azithromycin/vladfatu_so100_office

SO100 Office TsFile

This is an Apache TsFile representation of vladfatu/so100_office, a LeRobot v2.0 SO100 manipulation dataset. It contains numeric robot trajectories and frame timing. Camera videos stay in the original Hugging Face repository.

Source

• Author and repository owner: Vlad Fatu (vladfatu)
• License: Apache-2.0
• Task: grab the red object and place it in the box
• Robot: SO100; LeRobot version: v2.0
• Split: train; 50 episodes; 29,880 frames; 1 task; 30 fps
• Source data: 50 Parquet episode shards at data/chunk-000/episode_{episode_index:06d}.parquet
• Source card: https://huggingface.co/datasets/vladfatu/so100_office
• Paper and citation: not supplied by the source card

TsFile Data

• Path: data/vladfatu_so100_office_train.tsfile
• Table: vladfatu_so100_office_train
• Rows: 29,880 across 50 episode/task devices
• Time precision: milliseconds
• meta/ retains source metadata; meta/info.json points data_path to the TsFile

Time = round(timestamp * 1000) milliseconds and restarts for each episode. The redundant source timestamp field is not stored separately. Source index is named sample_index; frame_index is retained. No rows or numeric vector dimensions are omitted.

TsFile role	Columns	Type / storage
TIME	Time	INT64 timestamp, milliseconds
TAG	episode_index, task_index	TsFile table/device tags
FIELD	frame_index, sample_index	INT64
FIELD	action_0 through action_5	FLOAT
FIELD	observation_state_0 through observation_state_5	FLOAT

action[6] and observation.state[6] are flattened in source order, replacing the dot in the latter name with an underscore. The dimension names are main_shoulder_pan, main_shoulder_lift, main_elbow_flex, main_wrist_flex, main_wrist_roll, and main_gripper.

FLOAT and DOUBLE use GORILLA + LZ4. INT32 and INT64 fields use TS_2DIFF + LZ4. Time uses TS_2DIFF + LZ4. The BOOLEAN policy is RLE + LZ4, although this source has no BOOLEAN field. TAG columns use the TsFile table/device mechanism.

Videos

The 100 source MP4 files (about 492 MB total) are not included here. They are 640x480 AV1 at 30 fps without audio, with 50 episode files in each stream:

• observation.images.laptop
• observation.images.phone

Use episode_index and frame_index to align TsFile rows with the two original per-episode videos and meta/episodes.jsonl.

Read Example

from tsfile import TsFileReader

reader = TsFileReader("data/vladfatu_so100_office_train.tsfile")
columns = [
    "episode_index", "task_index", "frame_index", "sample_index",
    "action_0", "observation_state_0",
]
with reader.query_table("vladfatu_so100_office_train", columns, batch_size=65536) as result:
    print(result.read_arrow_batch().to_pandas().head())
reader.close()