Skip to Content

Teaching robots by demonstration: the job that is coming

Teleoperation, learning from demonstration and imitation learning: humans are becoming robot trainers.
November 15, 2026 by
Teaching robots by demonstration: the job that is coming

For a long time, robots were programmed line by line. In 2026, the most advanced ones learn differently: a human shows them the movement, guiding them remotely or doing it themselves in front of a camera, and an AI model learns to imitate it. The short answer: humans are not disappearing from robotics, they are changing roles and becoming the robot's trainer. It is already a job, and it will be a key skill in workshops.

Key takeaways

  • Teaching robots by demonstration has taken hold in research and among humanoid makers: teleoperation, hand-held grippers, motion-capture suits, video.
  • Researchers showed in 2023 that a robot could succeed at fine tasks 80 to 90% of the time with only ten minutes of demonstrations.
  • Major players are already hiring people to produce these movements: operators in motion-capture suits, teleoperators, data collectors.
  • The quality of the demonstrations makes the quality of the robot. The know-how of your best operators becomes a strategic resource.

What is happening

From code to gesture. In April 2023, researchers published ALOHA, a low-cost two-arm system driven by a human through two "leader" arms. The robot records the movements, then a neural network learns to reproduce them. With ten minutes of demonstrations, it opens a small condiment cup or slots in a battery with an 80 to 90% success rate. In early 2024, the mobile version, designed at Stanford, learned complete household tasks with 20 to 50 demonstrations per task, on a hardware budget under 32,000 dollars, comparable according to its authors to the price of a single industrial cobot. The same year, the UMI interface showed that these demonstrations can be collected with no robot at all, using a hand-held gripper.

Demonstrations on a global scale. Researchers are pooling their data. The Open X-Embodiment project, published in 2023 by 21 institutions, brings together more than a million trajectories from 22 types of robots. The DROID dataset, published in 2024, contains 76,000 demonstrations, or 350 hours, collected by 50 people in 564 different scenes. These data are used to train vision-language-action (VLA) models, such as π0 from the company Physical Intelligence, Google DeepMind's Gemini Robotics or NVIDIA's GR00T N1, presented in March 2025 as the first open foundation model for humanoids.

The corrector, after the demonstrator. In November 2025, Physical Intelligence described a method that combines three sources: demonstrations, the robot's own experience and the corrections of an expert who takes over remotely when the robot gets it wrong. The robot folds laundry, assembles boxes and makes coffee; on some of the hardest tasks, throughput more than doubles and the failure rate is roughly halved. The human is no longer only the one who shows: they are the one who corrects.

A new job is emerging. In 2024, Tesla was hiring "data collection operators" paid 25.25 to 48 dollars an hour, equipped with a motion-capture suit and a virtual reality headset, and able to walk more than seven hours a day. In 2025, Figure announced that its model had learned to move around homes solely from videos filmed at eye level by people, with no robot demonstrations at all. In China, according to the Hong Kong daily The Standard, at least 30 data collection centres existed by the end of 2025; in one of them, each robot has its own dedicated trainer, who guides its movements by teleoperation while the system records joint angles and grip force. The director of that centre estimates that learning a single basic skill takes thousands of recordings. At home, finally, 1X's NEO robot can be taken over remotely by an operator, with the owner's approval, for tasks it has not yet mastered.

Humans as robot trainers: what it means for you

For a workshop, the consequence is profound. Programming a conventional robot required a specialist, days of tuning and a part always placed in the same spot. Learning by demonstration opens the door to more variable tasks, and above all to the people who know how to do the movement: your operators.

Know-how therefore changes status. A setter's knack, the way an assembler picks up a fragile part, a welder's instinctive correction: these are exactly the data a robot needs. Demonstrating well is a skill. A hesitant or inconsistent demonstration produces a hesitant robot. The best robot trainers will often be the best professionals in the trade, not the IT people.

Two points to watch. First, these data have value: who owns them when a supplier collects them in your workshop? Second, the people: filming and recording an employee's movements touches on their privacy and on the recognition of their know-how. This is negotiated beforehand, not afterwards.

What I have learned

I have not yet trained a humanoid robot. But I know this logic in other forms. When I set up the Alquimista Café, I trained myself as a barista so that I could train the team, then wrote a manual for each role. Passing on a movement means breaking it down, making it consistent and explaining the exceptions. That is exactly what is asked of a demonstrator.

I also live it with AI. For my bookkeeping handed to an agent, I no longer enter the documents: I write the rules, I review, I correct the doubtful cases, and the agent improves. Demonstrate, supervise, correct: tomorrow's robot will require the same work, with movements instead of rules. And in my workshops, I see every day that a precise machine does not replace the eye of someone who knows what a well-made part is.

What to do in practice

1. Identify your reference movements

Find the tasks where one person does clearly better than the others, and ask yourself why. That know-how, implicit today, is your future raw material.

2. Document them now

Film, with the people's consent, and describe the steps, tolerances and special cases. Even without a robot, this documentation helps train newcomers.

3. Run a small-scale trial

A small open-source arm can be driven, can record and can learn for a few hundred dollars. An afternoon is enough to understand what a good demonstration is. I go through these options in robotics within reach of a small business.

4. Set the rules for the data

Before any project with a supplier, write down who owns the recordings, where they are stored and what the supplier may do with them. Involve the people being filmed in that decision.

Under the hood

For your technician, the typical imitation learning pipeline:

  • Teleoperate, record, train, deploy: this is the cycle offered by Hugging Face's open-source LeRobot library, with a leader arm moved by hand and a follower arm that reproduces the movement.
  • Each demonstration becomes a dataset: camera video synchronised with joint positions and commands.
  • To start, a lightweight policy such as ACT (the one from ALOHA); then fine-tuning a pretrained VLA model, such as π0 or SmolVLA, on your own demonstrations.
  • Quality matters more than quantity: stable lighting, fixed cameras, consistent movements from one demonstration to the next.

Frequently asked questions

Do you need to know how to code to train a robot?

To demonstrate a movement, no: above all you need to know the movement well. To set up the learning pipeline, a technician is still needed.

How many demonstrations does it take?

It depends on the task. Published work ranges from a few dozen demonstrations for a specific task to thousands for a general skill.

Are robot trainers making themselves redundant?

Not in the short term: the most recent methods rely precisely on an expert's corrections. Their role shifts from doing the work to demonstrating, supervising and quality control.

Going further

This article is part of a four-part series: where humanoid robots and cobots really stand, what robots automate and what stays human, and where a small business should start. Preparing your teams for this new role is as much a matter of organisation as of technology: see Strategy & advisory, or Engineering & fabrication for a prototype. The simplest way to discuss it: a 20-minute call.

Sources

Tags: teaching robots, learning from demonstration, imitation learning, teleoperation, robotics, know-how, new jobs

Humanoid robots and cobots in 2026: where do we stand?
Humanoids are in the pilot phase, cobots are ready: a dated, sourced state of the art, without the marketing.