drawer_v1 TsFile

This dataset is a TsFile conversion of the Hugging Face dataset danielsanjosepro/drawer_v1. The source dataset was created with LeRobot and stores Franka robot manipulation episodes for the task Insert the lego block in the drawer.

Modalities: Time-series. This converted repository contains the numeric robot state, action, time, and task data plus selected source metadata. The source camera videos are not mirrored here; they remain in the original dataset under videos/.

Dataset Summary

• Source dataset: danielsanjosepro/drawer_v1
• Source author: danielsanjosepro
• Source license: apache-2.0
• Source format: LeRobot v2.1 per-episode Parquet files plus metadata and videos
• Robot type: franka
• Task: Insert the lego block in the drawer
• Episodes: 76
• Frames / converted rows: 22,566
• Split: train (0:76)
• Sampling rate: 15 fps
• Video streams in source: 152 videos, observation.images.primary and observation.images.wrist
• Converted TsFile: data/drawer_v1_lerobot.tsfile
• TsFile size: 5,560,301 bytes
• Converted table: drawer_v1_lerobot
• Time precision: milliseconds

Converted Schema

The conversion writes one TsFile table, drawer_v1_lerobot, with all source episodes merged into a single TsFile.

• Time: synthesized as round(timestamp * 1000) in milliseconds. Time restarts within each episode, matching the source LeRobot timestamp behavior.
• TAG columns: episode_index, task_index. These are the TsFile device dimensions. Query one episode with a predicate such as WHERE episode_index = 0.
• FIELD columns: frame_index, sample_index, scalar robot state fields, flattened observation vectors, and flattened action vectors.

Schema roles:

Column	Role	Type
Time	TIME (ms)	INT64
episode_index	TAG	INT64
task_index	TAG	INT64
frame_index	FIELD	INT64
sample_index	FIELD	INT64
observation_state_gripper	FIELD	FLOAT
observation_state_cartesian_0 .. observation_state_cartesian_5	FIELD	FLOAT
observation_state_joints_0 .. observation_state_joints_6	FIELD	FLOAT
observation_state_sensors_ft_sensor_0 .. observation_state_sensors_ft_sensor_5	FIELD	FLOAT
observation_state_target_0 .. observation_state_target_5	FIELD	FLOAT
observation_state_0 .. observation_state_25	FIELD	FLOAT
action_0 .. action_6	FIELD	FLOAT

Vector Flattening

Vector columns are flattened into scalar FIELD columns by preserving the source column name, replacing . with _, and appending the element index.

• observation.state.cartesian with shape [6] becomes observation_state_cartesian_0 through observation_state_cartesian_5.
• observation.state.joints with shape [7] becomes observation_state_joints_0 through observation_state_joints_6.
• observation.state.sensors_ft_sensor with shape [6] becomes observation_state_sensors_ft_sensor_0 through observation_state_sensors_ft_sensor_5.
• observation.state.target with shape [6] becomes observation_state_target_0 through observation_state_target_5.
• observation.state with shape [26] becomes observation_state_0 through observation_state_25.
• action with shape [7] becomes action_0 through action_6.

The source metadata names the Cartesian state elements as x, y, z, roll, pitch, and yaw; the joint elements as joint_0 .. joint_6; and the action elements as x, y, z, roll, pitch, yaw, and gripper.

Dropped And Omitted Source Fields

• timestamp is not retained as a separate FIELD because it is the source used to synthesize Time (Time / 1000 seconds).
• index is renamed to sample_index.
• observation.images.primary and observation.images.wrist are video features and are not uploaded in this converted repository. Use the original dataset videos for frame-aligned camera data.

No numeric time-series rows are dropped.

Metadata

This repository includes selected source metadata:

• meta/info.json is rewritten so data_path points to data/drawer_v1_lerobot.tsfile; the original video path is preserved as video_path_original.
• meta/tasks.jsonl, meta/episodes.jsonl, meta/episodes_stats.jsonl, and meta/crisp_meta.json are mirrored from the source metadata.

The tsfile_conversion object inside meta/info.json records the source dataset, converted data path, table name, time mapping, TAG columns, row count, flattened features, renamed fields, dropped fields, omitted video fields, and video policy.

Read Example

Use the Apache TsFile SDK to read data/drawer_v1_lerobot.tsfile. The exact query API depends on the SDK version; the table name is drawer_v1_lerobot, and episode/task identifiers are TAG columns.

from tsfile import TsFileReader

path = "data/drawer_v1_lerobot.tsfile"

with TsFileReader(path) as reader:
    schemas = reader.get_all_table_schemas()
    table = schemas["drawer_v1_lerobot"]
    columns = [c.get_column_name() for c in table.get_columns()]
    print(columns)