What is teleoperation in robotics?

What the rigs actually look like

There is no single method. A VR headset and hand controllers is the cheapest. An exoskeleton records joint angles from a person's own limbs as they move. A leader-follower rig is a scaled-down replica of the robot arm that an operator moves by hand while the real arm mirrors it — clumsy to look at, but it puts the operator's sense of force in roughly the right place. Motion capture suits sit at the expensive end, and the market for them has grown specifically to serve robotics: Xsens now positions its Link suit at robotics labs facing the data bottleneck.

What they share is the hard part. The operator has to feel enough of what the robot feels to avoid crushing things, through a body that is not theirs — the embodiment gap, experienced live.

Two jobs that get confused

Collecting data. A policy trained by imitation learning needs examples of the task done correctly, and teleoperation is how most of them are produced. This is uncontroversial, and it is most of the teleoperation happening in the industry.

Catching failures in deployment. When a robot working in a real building gets stuck, someone takes over. This is the one people argue about, because a robot that needs rescuing is a robot doing less than the demo implied.

Almost every dispute about teleoperation is really a dispute about the second while citing evidence from the first.

The 1X argument

No company has made the tension more public. The NEO launch split the industry, with a 20,000-dollar home robot whose "Expert Mode" hands control to a remote human. Critics called it selling the dream. 1X's answer is that this was never hidden: the human-in-the-loop design is a decade-old bet rather than a pivot, on the theory that deployed robots generate the data that removes the need for the human.

CEO Bernt Børnich has defended it by comparison to Waymo and ChatGPT — remote assistance as a stage every autonomy business passes through — and argued a supervised robot is more privacy-preserving than a cleaner in your house. Whether you buy that, it is at least an argument in the open. The model has spread: Tau Robotics runs a 30-dollar-an-hour humanoid cleaning service blending autonomy with human teleoperation.

The only number that settles it

Not whether a robot is teleoperated, but how often. The intervention rate — the share of time or tasks a human has to take over — is the honest measure of a deployed system, and it is conspicuously missing from most demo videos.

When a company shows a robot doing something impressive, the questions are: was anyone driving, how often did they take over, and is that rate falling. A company confident in the third will usually tell you the first two.

The bet that it ends

The counter-position is that teleoperation is a scaffold to be removed, and that continuing to depend on it means the data problem was never solved. Sunday Robotics' founders argue exactly that in a conversation about the data bottleneck and the end of teleoperation, and the cheaper alternatives keep arriving: egocentric video, synthetic data, and marketplaces such as Kinetic Blocks, which is trying to make physical-AI datasets a licensable commodity.

Models that need less of it are appearing too. Skild AI's S1 handles unseen ten-minute tasks from a single video prompt, and Generalist AI's GEN-1.5 claims new physical tasks in seconds from one prompt. If that generalizes, the demonstration count per task collapses — and with it the economics that make teleoperation worth doing at scale.