THULab/hnghia517_3_items

Hnghia517 3 Items TsFile

This dataset is an Apache TsFile conversion of hnghia517/3_items, a LeRobot v3 robot dataset recorded with a LeKiwi client robot.

Modalities: Time-series. The converted repository contains numeric robot observations, actions, frame timing, and episode/task tags. The original camera videos remain in the source Hugging Face dataset.

Source Dataset and Author

• Source dataset: hnghia517/3_items
• Pinned source revision: f7c9691edb1432e63c5f65c825e1860410c81bfd
• Repository owner and uploader: Huy Nghia (hnghia517)
• Authorship evidence: the Hugging Face profile for hnghia517 displays the name Huy Nghia; the source card provides no separate formal author list.
• License: Apache-2.0
• Robot type: lekiwi_client
• LeRobot codebase version: v3.0
• Split: train
• Scale: 100 episodes, 430,043 frames, 1 task index, 30 fps
• Source frame files: 31 Parquet shards under data/chunk-000/
• Source frame layout: data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet
• Source video layout: videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4
• Paper/citation/task description: not supplied by the source dataset card

The source meta/tasks.parquet contains only task_index = 0 and does not include a textual task label, so this conversion does not invent one.

Converted Files

• TsFile: data/hnghia517_3_items_train.tsfile
• Table: hnghia517_3_items_train
• Rows: 430,043
• Episodes/devices: 100
• Tasks: 1
• Time precision: milliseconds
• Metadata: meta/ is mirrored from the source, with meta/info.json rewritten to describe the TsFile artifact and video policy.

TsFile Schema

Time is synthesized as round(timestamp * 1000) milliseconds and restarts from 0 for each episode.

TAG columns:

• episode_index
• task_index

Scalar FIELD columns:

• frame_index
• sample_index, renamed from source column index

Flattened FLOAT FIELD groups:

• action[9] -> action_0 ... action_8
• observation.state[9] -> observation_state_0 ... observation_state_8

For both vectors, the nine source dimensions are, in order: arm_shoulder_pan.pos, arm_shoulder_lift.pos, arm_elbow_flex.pos, arm_wrist_flex.pos, arm_wrist_roll.pos, arm_gripper.pos, x.vel, y.vel, and theta.vel.

Conversion Notes

• The shared config-driven lerobot converter is used; the dataset-specific convert_hnghia517_3_items.py script is a local orchestration and validation entry point.
• The train split is merged into one table-model TsFile. Filter by episode_index and task_index to select an episode or task.
• Vector columns are flattened to scalar TsFile fields. Full source prefixes are preserved, with . replaced by _.
• The source timestamp column is dropped after Time synthesis because it is redundant with Time / 1000 seconds.
• Source index is renamed to sample_index; frame_index is retained.
• All 31 source Parquet shards share the same schema. No numeric row, episode, task index, action dimension, or observation-state dimension is dropped.

Encoding and Compression

The final TsFile is rewritten by EncodedLeRobotTsFileImporter.java with an explicit compact policy:

• Time (INT64): TS_2DIFF + LZ4
• FLOAT/DOUBLE FIELD values: GORILLA + ZSTD
• INT32/INT64 FIELD values: TS_2DIFF + ZSTD
• BOOLEAN FIELD values: RLE + LZ4
• TAG values: TsFile table device/TAG mechanism

This avoids the larger default PLAIN/uncompressed output produced by the generic schema importer.

Videos

Videos are not duplicated in this converted repository. They remain in three frame-aligned camera streams in the original dataset:

• videos/observation.images.front/chunk-000/
• videos/observation.images.top/chunk-000/
• videos/observation.images.wrist/chunk-000/

The source path template is videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4. Source metadata describes 640x480 AV1 video at 30 fps without audio. Numeric rows remain aligned with the original videos through episode_index, frame_index, and the source episode metadata under meta/episodes/.

Validation

The generated TsFile is inspected with the Apache TsFile Python SDK and read back in full with the Apache TsFile Java 2.2.1 reader. The Java check verifies 430,043 query rows and the physical encoding/compression policy; metadata and device counts are cross-checked against staged Parquet. The table contains 100 episode/task devices, 2 TAG columns, and 20 FIELD columns.

The local files VALIDATION.md and validation_report.json contain exact hashes and check results; they are conversion artifacts and should not be uploaded with the dataset.

Usage

from tsfile import TsFileReader

path = "data/hnghia517_3_items_train.tsfile"
reader = TsFileReader(path)
table_name = "hnghia517_3_items_train"
columns = [
    "episode_index",
    "task_index",
    "frame_index",
    "sample_index",
    "action_0",
    "observation_state_0",
]

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