THULab/act_kikobot_block_real

ACT Kikobot Block Real (TsFile)

This dataset is a numeric time-series conversion of swarajgosavi/act_kikobot_block_real, which was created using LeRobot. The source is a LeRobot v2.0 dataset for an SO-100 robot performing the task pick and place. The source dataset is licensed under Apache-2.0.

Dataset Summary

• Modalities: Time-series
• Split: train (episodes 0:50)
• Episodes: 50
• Frames / TsFile rows: 21,685
• Tasks: 1 (pick and place)
• Sampling rate: 30 fps
• Source numeric files: 50 per-episode Parquet files
• Converted files: 1 TsFile

The converted file is data/act_kikobot_block_real_train.tsfile.

TsFile Schema

The TsFile table is act_kikobot_block_real_train and contains 19 columns:

Columns	TsFile role	Type	Description
Time	TIME	INT64	round(timestamp * 1000) milliseconds, restarting within each episode
episode_index, task_index	TAG	INT64	Original LeRobot episode and task identifiers
frame_index, sample_index	FIELD	INT64	Original frame index and renamed source index
action_0 ... action_6	FIELD	FLOAT	Flattened 7-element action vector
observation_state_0 ... observation_state_6	FIELD	FLOAT	Flattened 7-element observation.state vector

The action and state indices follow the source feature order: main_shoulder_pan, main_shoulder_lift, main_elbow_flex, main_wrist_flex_1, main_wrist_flex_2, main_wrist_roll, and main_gripper.

Conversion Notes

• The 50 source episode Parquet files were merged into one TsFile for the train split. All 21,685 numeric frame rows are retained.
• Source timestamp is replaced by the millisecond Time column and is not duplicated as a FIELD because timestamp = Time / 1000 seconds, subject to integer millisecond rounding.
• Source index is renamed to sample_index.
• The full 7 values of both action and observation.state are retained as scalar FLOAT fields.
• episode_index and task_index are retained as original numeric TAG columns. No synthetic episode or task aliases are added.
• The mirrored meta/info.json records the converted path, row count, time mapping, TAG columns, flattened features, source episode-file count, and source video references.

Videos

The camera streams observation.images.laptop and observation.images.phone are not duplicated in this converted repository. The 100 source videos remain available from the original dataset videos. Use episode_index and frame_index to align TsFile rows with the original video frames.

Reading the Data

Download the converted file and open it with an Apache TsFile reader:

from huggingface_hub import hf_hub_download
from tsfile import TsFileReader

path = hf_hub_download(
    repo_id="zjt24/act_kikobot_block_real",
    repo_type="dataset",
    filename="data/act_kikobot_block_real_train.tsfile",
)

reader = TsFileReader(path)
print(reader.get_all_table_schemas())

Source and Citation

• Original dataset: swarajgosavi/act_kikobot_block_real
• Source framework: LeRobot
• License: Apache-2.0

The source dataset card does not provide a paper or BibTeX citation.