data_real_4tasks

Teleoperated demonstrations of four tabletop tasks on an AgileX Piper arm: 450 episodes, 49,482 frames (27.5 min at 30 fps). Each task is a separate, complete LeRobot v3.0 dataset in its own folder.

folder	prompt	episodes	frames	mean episode
press_button/	press the {blue, orange, pink, red} button	100	5,533	1.8 s
put_spoon_in_bowl/	put the spoon in the bowl	100	13,024	4.3 s
stack_bowls/	stack the bowls	100	11,679	3.9 s
pick_can_place_plate/	pick up the can and place it on the plate	150	19,246	4.3 s

press_button/ holds four prompts, 25 episodes each, in this order:

episodes	prompt	task_index
0-24	press the blue button	0
25-49	press the orange button	1
50-74	press the pink button	2
75-99	press the red button	3

Loading

The repository root is not itself a LeRobot dataset, so download it and point root at one folder:

from huggingface_hub import snapshot_download
from lerobot.datasets.lerobot_dataset import LeRobotDataset

path = snapshot_download("anhdao69/data_real_4tasks", repo_type="dataset")
ds = LeRobotDataset("anhdao69/data_real_4tasks", root=f"{path}/stack_bowls")

To download a single task, add allow_patterns="stack_bowls/*" to snapshot_download.

Format

All four folders share the same layout and features:

<task>/
  meta/info.json, stats.json, tasks.parquet
  meta/episodes/chunk-000/file-000.parquet
  data/chunk-000/file-000.parquet
  videos/observation.images.front/chunk-000/file-000.mp4
  videos/observation.images.wrist/chunk-000/file-000.mp4

• action (7, float32): target position of joint_1 ... joint_6 in degrees (absolute, not deltas), then gripper.pos (0 = closed, about 8 = fully open).
• observation.state (14, float32): the six measured joint positions in degrees, gripper.pos, the end-effector pose (eef_pose_x/y/z, eef_pose_rx/ry/rz) and gripper_effort.
• observation.images.front, observation.images.wrist: 640x480, 30 fps, AV1 (yuv420p).
• robot_type is piper; fps is 30.

Each folder has one data file, one episodes file and one video per camera, and each video holds exactly one frame per data row, in row order: video frame k belongs to row index == k. The from_timestamp / to_timestamp columns in meta/episodes agree with that (dataset_from_index / 30 and dataset_to_index / 30).

In the press-button task the gripper stays closed for the whole episode. Episodes start near the home pose and end when the task is done; the arm does not return home.

Provenance

• pick_can_place_plate/ is an unchanged copy of jachansantiago/pick_can_place_plate_150_trimmed (Apache-2.0).
• press_button/ has the same 100 episodes, rows and pixels as anhdao69/press_button_task. Two things differ: the videos here no longer contain the unused footage between episodes that the other copy has, and the data/file_index and meta/episodes/file_index columns point at the files that exist.
• put_spoon_in_bowl/ and stack_bowls/ were recorded in November 2026 with the same robot, cameras and recorder.

The three recorded tasks were consolidated from one file per episode. Rows are unchanged. Videos were cut by copying compressed packets, without re-encoding, so every frame is pixel-identical to the recording.

Known imperfect demonstrations

Every episode completes its task. These six end less cleanly than the rest and are kept:

folder	episode	what happens
stack_bowls	13, 47	the top bowl ends tilted instead of seated flat
press_button	77	the red button slides sideways while it is pressed
pick_can_place_plate	0, 58, 124	the can tips over onto the plate

In a few stack_bowls episodes (for example 98 and 99) one bowl sits partly outside the front camera's view.