Teleoperation Data Collection
As a Service

Why Teleoperation Is Still the Reference Method

When a human drives a robot through a task, the recording contains the robot's own observations and the robot's own actions, aligned on one clock. There is no embodiment gap to close and no action recovery step to trust. The policy trains on the same distribution it will act in. That is why the imitation learning results that defined the current era, from fine manipulation with low-cost arms to the large multi-task corpora behind today's robot foundation models, were built on teleoperated demonstrations.

The cost is throughput. Teleop needs a robot, a rig, and a trained operator for every collection seat, and episodes accumulate one attempt at a time. Running that efficiently is an operations problem, which is precisely why it makes sense to buy it as a service rather than staff it inside a research team.

The Rigs

Leader-Follower Arms

The operator moves a small leader arm and the follower robot mirrors it joint for joint. This ALOHA-style setup is the workhorse for bimanual and fine manipulation work: the kinematic mapping is direct, latency is low, and operators develop real dexterity with practice. It is our default rig for tabletop and garment tasks.

VR and Motion-Tracked Teleop

Headset and controller rigs map hand pose to end-effector targets. They shine for mobile manipulators and humanoid upper bodies, where a leader arm has no natural mapping. The trade is a layer of retargeting between human motion and robot motion that must be tuned per embodiment and validated in the pilot batch.

What the Rig Records

Whatever the input device, the recording is the same: scene and wrist cameras, joint states, end-effector pose, gripper commands, and the action stream in the space you train in. If your controller expects end-effector deltas, we record end-effector deltas. Action space is a spec decision, and getting it wrong is the most common reason bought datasets disappoint.

Operators Are Half the Dataset

Demonstration quality is operator skill made visible. An experienced operator produces smooth, purposeful trajectories, consistent grasp choices, and natural recoveries from slips. A novice produces hesitation, jitter, and corrections the policy will faithfully learn. Our operators train on each task until their success rate and trajectory quality stabilize before a single production episode is recorded, and the protocol scripts what varies between episodes: object placement, distractors, lighting, initial states.

Recovery demonstrations deserve their own line in the spec. Policies fail in deployment at the edges of the state distribution, so we deliberately stage realistic error states and demonstrate the way out of them. A dataset of only clean successes teaches a policy that has never seen trouble.

QA That Happens Per Episode

Batch-level sampling misses the failure modes that actually poison training runs. Our checks run on every episode: timestamp monotonicity and cross-stream synchronization, dropped frame detection, calibration presence, action-observation alignment, and label consistency against the written success criteria. Episodes that fail are re-collected, not patched.

Throughput Is a Planning Problem, Not a Promise

Episodes per hour depends on task horizon, reset cost, and failure rate, which is why any vendor quoting a universal collection rate is guessing. A short tabletop grasp with a trivial reset accumulates fast; a bimanual garment task with a staged crumple reset accumulates slowly, and the reset often costs more clock time than the demonstration. We plan collections from a measured pilot: the pilot batch establishes the real episodes-per-seat-hour for your task, and the full-collection schedule is built from that number rather than from optimism. Resets get engineered too, with jigs, staged object pools, and two-station alternation so the robot never idles while a scene is rebuilt. This is the unglamorous half of capture as a service, and it is where the 60-day commitment is actually won.

Teleop or UMI-Style Capture?

Choose teleop when the target embodiment is fixed, when force interaction matters, or when the task needs the robot's real constraints in the data. Choose UMI-style handheld capture when you need volume and scene diversity fast and can accept recovered end-effector trajectories instead of on-robot states. Many buyers combine them: a broad handheld corpus for pretraining plus a focused teleop set on the target robot. The spec guide covers how to split a budget between the two.