element float32 vectors to action_0..action_6 and observation_state_0..observation_state_6. The element order is shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos, wrist_flex.pos, wrist_side.pos, wrist_roll.pos, and gripper.pos. The source index column is renamed to sample_index; frame_index is kept.

Time and storage

Time = round(timestamp * 1000) in milliseconds and restarts at 0 within each episode. The redundant source timestamp column is dropped because it equals Time / 1000 seconds. No numeric rows, state dimensions, action dimensions, or task values are dropped.

• FLOAT/DOUBLE: GORILLA + LZ4
• INT32/INT64: TS_2DIFF + LZ4
• Time: TS_2DIFF + LZ4
• BOOLEAN: RLE + LZ4 (the source has no BOOLEAN fields)
• TAG values: native TsFile table device/tag storage

The physical encodings and compression were read back from the generated file with the Apache TsFile Java reader, not inferred only from the YAML config.

Videos and episode alignment

Original AV1 videos are stored at https://huggingface.co/datasets/1g0rrr/cables_v6/tree/main/videos/chunk-000 in three camera directories: observation.images.laptop, observation.images.phone, and observation.images.side. Videos are not included in this TsFile dataset.

The source repository deleted original episodes 17, 18, 20, and 41, then renamed Parquet and video filenames to 0..45 without rewriting the internal Parquet episode_index. This conversion preserves the source column as the TsFile TAG. To align a TsFile row with the renamed video file, keep frame_index unchanged and map TAG episode_index as follows:

Source TAG episode_index	Video filename episode index
0..16	unchanged
19	17
21..40	source value minus 3
42..49	source value minus 4

The complete 46-entry mapping is also stored in meta/info.json under tsfile_conversion.video_episode_alignment.

Read example

from tsfile import TsFileReader

path = "data/cables_v6.tsfile"
reader = TsFileReader(path)
print(reader.get_all_table_schemas()["cables_v6"])

columns = ["episode_index", "task_index", "frame_index", "action_0"]
with reader.query_table("cables_v6", columns, batch_size=4096) as result:
    batch = result.read_arrow_batch()
    print(batch.slice(0, 5).to_pandas())

Attribution

• Original dataset: 1g0rrr/cables_v6
• Publisher/author: Igor (1g0rrr)
• License: Apache-2.0
• Original task: Insert cable into the connector
• Citation: the original card states [More Information Needed]