THULab/droid_merged_skills_scooping

Merged DROID Skill Dataset: scooping (TsFile)

This dataset is a TsFile conversion of jellyho/droid_merged_skills_scooping, a merged LeRobot v2.1 / DROID dataset for the scooping skill with Franka robot episodes. Modalities: Time-series. The source dataset license is apache-2.0.

Source Dataset

The original dataset merges a language-filtered DROID skill subset with newly collected environment data for the same skill.

Sources:

Source dataset	Episodes
jellyho/droid_subsets_scooping	300
jellyho/droid_scoop_candy	19

Source scale and split:

Item	Value
Episodes	319
Frames / converted rows	102,576
Unique tasks	301
Sampling rate	10 fps
Split	train 0:319
Source episode parquet files	319
Source videos	638

The compact LeRobot training fields are standardized as 8D vectors in the source data:

observation.state = [joint_position_0..6, gripper_position]
action            = [joint_velocity_0..6, gripper_position]

For original DROID subset rows, action is derived from action.joint_velocity plus action.gripper_position. For newly collected rows, action is derived from observation.joint_velocities[:7] plus the collected gripper position from the original action vector.

Converted Layout

This repository contains one TsFile for the train split:

data/droid_merged_skills_scooping.tsfile

Additional source metadata is included under meta/, plus episodes.csv and merge_summary.json. Source video files are not mirrored in this repository. They remain available in the original dataset under videos/. The original video keys are:

• observation.images.wrist_left
• observation.images.side_view_1_left

TsFile Schema

Time column:

• Time: integer milliseconds, computed as round(timestamp * 1000).

TAG columns:

• episode_index
• task_index

FIELD columns:

• frame_index
• sample_index, renamed from the source index column
• scalar task and language fields such as language_instruction, language_instruction_2, language_instruction_3, task_category, and prompt
• scalar metadata fields such as building, collector_id, and date
• logical boolean flags such as is_first, is_last, is_terminal, and is_episode_successful, retained as scalar fields
• scalar reward/discount and all flattened state, action, and camera extrinsic measurements

Flattened vector groups:

Source column	Converted fields
observation.state.cartesian_position	observation_state_cartesian_position_0 ... _5
observation.state.joint_position	observation_state_joint_position_0 ... _6
observation.state	observation_state_0 ... _7
action.cartesian_position	action_cartesian_position_0 ... _5
action.cartesian_velocity	action_cartesian_velocity_0 ... _5
action.joint_position	action_joint_position_0 ... _6
action.joint_velocity	action_joint_velocity_0 ... _6
action	action_0 ... _7
camera_extrinsics.wrist_left	camera_extrinsics_wrist_left_0 ... _5
camera_extrinsics.exterior_1_left	camera_extrinsics_exterior_1_left_0 ... _5
camera_extrinsics.exterior_2_left	camera_extrinsics_exterior_2_left_0 ... _5

Conversion Notes

• The conversion uses the generic LeRobot converter in script mode.
• Time = round(timestamp * 1000) in milliseconds and restarts per episode.
• The source timestamp column is not retained as a FIELD because it is equivalent to Time / 1000 seconds.
• The source index column is renamed to sample_index.
• Vector/list columns are flattened into scalar fields by preserving the source prefix, replacing . with _, and appending element indexes.
• Source columns episode_index and task_index are declared as TsFile TAG columns; remaining scalar columns are FIELD columns.
• The converted TsFile contains 102,576 rows, matching the source frame count and staged Parquet row count.
• Source videos are intentionally excluded from this converted repository; video alignment metadata remains in meta/info.json.

Minimal Read Example

from huggingface_hub import hf_hub_download
from tsfile import TsFileReader

path = hf_hub_download(
    repo_id="zjt24/droid_merged_skills_scooping",
    repo_type="dataset",
    filename="data/droid_merged_skills_scooping.tsfile",
)

reader = TsFileReader(path)
schemas = reader.get_all_table_schemas()
print(schemas.keys())

with reader.query_table(
    "droid_merged_skills_scooping",
    ["episode_index", "task_index", "frame_index", "observation_state_0"],
    batch_size=1024,
) as result:
    batch = result.read_arrow_batch()
    print(batch)