Datasets:

THULab
/

berkeley_mvp_lerobot

Modalities: Time-series. The original dataset also contains a hand-camera video stream. The videos are not included in this repository and remain available in the source dataset.

Source Dataset

Original dataset: IPEC-COMMUNITY/berkeley_mvp_lerobot
License: Apache-2.0
Task category: robotics
Robot type: xarm
LeRobot codebase version: v2.1 in meta/info.json
Split: single train split (0:480)
Scale: 480 episodes, 45,308 frames, 6 tasks, 480 source frame Parquet files
Sampling rate: 5 fps
Original video feature: observation.images.hand_image (480 x 640 RGB, AV1, 5 fps)

The six source tasks are:

push wooden cube
pick detergent from the sink
reach red block
pick yellow cube
close fridge door
pick fruit

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

Converted Files

TsFile: data/berkeley_mvp_lerobot.tsfile
TsFile size: 2,614,379 bytes
Rows: 45,308
Episodes: 480
Tasks: 6
Table name: berkeley_mvp_lerobot
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 (`0` through `5`)
`frame_index`	FIELD	INT64	Source frame index, preserved
`sample_index`	FIELD	INT64	Renamed from source `index`
`observation_state_0` ... `observation_state_7`	FIELD	FLOAT	Flattened from `observation.state[8]`
`action_0` ... `action_7`	FIELD	FLOAT	Flattened from `action[8]`

The observation-state elements are x, y, z, rx, ry, rz, rw, and gripper. The action elements are motor_0 through motor_6 and gripper.

episode_index and task_index are TAG columns, so an individual episode can be selected by filtering on episode_index, and a task can be selected by filtering on task_index.

Conversion Notes

All 480 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. Floating-point values remain single-precision FLOAT fields.
The source timestamp column is dropped because it is redundant with Time / 1000 seconds.
The source index column is renamed to sample_index.
The camera video feature is omitted from the converted schema and videos are not uploaded here. Use the original dataset videos: https://huggingface.co/datasets/IPEC-COMMUNITY/berkeley_mvp_lerobot/tree/main/videos
meta/info.json records the original and converted data paths, Time mapping, TAG columns, row count, flattened fields, renamed/dropped fields, and video alignment policy.

Validation

The output directory contains exactly one .tsfile file.
Apache TsFile SDK readback returned 45,308 rows, matching the staged Parquet.
Readback confirmed both TAG columns and all 18 FIELD columns.

Read Example

from tsfile import TsFileReader

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

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

Berkeley MVP LeRobot TsFile