THULab/so100_office

SO100 Office (TsFile)

This dataset contains the numeric time-series data from vladfatu/so100_office, converted from the LeRobot v2.0 layout to one Apache TsFile. The source was recorded with an SO-100 robot performing one task: "Grab red object and place it in the box."

• Modalities: Time-series (this repository); video (original repository only)
• Split: train
• Scale: 50 episodes, 29,880 frames/rows, 1 task
• Sampling rate: 30 FPS
• Converted files: 1 TsFile from 50 source episode Parquet files
• License: Apache-2.0

Data File

The complete numeric dataset is stored at data/so100_office.tsfile. It contains the table so100_office. Each source frame corresponds to one row, and episodes are represented as separate TAG combinations inside the same TsFile.

TsFile Schema

Column	TsFile role	Type	Description
Time	TIME	INT64 timestamp (ms)	round(timestamp * 1000); starts from the source timestamp within each episode
episode_index	TAG	STRING (source INT64)	Source episode index, 0 through 49
task_index	TAG	STRING (source INT64)	Source task index; 0 for this dataset
frame_index	FIELD	INT64	Frame index within the episode
sample_index	FIELD	INT64	Source global sample index
action_0 ... action_5	FIELD	FLOAT	Six-dimensional robot action vector
observation_state_0 ... observation_state_5	FIELD	FLOAT	Six-dimensional robot state vector

For both six-dimensional vectors, indices 0 through 5 preserve the source feature order: main_shoulder_pan, main_shoulder_lift, main_elbow_flex, main_wrist_flex, main_wrist_roll, and main_gripper.

Conversion Notes

• Source timestamp is represented by the TsFile Time column in milliseconds and is therefore not duplicated as a FIELD.
• Source index is renamed to sample_index to avoid ambiguity.
• action is flattened to action_0 through action_5 as single-precision FLOAT fields.
• observation.state is flattened to observation_state_0 through observation_state_5 as single-precision FLOAT fields.
• episode_index and task_index are TAG columns, so an episode can be selected without creating a separate TsFile for every episode.
• The conversion metadata and original frame/video alignment information are in meta/info.json.

Videos

The source contains 100 AV1 videos: one observation.images.laptop stream and one observation.images.phone stream for each episode. Both streams are 480 x 640 at 30 FPS and have no audio. Videos are intentionally not copied into this repository. They remain available in the videos/ tree of the original dataset. The preserved episode, frame, task, and sample indices allow converted rows to be aligned with the original frames.

Reading the TsFile

Install the Apache TsFile Python SDK, then open the actual uploaded path and query the table:

from tsfile import TsFileReader

reader = TsFileReader("data/so100_office.tsfile")
columns = [
    "episode_index",
    "task_index",
    "frame_index",
    "sample_index",
    "action_0",
    "observation_state_0",
]

with reader.query_table("so100_office", columns, batch_size=65536) as result:
    first_batch = result.read_arrow_batch()
    print(first_batch)

Source and Attribution

The original dataset was created using LeRobot. Dataset context, source videos, and the original Parquet representation are available at vladfatu/so100_office. No paper or citation was provided in the source dataset card.