openrob-furg/openarm-single-pick-place-lego

OpenArm Single Arm Pick and Place (Lego)

This dataset contains 200 teleoperated demonstrations of a pick and place task recorded on the OpenArm platform. Only the right arm is used. The data is stored in the LeRobot format (codebase version v3.0). It was collected by the OPENROB project at FURG and used to fine tune SmolVLA policies.

Task

A single white Lego piece is picked up from the table and placed inside a container. One Lego piece and one container are present in each episode. All episodes share a single task.

Language instruction stored in the task field: put the lego brick in the box

Hardware

Component	Description
Robot	OpenArm follower (openarm_follower), right arm only
Wrist camera	Intel RealSense D405, mounted on the right arm wrist (observation.images.right_cam_wrist)
Chest camera	Intel RealSense D435i, mounted on the chest (observation.images.cam_chest)

Dataset summary

Property	Value
Episodes	200
Frames	47,728
Frame rate	30 fps
Total duration	about 26.5 minutes (about 8 seconds per episode on average)
Tasks	1
Camera streams	2 RGB streams, 640 × 480, AV1 encoded
Depth	Not included
Split	train: episodes 0 to 199

Features

Feature	Shape	Description
observation.state	8	Positions of right arm joints 1 to 7 and the right gripper
action	8	Target positions of right arm joints 1 to 7 and the right gripper
observation.images.right_cam_wrist	480 × 640 × 3	RGB video from the D405 wrist camera
observation.images.cam_chest	480 × 640 × 3	RGB video from the D435i chest camera
timestamp, frame_index, episode_index, index, task_index	1	Standard LeRobot indexing fields

The state and action vectors share the same joint ordering: right_joint_1.pos to right_joint_7.pos, followed by right_gripper.pos.

Videos are encoded with AV1. Decoding requires an FFmpeg or PyAV build with AV1 support.

Subsets used in the paper

The dataset was recorded in independent sessions and merged into a single repository. The training subsets used in the paper are defined by episode index.

Subset	Episode indices	Episodes
n50	0 to 49	50
n100	listed here when the n100 model is released	100
n200	0 to 199	200

Usage

Load the full dataset:

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("openrob-furg/openarm-single-pick-place-lego")

Load a single training subset, for example n50:

dataset = LeRobotDataset(
    "openrob-furg/openarm-single-pick-place-lego",
    episodes=list(range(0, 50)),
)

The same subset can be selected during training with the dataset.episodes option of the LeRobot training script.

Related models

Model	Training subset
openrob-furg/smolvla-openarm-single-pick-place-lego-n50	n50
openrob-furg/smolvla-openarm-single-pick-place-lego-n100	n100
openrob-furg/smolvla-openarm-single-pick-place-lego-n200	n200

Changelog

• 2026-10-06: two episodes with failed grasp attempts were removed and replaced by two new recordings (episodes 198 and 199). Episodes 0 to 99 are unchanged; later episodes were renumbered.

License

This dataset is released under the Apache 2.0 license.