rhoyn.

Teleop data your policy can learn from.

Your task, recorded on our in-house VR teleop setup.

OpenArm · Jenga · sample episode single clock · 30 Hz00:00.000frame 000Unitree G1 · hand latency20.3 ms
Left wrist · D405 Column · D435F Right wrist · D405
Left wrist · D405Column · D435FRight wrist · D405
Unitree G1 · hand latency20.3 ms

Three cameras. One clock.

Every row stamped on one system clock · colour + IR · LeRobotDataset v3.0 · 30 Hz

No room for error

Jenga.

Pull one block out without moving the tower.

Connect four.

Pass a small disc between its hands and drop it into a narrow slot.

Eggs.

Hold a raw egg firmly enough to carry, gently enough not to crack it.

Socks.

Grip a soft sock that folds under the fingers, pass it across and pair it with the other.

01 / 04

Recorded on OpenArm 2.0

See the data for yourself.

Open the sample data →
Format
LeRobotDataset v3.0
Rate
30 Hz · one system clock
Streams
RGB-D + IR
Delivered
Hugging Face or your S3
Licence
Exclusive or non-exclusive

Robots we record

  • Unitree G1 humanoid with two grippers
    Unitree G1 EDU+
  • OpenArm: two arms on a column with a camera on top
    OpenArm 2.0
  • YAM Ultra: a single arm on a base with a gripper
    YAM Ultra 2
  • reBot Arm B601-RS: a single arm on a base with a gripper
    reBot Arm B601-RS
  • Stanford UMI: a handheld gripper with a GoPro, two side mirrors and two soft fingers
    Stanford UMI

20.3 ms hand latency.

Less than one frame at 30 Hz.

From the operator’s hand moving to the robot’s hand moving.

How we measured it →

A 50-episode data pilot.

Your robot, your task. Train on them, keep them.

You’ll talk to founders: Adam and Amelia Jochna.

We publish our measurements.