portal-rebot-pick-microcontroller

A LeRobot dataset of a 6-DOF arm picking a microcontroller and placing it in a box. Episodes were collected by teleoperating a real robot over the network with LiveKit Portal, using a human-in-the-loop recording loop that aligns camera frames and joint state into clean, synchronized trajectories.

Task: Pick the microcontroller and place it in the box

At a glance

Robot	Seeed reBot Arm B601-DM (seeed_b601_dm_follower)
Episodes	50
Frames	27,502
FPS	30
Cameras	arm_camera, overhead_camera (640×480 RGB, AV1)
LeRobot codebase version	v3.0
License	Apache-2.0

Features

• observation.state — 7 absolute joint positions (degrees): shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_yaw, wrist_roll, gripper.
• action — same 7 joint positions, the commanded targets (from human teleop or a policy during handoff).
• observation.images.arm_camera — arm-mounted RGB camera, 640×480, 30 fps.
• observation.images.overhead_camera — overhead RGB camera, 640×480, 30 fps.

How it was collected

Data was recorded with LiveKit Portal's human-in-the-loop pipeline: a robot publishes camera frames and joint state, an operator (human teleoperator and/or policy) drives it, and every executed action is recorded tagged with its sender. The synchronizer stitches the separately-arriving camera and state streams back into one coherent snapshot per tick before writing each frame to the LeRobotDataset episode.

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("binhpham/portal-rebot-pick-microcontroller")
print(dataset)
print(dataset[0])

Versions

• v3 — 50 episodes. A stray, partially-recorded frame fragment (from an accidental record toggle) was dropped during repacking; all 50 intended episodes are intact.