CALBEAR

Video in. Robot training data out.

CalBear turns video of people doing skilled work into robot training data, in minutes. Head-mounted, chest-mounted or third-person — any camera that sees the hands.

See how it works

Upload video. Pick a robot. Download the dataset.

Six stages run on our GPUs: blur, segment, reconstruct hands, retarget, align, score. You review the episodes; we package LeRobot, RLDS, HDF5 or Zarr.

Explore the pipeline

From footage to policy in an afternoon

  1. 1Upload video

    Head-mounted, chest-mounted or third-person footage goes straight from your browser to storage.

  2. 2Process

    Faces and screens are blurred first. Then segmentation, hand reconstruction and retargeting to your robot.

  3. 3Review

    Every episode gets a quality score and an instruction. Keep what's good, drop what isn't.

  4. 4Download

    The dataset in the formats you chose, plus an offline toolkit, in one kit with checksums.

  5. 5Test on your robot

    Replay, train and evaluate on your own machine. Nothing about your robot connects to us.

Robots

93 robots. One upload.

Arms, bimanual rigs, humanoids and mobile manipulators. Pick the embodiment when you create a dataset; the same footage can be retargeted to another robot later without re-uploading. Joint names, limits and gripper conventions match each robot's LeRobot layout.

See the full list →

Missing yours? A profile is a 20-line file. Request a robot and we usually add it within a few days.

Privacy is a pipeline stage, not a policy.

Faces, badges and screens are blurred at ingest, before any model sees a frame, and every episode carries a chain of custody back to its source video.

Nothing about your robot ever connects to us.

The download includes an offline toolkit: verify the files, replay trajectories with safety limits, train with LeRobot, and evaluate on your own hardware.

Start with one clip.

Create a workspace, upload a short clip from any camera, and review your first episodes a few minutes later.