Robot teleoperation data. Demonstrations worth imitating.
A policy trained by imitation inherits every hesitation, every inconsistent grasp, and every failure somebody labeled a success. eeko commissions vetted operators on leader arms, VR rigs, and custom setups to record consistent demonstrations and deliberate recoveries, label every outcome honestly, and put a second reviewer on each episode before it ships.
What is robot teleoperation data?
Robot teleoperation data is a library of recorded episodes in which a person drives a robot through a task by hand, through a leader arm, a VR controller, or a haptic device, while the system logs what the cameras saw, where each joint was, what action was commanded, and what the gripper did, timestep by timestep. Imitation learning trains on those episodes directly. There is no reward function standing between the operator's habits and the policy's behavior.
That is why demonstration datasets fail in such specific ways. Jerky motion from an operator still learning the rig. Three different grasp strategies for the same mug across thirty episodes. A dropped object that nobody marked as a failure, so the policy learns that dropping is fine. And, almost universally, no examples of what to do after a slip, because clean demonstrations are what people record and recoveries are what they delete.
eeko's data practice treats operators as the specialists they are. We vet them for steadiness and consistency, give them a written task specification for every episode, commission recoveries on purpose, and confirm every outcome label with a second reviewer. It sits inside our AI training data practice and is the human side of our robotics teleoperation data solution, recorded through the same data capture process we use for every modality.
Teleoperation datasets, built to a task spec.
Six lines of work. Each starts with a written task specification, a confirmed rig, and a paid pilot batch, and each ends with a signed receipt listing every episode.
Four defects that poison an imitation dataset.
Each of these is invisible in a dataset summary and obvious in the policy. Each is prevented by the operator, the specification, or the reviewer rather than by post-processing.
What the operator controls and the model inherits
- Inconsistent strategy — the same task solved three ways across the dataset leaves the policy averaging between them. The task spec fixes approach, grasp, and pacing, and reviewers reject episodes that drift.
- No recovery data — a policy that has only seen clean runs freezes or flails after its first slip. We script disturbances and record the recovery as its own task.
- Mislabeled outcomes — a failure marked as a success teaches the model that failing is the goal. Every label is confirmed by a second reviewer who watches the episode.
- People in frame — a bystander, a face, a badge on a lanyard. Consent and releases are collected before recording, and video is screened before anything is delivered.
None of these can be fixed later with a filter. They are fixed at the rig, by a trained operator working to a written specification.
Paid teleoperation work for steady hands.
The eeko expert network pays operators to record demonstrations, perform recoveries, label episodes, and review other operators' work. The hardware, the setup, the rate, and the schedule are stated on every job before you accept it.
Who we look for
- Teleoperators — hands-on time driving robot arms, mobile robots, or similar systems.
- VR and leader-arm fluency — comfortable with controllers, headsets, or leader-follower setups.
- Precision manual workers — assembly, lab technique, surgical or dental assisting, or a skilled trade.
- Remote equipment operators — drones, cranes, excavators, or other remotely driven machinery.
- Patience — the willingness to repeat a task forty times to the same written specification.
What the jobs involve
- Record — perform the task through the rig, episode after episode, to spec.
- Recover — demonstrate safe, deliberate recoveries from scripted disturbances.
- Label — mark outcomes, failure types, step boundaries, and instructions.
- Review — watch another operator's episodes and confirm or correct their labels.
Tradespeople and machine operators without robotics experience often qualify. Our AI training jobs for skilled trades page covers related capture work, and the expert data annotation jobs page explains onboarding, rates, and payouts across the network.
Consent, fingerprints, and provenance on every episode.
From a task description to episodes in your pipeline.
Fixed price per accepted episode, agreed on the pilot and held for production. Recording does not scale until your training pipeline has ingested real episodes.
Robot teleoperation data questions.
What is robot teleoperation data?
Robot teleoperation data is a set of recorded episodes in which a human operator drives a robot through a task with a leader arm, VR controllers, or a similar interface, while the system logs what the cameras saw, where every joint was, what action was commanded, and what the gripper did at each timestep. Imitation learning trains directly on these episodes, so the policy inherits whatever the operator did, including hesitation and mislabeled failures.
How much does teleoperation data cost?
We price per accepted episode, and the unit price depends on task complexity, the rig, the number of synchronized streams, and how much labeling each episode needs. A single-arm pick-and-place costs less than a bimanual deformable-object task with step-level language annotation. After scoping we quote a fixed price for a pilot batch, and the same price holds for the production run. Episodes that fail second review are not billed.
Which robots and teleoperation rigs do you support?
Requests are scoped by robot platform, end effector, camera and sensor layout, and teleoperation interface, whether leader-follower arms, VR controllers, haptic devices, or a custom setup. We confirm feasibility during scoping, including whether recording happens on your hardware, at an operator site, or in a session we run, then staff it with operators who have logged time on that class of rig.
What does a delivered episode contain, and in what format?
Each episode bundles time-aligned camera video, joint states, commanded actions, and gripper states, plus an outcome label, a failure type where relevant, step boundaries, and any language instruction. Episodes ship as WebDataset or Parquet with Croissant metadata, or mapped to your schema. Every episode is fingerprinted with frame hashes and listed by identifier and hash on the signed receipt.
Do I need robotics experience, and how are operators paid?
Not necessarily. Time on robot arms is the most direct qualification, but VR fluency, precision manual work, a skilled trade, and remote equipment operation all transfer. What we test is steady hands, consistency across repeated episodes, and patience with a written specification. Each job states the rate, the hardware, and the block schedule before you accept it, and approved episodes are credited to your account with payouts after identity and tax verification.
Can we license demonstrations exclusively or set up a standing order?
Yes to both. Commissioned episodes can be licensed to a single buyer, with scope, term, and exclusivity written into the agreement before recording begins. A standing order keeps the same operators, rig, and task specification producing new episodes on a cadence, which is the usual arrangement for a team that keeps adding tasks. Every delivery carries its own signed receipt and audit trail.
Related solutions, industries, and expertise.
Tell us the task your policy keeps failing.
Send the robot, the rig, and the task list. A data lead replies inside one business day with a feasibility view, a spec outline, and a pilot price. Operators: the application takes ten minutes, and the listing tells you the hardware before you commit. No deck. A ranked list.
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: