OhhO Data

Collect. Label. Ship.

LeRobot dataset recording scripts in the public OmniBot repository. The browser episode viewer is a simulated console (In progress). A record / inspect / push CLI is Vision.

Overview (hero)

Works today: LeRobot dataset recording and processing scripts in the public OmniBot repository. Multi-camera sync and MCAP bags are In progress. The episode viewer on this site is a simulated console (In progress). A Data CLI is Vision.

Highlights

  • LeRobot recording scripts in the public repository
  • MCAP + multi-camera sync (In progress)
  • Episode viewer — simulated console (In progress)
  • CLI record / inspect / push (Vision)

What you get

  • Robot intelligence is only as good as the demonstrations it learns from. OhhO Data is the pipeline that turns teleoperation sessions into clean, training-ready datasets — but only the recording scripts in the public repository ship today.
  • Record episodes from a leader arm and mobile base into standard LeRobot dataset format (Parquet plus MP4) using the two-camera recorder (wrist + bird's-eye) in the public OmniBot repository. Demonstrations capture joint positions on a position-controlled 6-DOF servo arm with a parallel jaw gripper; the physical robot has no force sensors or tactile skin. Multi-camera time sync and MCAP bags are In progress. Curation on this site is a simulated episode viewer (In progress), not a live dataset from your robot. A record / inspect / push CLI is Vision.
  • The planned loop keeps one schema from recorder to trainer to policy. That end-to-end path is not fully built on this site yet. Pick-and-place demonstration evaluation is tracked In progress (OHH-101) without fabricated success tables. Force limiting, tactile feedback, and in-hand manipulation channels are Vision: not built.

Features

  • Synchronized recording — Two-camera recording (wrist + bird's-eye) exists in the public OmniBot repository. Multi-camera time sync across arbitrary camera rigs is In progress.
  • Standard format — LeRobot-compatible Hugging Face datasets (Parquet + MP4) for imitation learning are supported today in the public repository. MCAP recording for raw ROS 2 topic bags is In progress / Vision.
  • Episode viewer — In progress. The /data console scrubs simulated episodes — not recordings from your robot.
  • One schema, end to end — A 9-DOF mobile-manipulation state/action (position-controlled 6-DOF servo arm + base ×3) that matches the recorder, the trainer and the policy — planned across the stack. No force or tactile sensors on the physical arm today.
  • CLI — Vision. Scriptable record / inspect / push commands are not shipped.

How it works

  1. Record — Use the LeRobot recording scripts in the public OmniBot repository on hardware or in simulation.
  2. Review — In progress: open the simulated episode viewer on this site to see the curation UX — not your live dataset.
  3. Ship — Push datasets in LeRobot format to training with your own tooling today; a Data CLI is Vision.
  4. Close the loop — Fine-tune with the OmniVLA engine and deploy via OhhO Serve when those pieces are built for your rig.

Specs

Spec Value
Training format LeRobot HF dataset (Parquet + MP4) — public repo scripts
ROS 2 recording MCAP (In progress / Vision)
State / action 9-DOF (position-controlled arm ×6 + base ×3) — target schema
Tactile / force Vision: not built (position-controlled servo + parallel jaw today)
Cameras Two-camera (wrist + BEV) today; multi-camera sync In progress
Sync tolerance ~50 ms (In progress)
Episode viewer In progress — simulated console on /data
CLI Vision — record / inspect / push not shipped
Targets Real robot, Gazebo (Isaac Sim: Vision / not built)

Plans

There is no paid plan. The console demo is free. Hosted cloud is a waitlist, not a checkout.

FAQ

Is my data portable? Completely. It's standard LeRobot format — train with OhhO's engine or any compatible toolchain.

Can I collect in simulation? Yes. Data records from Gazebo with the same schema as the real robot. Isaac Sim collection is Vision and not built.

Can Data record force or tactile demonstration channels? No. The physical robot uses a position-controlled 6-DOF servo arm with a parallel jaw gripper without force sensors or tactile skin. Force limiting, tactile telemetry, and in-hand manipulation are Vision: not built.

Related products