jzpeng/1g0rrr_test-rtc-dagger-5-merged-igor

Test RTC DAgger 5 Merged Igor TsFile

This dataset is an Apache TsFile conversion of 1g0rrr/test-rtc-dagger-5-merged-igor, a LeRobot robot-manipulation dataset for the task “Insert cable into the connector.”

Modalities: Time-series and tabular. The converted repository contains numeric robot states, actions, frame timing, and episode/task tags. The source camera videos remain in the original Hugging Face dataset and are not duplicated here.

Source Dataset

• Source dataset: 1g0rrr/test-rtc-dagger-5-merged-igor
• Original author/repository owner: Igor (1g0rrr)
• Source repository contributors shown by Hugging Face: 1 (1g0rrr)
• License: not declared on the source dataset page or in the downloaded metadata
• Robot type: eyou_ft7_follower
• LeRobot codebase version: 2.1
• Split: train, episodes 0:102
• Scale: 102 episodes, 37,773 frames, 30 fps
• Task metadata: one task definition, task_index=0, “Insert cable into the connector”
• Source numeric layout: data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet
• Source video layout: videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4
• Video scale recorded in meta/info.json: 306 AV1 videos (three streams for each of 102 episodes), 480 x 640, 30 fps, no audio

The downloaded source meta/info.json reported total_tasks: 0, while meta/tasks.jsonl contains the task above. The converted metadata sets total_tasks: 1 so the count matches the task file.

Converted Files

• TsFile: data/1g0rrr_test_rtc_dagger_5_merged_igor.tsfile
• Table: 1g0rrr_test_rtc_dagger_5_merged_igor
• Rows: 37,773
• Episode devices: 102
• Time precision: milliseconds
• TsFile size: 1,928,822 bytes
• Source numeric Parquet size: 3,630,944 bytes across 102 files
• TsFile/source-Parquet size ratio: 0.5312
• Metadata: meta/ mirrors the LeRobot metadata; meta/info.json is updated for the TsFile artifact and retains the original video path information

TsFile Schema

Time is synthesized as round(timestamp * 1000) milliseconds and restarts at zero for every episode. The Python SDK exposes the physical time column as time when reading the table.

TAG columns, stored with the native TsFile table device/tag mechanism:

• episode_index
• task_index

Integer FIELD columns:

• frame_index
• sample_index, renamed from source column index

FLOAT FIELD groups:

• action[7] -> action_0 ... action_6
• observation.state[7] -> observation_state_0 ... observation_state_6

The seven dimensions in both vectors are, in order: shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos, wrist_flex.pos, wrist_side.pos, wrist_roll.pos, and gripper.pos.

Encoding and Compression

• FLOAT/DOUBLE: GORILLA + LZ4
• INT32/INT64: TS_2DIFF + LZ4
• Time: TS_2DIFF + LZ4
• BOOLEAN: RLE + LZ4 policy; this dataset contains no BOOLEAN fields
• TAG: native TsFile device/tag storage

This type-aware profile is selected explicitly by the conversion script. The resulting TsFile is smaller than the source numeric Parquet files.

Conversion Notes

• All 102 episode Parquet files are merged into one table-model TsFile.
• episode_index and task_index remain the original source columns and are declared as TAGs; no synthetic episode or task aliases are created.
• Vector columns are flattened to scalar FLOAT fields. The full source column prefix is preserved, with . replaced by _.
• Source column timestamp is intentionally omitted after Time synthesis because it is redundant with Time / 1000 seconds.
• Source column index is renamed to sample_index; frame_index is retained.
• No numeric rows, episodes, action dimensions, or state dimensions are dropped.
• Use episode_index and task_index TAG filters to select a device/episode.

Source Videos

Videos are intentionally not included in this TsFile repository. They remain available in the original dataset and align with numeric rows through episode_index, frame_index, and the 30 fps timestamps:

• observation.images.laptop
• observation.images.phone
• observation.images.side

The complete source video tree is under videos/chunk-000/.

Validation

The generated file was opened with the Apache TsFile Python SDK. TsFile chunk metadata contained 37,773 rows across 102 devices, matching the 37,773 staged rows and the 37,773 source Parquet rows. A full table query also read back all 37,773 rows. The local conversion report is intentionally not part of the upload payload.

Usage

from tsfile import ColumnCategory, TsFileReader

path = "data/1g0rrr_test_rtc_dagger_5_merged_igor.tsfile"
table_name = "1g0rrr_test_rtc_dagger_5_merged_igor"
reader = TsFileReader(path)

schema = reader.get_all_table_schemas()[table_name]
columns = [
    column.get_column_name()
    for column in schema.get_columns()
    if column.get_category() in (ColumnCategory.FIELD, ColumnCategory.TAG)
]

with reader.query_table(table_name, columns, batch_size=65536) as result:
    first_batch = result.read_arrow_batch()
    print(first_batch)