THULab/pickup_single_cotton_v2

Pickup Single Cotton V2 TsFile

This dataset is a TsFile conversion of konstantinZ/pickup-single-cotton-V2, a LeRobot v2.1 Panda follower robotics dataset.

Modalities: Time-series. The original camera videos are not uploaded here; they remain available in the source dataset under videos/.

Source Dataset

• Source: konstantinZ/pickup-single-cotton-V2
• License: apache-2.0
• Created using: LeRobot
• Codebase version: v2.1
• Robot type: panda_follower
• Split: train 0:50
• Scale: 50 episodes, 24,994 frames, 1 task, 1 data chunk, 150 videos
• Sampling rate: 30 fps
• Task: "Pick and place sorting task of medical equipment with the Eranka Emika Panda robot and lerobot."

Converted Files

• TsFile path: data/pickup_single_cotton_v2.tsfile
• Table name: pickup_single_cotton_v2
• Row count: 24,994
• Time precision: milliseconds
• Metadata: meta/ is mirrored from the source dataset, with meta/info.json updated to describe the converted TsFile artifact and video policy.

Schema

Time is generated as round(timestamp * 1000) in milliseconds. Time restarts within each episode, and episode_index plus task_index identify the series.

TAG columns:

• episode_index
• task_index

FIELD columns:

• frame_index
• sample_index converted from the source index
• action_0 to action_8
• observation_state_0 to observation_state_8

The action and observation vectors use these source element names in order:

• panda_joint1.pos
• panda_joint2.pos
• panda_joint3.pos
• panda_joint4.pos
• panda_joint5.pos
• panda_joint6.pos
• panda_joint7.pos
• panda_finger_joint1.pos
• panda_finger_joint2.pos

Conversion Notes

• All 50 train episodes are stored in a single TsFile using the TsFile table model, with episode_index and task_index as TAG columns.
• The source timestamp column is not retained as a FIELD because it is the source for Time and equals Time / 1000 seconds after conversion.
• The source index column is renamed to sample_index.
• action[9] is flattened to scalar FLOAT fields action_0 to action_8.
• observation.state[9] is flattened to scalar FLOAT fields observation_state_0 to observation_state_8.
• Video features are omitted from this repository: observation.images.overhead, observation.images.front, and observation.images.wrist. Use the original dataset videos linked above for frame-aligned visual data.

Minimal Read Example

from tsfile import TsFileReader

path = "data/pickup_single_cotton_v2.tsfile"
reader = TsFileReader(path)

columns = [
    "episode_index",
    "task_index",
    "frame_index",
    "sample_index",
    "action_0",
    "observation_state_0",
]

with reader.query_table("pickup_single_cotton_v2", columns, batch_size=1024) as result:
    batch = result.read_arrow_batch()
    if batch is not None:
        df = batch.to_pandas()
        print(df.head())