Robotics teleoperation data. Episodes a policy can learn from.
Teleoperated demonstrations recorded by operators trained on your task, not whoever was free that afternoon. Joint states, actions, poses, and camera streams aligned per episode, graded against written success criteria, with consent on file for anyone in frame and signed provenance on every episode.
What is robotics teleoperation data?
Robotics teleoperation data is a collection of recorded episodes in which a person drives a robot through a task, using a leader arm, a VR rig, an exoskeleton, or a game controller, while the robot logs what it sensed and what it was told to do at every timestep. Enough well-chosen episodes and a policy can learn the task by imitation.
The catch is that a policy learns everything in the episodes, including the bad habits. A nervous operator teaches a nervous policy. Three operators with three strategies for the same drawer teach a model that cannot decide. Variation that came from mistakes rather than from the protocol teaches noise.
eeko's data practice records episodes the other way around: a written task protocol first, operators qualified against it, then production, with a review sample on everything that ships. It is the discipline behind our enterprise AI agents, pointed at the physical world. Operators are paid through our paid AI-training programs, and every session follows the rules on our data capture page.
Teleoperation datasets, stream by stream.
An episode is only as useful as the streams inside it and the metadata around it. Each piece below is agreed in the specification before the first session.
Robot demonstration data starts with the operator.
The policy inherits the operator's habits, so the operator is the first thing we get right.
From a task description to a qualified operator
- Protocol — initial conditions, object set, allowed strategies, variations, success criteria, and the rule after a failed attempt, written with your engineers and versioned.
- Interface training — operators from the eeko teleoperation pool learn your rig until only the task remains.
- Qualification — scored episodes per operator. Hesitation, off-protocol strategies, and sloppy resets fail the gate.
- Production — sessions follow a variation schedule, so coverage comes from the plan, not from chance.
- Live review — reviewers sample as it happens and pull an operator for retraining the first time an episode drifts.
- Records — every episode names its operator, reviewer, protocol version, and rig.
Operators are paid per accepted hour at a rate stated before they start, so they stay for the whole project and your episodes look like they came from one hand.
Imitation learning data, scoped to your embodiment.
Spec and pilot batch first, then production in whichever of these forms you need.
Six things we produce for robot learning teams
- Teleoperated episodes — tabletop, bimanual, and mobile manipulation recorded to your protocol on your platform or ours.
- Synchronized streams — proprioception, actions, poses, cameras, and contact sensors aligned per episode, one table plus linked media.
- Outcome labels — success, failure, and recovery, with failure reasons where the spec calls for them.
- Phase and event annotation — sub-task boundaries, grasps, releases, and contacts marked by reviewers.
- Egocentric human video — head- or chest-mounted recordings of people doing the task by hand, consent recorded per session, for cross-embodiment pretraining.
- Protocol design — if you have a task but no protocol yet, we write one with your team and pilot against it.
Adjacent needs: licensed video datasets for world-model pretraining, and 3D and CAD datasets for the objects your robot will meet.
Every episode fingerprinted, every delivery signed.
A checkpoint should be traceable to the demonstrations behind it. That takes a record on each episode, not a line in a contract.
Teleoperation datasets by industry.
From task description to a sealed dataset.
Fixed scope, fixed price per accepted episode, agreed after the pilot. A real person replies inside one business day.
Robotics teleoperation data questions.
What is robotics teleoperation data?
Robotics teleoperation data is a set of recorded episodes in which a human operator drives a robot through a task with a controller, VR rig, exoskeleton, or leader arm, while joint states, commanded actions, end-effector poses, gripper state, and camera views are logged in sync. Each episode is a worked example a policy can imitate, and the set is only as good as the operators' consistency, the chosen variations, and the labels on each episode.
How much does robotics teleoperation data cost?
Cost is set per accepted episode and moves with task length and difficulty, the number of protocol variations, the sensors recorded, annotation depth such as phase boundaries, whether sessions run on your hardware or ours, and exclusivity. A short tabletop pick-and-place costs far less per episode than a long, contact-rich assembly. We price the production run only after a pilot batch, so the quote reflects measured throughput rather than a guess.
Can you record on our robot, or do you supply the hardware?
Either. If the embodiment matters, and for most policy work it does, we record on your platform and interface, at your site or on units shipped to our operators. Where the task is embodiment-agnostic, or you need human demonstrations rather than robot ones, we record with our own rigs and egocentric cameras. Setup, calibration routine, and episode schema are fixed during scoping so every session logs the same thing the same way.
What format do teleoperation episodes come in?
Episodes ship as WebDataset shards for streaming into training, as Parquet tables for the numeric streams with linked media files, or as JSON Lines with one record per episode, with Croissant metadata describing the set. If your stack expects a particular episode layout, such as specific key names or a timestep convention, we write it into the specification and check every delivery against it before it leaves us.
Do you include failed attempts in the dataset?
Only when your specification asks for them. Many teams keep failures and recoveries, labeled with the failure reason and the phase where it happened, because some methods learn from them and they make useful evaluation cases. Others want clean successes only. Either way, the signed receipt lists exactly which episodes were delivered, so an unlabeled failure never slips into a set that was meant to be successes only.
How do you handle people who show up in the camera streams?
Scene and head-mounted cameras capture operators, bystanders, and places. Anyone identifiable in an episode offered for training has consent recorded against that episode, and participants in egocentric sessions agree in advance to what is captured and how it may be used. Location fields and other personal metadata are removed before delivery. A recording that includes someone without consent on file is held back, not shipped.
Explore related data.
Tell us the task. We will write the protocol.
Send the task, the embodiment, and the sensors you log. A robotics data lead replies inside one business day with a protocol outline and a pilot plan. No deck. A ranked list of what to record first.
Experienced within
Markets served.
As an enterprise AI agency, eeko systems delivers production AI systems remote-first across the United States and internationally — including these markets: