LeRobot OMX

Real-robot datasets and policies for the ROBOTIS OMX arm (OMX-F follower + OMX-L leader), collected and trained with LeRobot.

What's here

• Datasets — teleoperated demonstration episodes recorded with the OMX leader/follower setup. Current tasks: pick-and-place of small objects (e.g. omx_pick_place_orange_doll). Stored in the LeRobot dataset format, ready for training.
• Models — ACT policies trained on these datasets, plus the omx-setup repo with calibration and configuration files used to reproduce our rig. Fine-tuning vision-language-action models (e.g. SmolVLA) on the same data is planned.

Hardware

• ROBOTIS OMX-AI: OMX-F follower arm (5 DOF + gripper, 400 mm reach, DYNAMIXEL XL430/XL330) and OMX-L leader device for teleoperation
• LeRobot integration: huggingface.co/docs/lerobot/omx
• Cameras: 2× Innomaker USB cameras — one top-view camera over the workspace (observation.images.top) and one wrist-mounted camera on the OMX-F gripper (observation.images.wrist), both 640×480 @ 30 fps
• Control loop: 30 Hz (teleoperation and recording)

Software

• Ubuntu 24.04 LTS · LeRobot 0.6.2 for teleoperation, data recording and policy training.

Quick start

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("lerobot-omx/omx_pick_place_orange_doll")
print(ds.meta.info)

Team

• Yumin Kim
• yanwoo

Questions and collaboration requests are welcome via the Community tab of any repo.