THULab/LSY-lab_drawer_v1

Drawer V1 TsFile

This dataset is an Apache TsFile conversion of LSY-lab/drawer_v1, a LeRobot v2.1 dataset of Franka robot demonstrations for the task "pick the lego block."

Modalities: Time-series and tabular numeric robot data. The converted repository contains Cartesian, joint, force/torque, tactile, target, gripper, action, frame, episode, and task signals. Camera videos remain in the original Hugging Face dataset.

Source Dataset and Provenance

• Original dataset: LSY-lab/drawer_v1
• Pinned source revision: f53bc15b51e833f3a2103cb4e9d0474db0527f6c
• Original repository contributor/uploader: Daniel San Jose Pro (danielsanjosepro)
• Source organization: Learning Systems Lab (LSY-lab)
• License: Apache-2.0
• Robot type: franka
• LeRobot codebase version: v2.1
• Task: pick the lego block. (task_index = 0)
• Split: train
• Sampling rate: 15 fps
• Scale: 75 episodes, 25,156 frame rows, 1 task, 75 source Parquet files, and 225 source videos
• Source frame layout: data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet
• Source video layout: videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4

The source README embeds a stale one-episode / 381-frame meta/info.json example. At the pinned revision, the repository meta/info.json, 75 Parquet files, episode metadata, and video tree all describe the current 75-episode, 25,156-frame dataset; those files are authoritative for this conversion.

Converted Files

• TsFile: data/drawer_v1_train.tsfile
• Table: drawer_v1_train
• Rows: 25,156
• Episodes/devices: 75
• TsFile size: 9,182,224 bytes (compact encoded build)
• Previous uncompressed/plain build: 10,484,129 bytes; compact build is 1,301,905 bytes (12.4%) smaller.
• SHA-256: 834d0d15f54e8c0d90b69109c5498e335344035fb8823e2bbe085b74d3f68a87
• Time precision: milliseconds
• Metadata: meta/ is mirrored from the source, with meta/info.json rewritten to describe the TsFile artifact and conversion mapping.
• Reproducibility config: LSY-lab_drawer_v1.yaml

TsFile Schema

Time is an INT64 millisecond value computed as round(timestamp * 1000) and restarts for each episode.

Role	Columns	Representation
TIME	Time	INT64 milliseconds
TAG	episode_index, task_index	Source episode and task dimensions
FIELD	frame_index, sample_index, observation_state_gripper	Scalar identifiers and gripper state
FIELD	observation_state_cartesian_0 ... observation_state_cartesian_5	6 FLOAT Cartesian values
FIELD	observation_state_joints_0 ... observation_state_joints_6	7 FLOAT joint values
FIELD	observation_state_sensors_ft_sensor_0 ... observation_state_sensors_ft_sensor_5	6 FLOAT force/torque values
FIELD	observation_state_sensors_tactile_sensor_0 ... observation_state_sensors_tactile_sensor_14	15 FLOAT tactile values
FIELD	observation_state_target_0 ... observation_state_target_5	6 FLOAT target values
FIELD	observation_state_0 ... observation_state_40	Full 41-element FLOAT source state
FIELD	action_0 ... action_6	7 FLOAT action values

Conversion Notes

• The shared config-driven lerobot converter is used. The dataset-specific conversion script and validation reports are retained locally and are not uploaded.
• The train split is merged into one table-model TsFile. Filter by episode_index and task_index to select an episode or task.
• Vector columns are flattened to scalar FLOAT fields. Full source prefixes are preserved, with periods replaced by underscores.
• The source timestamp column is dropped after Time synthesis because it is redundant with Time / 1000 seconds.
• The source index column is retained as sample_index; frame_index is retained unchanged.
• All 25,156 source rows and all 88 source vector elements per row are retained. Together with three scalar FIELD columns, the TsFile contains 91 FIELD columns, two TAG columns, and one TIME column.
• The generated file passed full Apache TsFile Java SDK readback: 25,156 rows, 75 episode/task devices, and no duplicate (episode_index, task_index, Time) keys.

Encoding and Compression

The TsFile uses an explicit compact policy so the binary remains smaller than the source-oriented plain build:

TsFile category/type	Encoding	Compression
FLOAT, DOUBLE fields	GORILLA	LZ4
INT32, INT64 fields	TS_2DIFF	LZ4
Time (INT64)	TS_2DIFF	LZ4
BOOLEAN fields (if present)	RLE	LZ4
TAG/device segments (episode_index, task_index)	TsFile table TAG/device mechanism; PLAIN tag values	LZ4

The policy was applied to the generated schema and confirmed with the Apache TsFile 2.2.1 reader. This dataset has no BOOLEAN field after LeRobot normalization; the BOOLEAN rule is retained in the reproducibility config for consistent conversion of related datasets.

Videos

Videos are not duplicated in this converted repository. At the pinned source revision, the videos tree contains 225 frame-aligned MP4 files (75 per stream) and is approximately 295 MB:

• observation.images.primary
• observation.images.tactile
• observation.images.wrist

Numeric rows remain aligned with the original per-episode videos through episode_index, frame_index, and the source episode metadata.

Minimal Read Example

from tsfile import TsFileReader

reader = TsFileReader("data/drawer_v1_train.tsfile")
table_name = "drawer_v1_train"
columns = [
    "episode_index",
    "task_index",
    "frame_index",
    "sample_index",
    "action_0",
    "observation_state_cartesian_0",
]

with reader.query_table(table_name, columns, batch_size=65536) as result:
    batch = result.read_arrow_batch()
    print(batch.to_pandas().head())

reader.close()

Citation

The source dataset card does not provide a paper or completed citation. Cite the original LSY-lab/drawer_v1 dataset, Daniel San Jose Pro (danielsanjosepro), and Learning Systems Lab when using this converted artifact.