An operator, a robot, and a task the policy has to learn.
Trained operators drive leader–follower arms or VR-controlled robots through repeatable tasks. Every episode logs synchronised camera streams with joint states and actions — the data a policy is actually trained on.
- Environment
- Small robotics workshop, LED work lamp
- Captured with
- Bimanual leader–follower arms, two camera views
- Task
- Stack three bowls
Perit has shipped 4,000+ hours of speech through this bench. Physical-AI capture runs on the same recruiting, consent and QA machinery, and the first collections are being scoped with early-access partners now. Nothing on these pages is a delivered volume — the first hours will be listed here the way our speech hours are, after they ship.
What gets captured, and why it is worth having.
Teleoperation is the embodiment-matched end of the spectrum: the demonstration is performed on the robot the policy will run on, so there is no transfer gap to close. It is slower and dearer per episode than human video, which is why the protocol matters — every episode has to count.
Operators pass the same calibration gate as the rest of the bench, on the task and the rig, before a paid episode is recorded. Interventions and failed takes are logged, not discarded, because they are training data too.
- Behaviour cloning and diffusion policies
- Policy evaluation and intervention data
- Recovery-from-failure demonstrations
- Embodiment-matched fine-tuning
- Primary rig
- Leader–follower bimanual rigs (ALOHA-class)
- Also
- VR teleoperation — headset and controllers
- Also
- Your robot, our operators: bring-your-own-stack
- Also
- Multi-view cameras: wrist, front, overhead
- Cameras
- 2–4 synchronised RGB streams, 480×640 to 1080p
- Frame rate
- 30 fps
- State
- Joint positions, velocities and gripper state per step
- Actions
- Leader-arm or controller commands per step
- Format
- LeRobot- or RLDS-style layout on request; HDF5 or Parquet
- Episode
- 20 seconds – 3 minutes
Where it is recorded.
What people are asked to do.
Tasks are recorded where they actually happen — not in a studio dressed to look like a kitchen. Hover a room to see the kind of setting the protocol calls for.
- Synchronised camera streams per episode
- State and action logs per step
- Task and environment metadata per episode
- QC status and reviewer chain per episode
- Consent reference on every file
- Quality report per batch: accepted, rejected, and why
- Weekly drops to your bucket, region pinned at kickoff
- Success and failure labels per episode
- Sub-task boundaries
- Language instruction per episode
- Operator interventions flagged
Labelled on the same bench as our speech work, to a written guide, behind the same calibration test.
Small first, on purpose.
- 01Briefday 0
The task family, the environments, the rig, the sensors, the episode count. One page, agreed before anything is bought or anyone is recruited.
- 02SOP and kitweek 1
A written capture protocol — framing, lighting, where a task starts and ends, what counts as a failed take — and the kit list. Operators are trained and tested on it before a single episode is paid for.
- 03Pilot episodesweek 2
A small batch, ten to thirty episodes, delivered with QC status and a quality report. This is where the protocol breaks — on purpose, and cheaply.
- 04Review and scaleongoing
You review the pilot, we fix the protocol, then the run scales on a weekly delivery cadence with the same report attached to every batch.
Scope a teleoperation pilot.
The task family, the environment, the rig and the episode count — one page, and we come back with a protocol and a quote.