G1_Dex1_Pack_PingPong

This dataset was converted from unitreerobotics/G1_Dex1_Pack_PingPong to Apache TsFile format.

Dataset Description

The original dataset was created using LeRobot.

• Homepage: UnifoLM-VLA-0
• License: apache-2.0
• Task Objective: Store the ping-pong paddle in its case and zip it.
• Operation Duration: Each operation takes approximately 20 to 40 seconds.
• Recording Frequency: 30 Hz.
• Robot Type: 7-DOF dual-arm G1 robot.
• End Effector: Gripper.
• Dual-Arm Operation: Yes.
• Image Resolution: 640x480.
• Camera Positions: Wrist-mounted monocular camera plus head-mounted binocular cameras.
• Data Content: robot current state, robot next action, and current camera view images in the original dataset.
• Robot Initial Posture: The first robot state in each dataset entry.
• Object Placement: Randomly placed within the robot arm motion range and the field of view of the head-mounted camera.
• Camera View: Follow Part 5 of the AVP Teleoperation Documentation.

Important notes from the original dataset:

1. Due to the inability to precisely describe spatial positions, adjust the scene to closely match the first frame of the dataset after installing the hardware as specified in Part 5 of the AVP Teleoperation Documentation.
2. Data collection was not completed in a single session, and variations between data entries exist. Account for these variations during model training.

Source Dataset

• Original dataset: unitreerobotics/G1_Dex1_Pack_PingPong
• Source scale: 200 episodes, 160,593 frames, 30 Hz, 1 task, 800 videos.
• Source robot type: G1_ALL.
• Source data path: data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet
• Source video path: videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4

TsFile Conversion

The numeric robot state/action/time/task data was converted into one TsFile:

data/g1_dex1_pack_pingpong.tsfile

Conversion details:

• All 200 source episodes are stored in one TsFile table named g1_dex1_pack_pingpong.
• episode_index and task_index are TsFile TAG columns. Query a single episode with WHERE episode_index=N.
• Time = round(timestamp * 1000) in milliseconds. The source timestamp column is not retained because it equals Time / 1000 seconds.
• frame_index is kept. Source index is renamed to sample_index.
• Vector columns are flattened by preserving the source column name, replacing . with _, and appending the element index.
• Values from flattened vector columns are stored as single-precision FLOAT.

Flattened vector features:

• observation.left_arm[7] -> observation_left_arm_0..observation_left_arm_6
• observation.right_arm[7] -> observation_right_arm_0..observation_right_arm_6
• observation.left_gripper[1] -> observation_left_gripper
• observation.right_gripper[1] -> observation_right_gripper
• observation.left_ee[6] -> observation_left_ee_0..observation_left_ee_5
• observation.right_ee[6] -> observation_right_ee_0..observation_right_ee_5
• observation.body[29] -> observation_body_0..observation_body_28
• action.left_arm[7] -> action_left_arm_0..action_left_arm_6
• action.right_arm[7] -> action_right_arm_0..action_right_arm_6
• action.left_gripper[1] -> action_left_gripper
• action.right_gripper[1] -> action_right_gripper
• action.left_ee[6] -> action_left_ee_0..action_left_ee_5
• action.right_ee[6] -> action_right_ee_0..action_right_ee_5
• action.body[7] -> action_body_0..action_body_6

Videos

Videos are not included in this TsFile repository. They remain in the original dataset under videos/.

The mirrored meta/info.json records the original video path and includes a tsfile_conversion section describing the converted schema, row count, tag columns, dropped timestamp column, omitted video features, and time mapping.