When a humanoid robot folds a shirt in a demonstration video, the natural assumption is that it's operating autonomously — the robot perceiving the world, making decisions, moving its hands. That assumption is often wrong. A significant portion of what gets shown in humanoid robotics videos involves teleoperation: a human operator wearing motion-capture equipment or using a control interface, guiding the robot's movements in real time, sometimes from across the room and sometimes from across the country.

This isn't deception, exactly. But it's routinely underexplained. Understanding what teleoperation is, why robotics companies rely on it, and what role it plays in the path to genuine autonomy is one of the more useful lenses for reading the humanoid industry honestly right now.

What Teleoperation Actually Means

Teleoperation — sometimes shortened to teleop — means a human operator controlling a robot remotely. The word covers a wide range of setups. At one end: a person wearing a full-body exoskeleton suit whose movements are mirrored precisely by the robot in real time, with haptic (touch-feedback) gloves letting the operator feel what the robot's hands are touching. At the other end: a person with a joystick or keyboard, directing coarse movement from a control panel.

Modern humanoid teleoperation systems used in research and commercial development tend toward the more sophisticated end. Companies including Figure AI, Physical Intelligence, 1X Technologies, and others use whole-body teleoperation rigs where a human demonstrator's head, torso, arm, and hand movements are captured and translated into corresponding robot motions. The human isn't just pressing buttons — they're physically performing the task, and the robot's body executes what the human body does.

The latency involved matters enormously. For tasks within the same building, latency can be held below 50 milliseconds, which feels nearly real-time to an operator. Remote teleoperation over the internet introduces more delay, and any latency above roughly 150 to 200 milliseconds starts to degrade the quality of fine manipulation — the operator's movements and the robot's responses fall out of sync in ways that make delicate tasks difficult.

Why Teleoperation Exists in the Development Pipeline

The reason teleoperation sits at the centre of so much humanoid development right now isn't primarily about demonstrations. It's about data.

Training a robot to perform a task autonomously requires large quantities of example data: recordings of the task being completed correctly, across enough variations in object position, lighting, orientation, and context that the robot's learning system can generalise. Getting that data is hard. You can simulate it in software — and most companies do — but simulation has gaps. Physics engines don't perfectly replicate real-world contact, friction, and material deformation. A robot trained entirely in simulation often struggles when it encounters the real world, a problem researchers call the sim-to-real gap.

Teleoperation provides a way to generate real-world demonstration data efficiently. A skilled human operator can complete a task correctly, repeatedly, while the robot records every sensor reading — joint positions, camera feeds, force measurements — alongside the operator's control inputs. Those recordings become training examples. The robot learns, essentially, by watching recordings of itself being guided by a human expert.

This approach is sometimes called learning from demonstration, or imitation learning. It doesn't produce instant autonomy — the robot still needs to generalise from demonstrations to new situations it hasn't seen before — but it provides a much richer signal than programming explicit rules, and it can be scaled by hiring human operators to collect data across a wider range of tasks and environments.

The Scale of the Data Problem

How much demonstration data does a robot need to learn a task reliably? The honest answer is that no one is certain, and the numbers vary significantly depending on the task's complexity, the robot's architecture, and the learning approach being used.

Some research groups have reported achieving reasonable performance on specific manipulation tasks with a few hundred teleoperated demonstrations. More complex tasks, or tasks requiring robust generalisation across varied conditions, may require thousands. Physical Intelligence, a San Francisco-based robotics AI company, has described training models on datasets drawn from multiple robot embodiments across diverse tasks — implying an effort to build generalised manipulation capability rather than single-task competence, which demands far more data.

The data collection bottleneck is one reason companies are investing in making teleoperation rigs faster to set up and easier to use. If a trained human operator can collect twenty high-quality demonstrations per hour, and you need ten thousand demonstrations for a robust policy, that's five hundred operator-hours of work — before you account for data quality filtering, system downtime, and the cost of the operators themselves. The economics of data collection are not trivial, and they don't show up in the press releases announcing new capabilities.

Reading Demonstrations More Carefully

Teleoperation's presence in the development pipeline has a direct implication for how to read humanoid robot demonstrations.

When a company releases a video of their robot completing a task, the relevant questions are: Is this autonomous, or is an operator involved? If autonomous, how was the task scope constrained — what aspects of the environment were controlled to make success more likely? Has this been tested across varied conditions, or is this a best-take recording of a specific setup?

Most companies are not transparent about these questions in their public communications, which is why they're worth asking. A robot that completes a task autonomously in a highly controlled environment is genuinely interesting. A robot completing the same task with real-time human guidance is also genuinely interesting — as a data-collection exercise — but it's a different kind of interesting, and conflating the two produces a distorted picture of where autonomy actually is.

There's also a middle ground worth knowing about: shadow mode operation, where a robot runs autonomously but a human operator monitors and can intervene. This is closer to true autonomy than teleoperation, but the intervention rate — how often the human needs to step in — is the number that matters, and it's rarely disclosed.

Teleoperation as Its Own Application

Separate from its role in training data collection, teleoperation has a near-term commercial application that often gets overlooked in the focus on full autonomy.

Some tasks are genuinely better suited to remote human operation than to autonomous robots, at least for now. Hazardous environments — nuclear facilities, disaster sites, certain manufacturing processes — where placing a human physically is dangerous or impractical. Precision tasks in remote locations. Situations where real-time human judgment is essential but physical presence isn't feasible.

1X Technologies, a Norwegian robotics company backed in part by OpenAI, has been explicit about a hybrid model: their humanoid robots are designed to operate under human teleoperation as a near-term revenue-generating service while autonomous capability develops alongside. The robot isn't waiting to become fully autonomous before it earns its keep — it earns its keep as a capable remote body for a human operator, and autonomy is layered in incrementally as the training data accumulates.

This framing is more commercially grounded than the "full autonomy by X date" claims that dominate coverage, and it's worth taking seriously as a model for how the transition from teleoperated to autonomous operation may actually unfold in practice.

The Operator Behind the Robot

There's a human dimension to the teleoperation story that's almost entirely absent from industry coverage: the people doing the operating.

Skilled teleoperation of a humanoid robot is not trivial. It requires spatial reasoning, fine motor coordination, and the ability to interpret sensor feedback from an unfamiliar embodiment — to feel, through haptic feedback or visual cues, what a robot hand is touching and how much force to apply. Training a competent teleoperator takes time, and collecting high-quality demonstration data requires someone who can complete tasks correctly and consistently across extended sessions.

As teleoperation data collection scales up, the workforce doing this work grows. It is, in a structural sense, a new category of skilled labour — one that exists precisely because the robots it trains aren't yet capable of doing the work themselves. What happens to that workforce as autonomous capability improves is an open question that the industry hasn't addressed, and probably won't until it becomes urgent.

For now, teleoperation occupies an unusual position: essential infrastructure for building autonomous robots, often invisible in public communications about those robots, and a genuine near-term commercial application in its own right. Keeping those three things distinct is useful for anyone trying to understand what the humanoid industry is actually building, rather than what its announcements suggest.