THULab/8ml_vial_place_30fps_fixed

8ml Vial Place 30fps Fixed TsFile

This dataset is a TsFile conversion of the corrected Hugging Face dataset Sichang0621/8ml_vial_place_30fps_fixed, which uses the LeRobot v3.0 format.

Modalities: Time-series. The original dataset also includes three camera video streams: observation.images.head_camera, observation.images.left_wrist_camera, and observation.images.right_wrist_camera. Videos are not included in this repository and remain available in the source dataset.

Source Dataset

• Original dataset: Sichang0621/8ml_vial_place_30fps_fixed
• Corrected fork of: ttotmoon/8ml_vial_place_30fps
• License: Apache-2.0
• Paper/homepage: https://github.com/TToTMooN/limb
• Format: LeRobot v3.0
• Robot type: yam
• Task: place the vial into the stand
• Split: single train split (0:197)
• Scale: 197 episodes, 229,860 frames, 1 task
• Sampling rate: 30 fps
• Original video codec: HEVC

The source dataset fixes the original wrist-camera labeling bug: the left and right wrist video streams were swapped back so their labels match the physical camera mounts. The source README states that data/*.parquet carries only state, action, episode/frame/time/index/task columns, so the numeric time-series data converted here is independent of that video-label fix.

Converted Files

• TsFile: data/vial_place_8ml_30fps_fixed.tsfile
• Rows: 229,860
• Episodes: 197
• Tasks: 1 (task_index = 0)
• Table name: vial_place_8ml_30fps_fixed
• Time precision: milliseconds
• Metadata: meta/ is mirrored from the source dataset, with meta/info.json rewritten to describe the TsFile artifact and conversion mapping.

Schema

Column	Role	Type	Notes
Time	TIME	INT64	round(timestamp * 1000), in milliseconds; restarts per episode
episode_index	TAG	INT64	Source episode identifier
task_index	TAG	INT64	Source task identifier
frame_index	FIELD	INT64	Source frame index, preserved
sample_index	FIELD	INT64	Renamed from source index
observation_state_0 ... observation_state_13	FIELD	FLOAT	Flattened from observation.state[14]
action_0 ... action_13	FIELD	FLOAT	Flattened from action[14]

The 14 state and action dimensions are aligned with the source metadata: left_joint_0 through left_joint_5, left_gripper, right_joint_0 through right_joint_5, and right_gripper.

episode_index and task_index are TAG columns, so they form the TsFile device dimension. To read one episode, filter by episode_index.

Conversion Notes

• The LeRobot frame Parquet files under data/chunk-000/ were merged into one TsFile for the train split.
• Vector columns were flattened by preserving the source column name, replacing . with _, and appending the element index.
• The source timestamp column is dropped because it is redundant with Time / 1000 seconds.
• The source index column is renamed to sample_index.
• Videos are not uploaded here. Use the source dataset videos: https://huggingface.co/datasets/Sichang0621/8ml_vial_place_30fps_fixed/tree/main/videos

Read Example

from tsfile import TsFileReader

path = "data/vial_place_8ml_30fps_fixed.tsfile"
table = "vial_place_8ml_30fps_fixed"

with TsFileReader(path) as reader:
    columns = [
        "episode_index",
        "task_index",
        "frame_index",
        "sample_index",
        "observation_state_0",
        "action_0",
    ]
    with reader.query_table(table, columns, batch_size=4096) as result:
        batch = result.read_arrow_batch()
        print(batch)