Episodes Preview omx_follower Visualizer
50 episodes · 30 fps · 2 cameras · 640×480 av1

omx_tomato

50 teleoperated demonstrations of a ROBOTIS OMX arm picking up a tomato and putting it into a bin. Recorded with LeRobot 0.6.1 (dataset format v3.0).

OMX 리더로 원격 조종해서 팔로워가 토마토를 집어 통에 넣는 시연 50개다. 수업 실습용 소규모 데이터셋이다.

Summary

Item Value
Task "Pick the tomato and place it in the bin"
Episodes / frames 50 episodes, 34,432 frames, 30 fps
Robot OMX follower (omx_follower), teleoperated with an OMX leader arm
State and action 6 joints each: shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper
Cameras top (fixed, looking down at the table) and wrist (on the gripper), 640x480, 30 fps
Video storage AV1, several episodes concatenated per mp4 file

Tomato positions

The tomato was placed at 5 fixed positions on the table, 10 episodes each, recorded in this order. Positions are described as seen from the top camera.

Episodes Position
0-9 1 (center)
10-19 3 (7 o'clock)
20-29 4 (11 o'clock)
30-39 5 (1 o'clock)
40-49 6 (5 o'clock)

The start pose of the arm is nearly the same in every episode (mean over the 50 episodes: pan 2.4, lift -63.7, elbow 55.7, wrist_flex 45.0, wrist_roll 1.6, gripper 50.6). Wrist roll barely changes during the demonstrations.

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("minodori/omx_tomato")
item = ds[0]  # observation.state, action, observation.images.top / wrist, ...

Train an ACT policy with lerobot-train on this dataset. A model trained this way (30,000 steps) is at minodori/omx_tomato_act. The recording, training, and rollout scripts are at minodori/omx-tomato-act on GitHub.

Notes

  • Video files are AV1. Some default players cannot open them (for example the Ubuntu default player). Use ffplay or VLC, or load them through LeRobot.
  • A model trained on this data was sensitive to the top camera position, the start pose, and the background in the top image when run on the real robot. See the model card for details.
  • The top camera shows only the table, the robot, and the background. No faces are visible in the videos.

Structure

meta/info.json holds the feature list and video settings, meta/stats.json the normalization statistics, data/ the state and action tables, and videos/ one folder per camera.

Downloads last month
105

Models trained or fine-tuned on minodori/omx_tomato