THULab/iiyudana_random50_v3

Random50 V3 TsFile

This dataset is an Apache TsFile conversion of iiyudana/random50_v3, a LeRobot v2.1 ALOHA robot-manipulation dataset.

Modalities: Time-series and tabular. The converted repository contains numeric robot state, action, frame timing, episode/task identifiers, and mirrored source metadata. The camera videos remain in the original Hugging Face dataset.

Source Dataset and Attribution

• Original dataset: iiyudana/random50_v3
• Original repository owner and uploader: iiyudana
• Pinned source revision: 5952489ab250b47121cb7c47a9fa89f7f26696aa
• Numeric data and video upload commit: cd797523d0708129e47b87372f386618e9dcf9cc
• License: Apache-2.0
• Robot type: ALOHA
• LeRobot codebase version: v2.1
• Task: Pick up the cube with the right arm and transfer it to the left arm.
• The source card does not provide a paper, formal author list, or completed citation.

Dataset Scale

• Split: train
• Episodes: 50
• Frames / TsFile rows: 20,000
• Rows per episode: 400
• Tasks: 1
• Sampling rate: 50 fps
• Duration per episode: 8 seconds
• Source Parquet shards: 50
• Source video files: 50
• Source video stream: observation.images.top
• Source video directory size: 65.2 MB as displayed by the Hugging Face file tree

Converted Files

• TsFile: data/random50_v3_compact.tsfile
• Table: random50_v3
• File size: 2,195,269 bytes
• SHA-256: a2928b7f75d649485aabc555075ea355270c9b290414efc85b658b6416dbcd4a
• Metadata: meta/ is mirrored from the source, with meta/info.json rewritten for the TsFile schema and video policy.

TsFile Schema

Time is the implicit INT64 TsFile time column in milliseconds. It is computed as round(timestamp * 1000), ranges from 0 to 7,980 ms, and restarts for every episode.

TAG columns (stored as STRING while preserving the source integer text):

• episode_index
• task_index

Scalar FIELD columns:

• frame_index
• sample_index, renamed from source index

Flattened FLOAT FIELD groups:

• observation.state[14] becomes observation_state_0 through observation_state_13
• action[14] becomes action_0 through action_13

For both 14-dimensional vectors, the element order is:

1. left_waist
2. left_shoulder
3. left_elbow
4. left_forearm_roll
5. left_wrist_angle
6. left_wrist_rotate
7. left_gripper
8. right_waist
9. right_shoulder
10. right_elbow
11. right_forearm_roll
12. right_wrist_angle
13. right_wrist_rotate
14. right_gripper

Conversion Notes

• The shared config-driven LeRobot converter was used.
• All 50 episode Parquet files are merged into one table-model TsFile.
• Compact encoding policy: FLOAT/DOUBLE use GORILLA + LZ4; INT32/INT64 use TS_2DIFF + LZ4; the implicit Time column uses TS_2DIFF + LZ4; BOOLEAN uses RLE + LZ4. TAG values remain stored through the table device/tag mechanism.
• episode_index and task_index are the device TAG columns. For example, episode 0 has episode_index equal to "0" and task_index equal to "0".
• Full source vector prefixes are preserved; dots are replaced with underscores.
• The source timestamp column is dropped after Time synthesis because it is redundant with Time / 1000 seconds.
• The source index column is retained as sample_index.
• No numeric rows, episodes, tasks, state dimensions, or action dimensions are intentionally dropped.
• The video feature observation.images.top is not written to TsFile.

Videos

The videos are not duplicated in this converted repository. The original 50 frame-aligned MP4 files are available at:

Original observation.images.top video tree

The source path pattern is videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4, covering episode_000000.mp4 through episode_000049.mp4. Numeric rows remain aligned with the source video frames through episode_index and frame_index.

Validation

The conversion was validated against the staged Parquet and read back with the Apache TsFile Java library:

• Staged Parquet rows: 20,000
• TsFile metadata rows: 20,000
• Devices: 50
• Rows per episode: 400
• Duplicate (episode_index, task_index, Time) keys: 0
• Time range per episode: 0 to 7,980 ms
• Table count: 1
• Explicit schema columns: 32 (2 TAG and 30 FIELD)
• TsFile head/tail magic: TsFile / TsFile
• Complete-file check: passed
• Encoding readback: Time TS_2DIFF + LZ4; INT64 fields TS_2DIFF + LZ4; FLOAT fields GORILLA + LZ4.

Minimal Read Example

from tsfile import TsFileReader

path = "data/random50_v3_compact.tsfile"
reader = TsFileReader(path)
table_name = "random50_v3"

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())

reader.close()

Citation

No completed citation is provided by the source dataset. Cite the original Hugging Face dataset and LeRobot when using this conversion.