Training the Next Maintenance Technician: Can Robots Learn Industrial Tasks
Ahmed Rezika, SimpleWays OU
Posted 9/3/2026
We have seen humanoid robots dance, run marathons, and compete in athletic events. Is it real? Is there some trick? Can we really train a robot for this? If so, why don’t we train them for something more helpful to maintenance, not just as a show? This sparked my curiosity and I started thinking “Sure those performances did not emerge from nowhere.” Looking around I found that behind them are real training methods—humans demonstrating, operators controlling, simulations creating millions of practice attempts, and algorithms learning through repeated feedback.
The question for maintenance is: Can we use the same principles to teach a humanoid to perform useful industrial work?
In the last article, we explored the subsystems inside a humanoid through a fictional journey[8]. Yet the technologies behind that journey already exist in today’s advanced robots. We explored what gives a humanoid its physical capabilities, why its components exist, and how different maintenance disciplines come together to keep it operating.
Before that, we examined where humanoids and physical AI could realistically fit on the shop floor[9].
Now, we move one step closer to reality.
Can this new colleague do useful maintenance work and help take maintenance to the next level?
That journey starts with a fundamental question: how do we train a robot to perform a real maintenance task?
Let’s delve into what is already happening—and what it could mean for maintenance.
Learning by Watching: Can a Robot Learn from a Technician?
This is where I would start because it immediately sounds familiar to every maintenance professional.
How did most of us learn our first practical maintenance tasks?
We watched someone.
An experienced technician did the job. We observed where they stood, which tool they selected, how they approached the component, and sometimes the small details that never appeared in any procedure. Then we tried it ourselves.
Teaching robots through human demonstration is not science fiction. It is an active area of robotics research known broadly as learning from demonstration or imitation learning.
And recent developments are becoming increasingly relevant to humanoids.
One particularly interesting example came from Figure in 2025. Its Helix system demonstrated human video-to-robot transfer. The company reported training its model on first-person human video and transferring the learned behavior to a humanoid robot, allowing it to navigate previously unseen, cluttered environments in response to natural-language instructions. Figure described this as Project Go-Big[1]: the robot behavior was transferred without collecting robot demonstrations for that particular environment.
This does not mean that we can simply record a technician replacing a pump bearing and expect a humanoid to perform the same job tomorrow.
There is a difficult problem in between.
The Robot Does Not Have Your Body
A technician and a humanoid robot may both have two arms and two legs, but their embodiments are not identical. They can differ in joint ranges, link lengths, hand geometry, actuator strength and speed, balance dynamics, and contact-sensing capabilities. These differences affect which actions are reachable, stable, and physically executable by the robot. In robot learning, the mismatch between transferable behavior and the target robot’s morphology, sensing, control, and contact conditions is commonly described as the embodiment gap.
A technician may lean over a machine, rest one hand on the structure, and use the other to manipulate a connector. A humanoid may need an entirely different posture to achieve the same objective without losing balance.
So the robot should not necessarily copy the human movement.
It must learn what the human was trying to achieve.
Recent research is directly tackling this problem. The 2025 Human2Sim2Robot work demonstrated a framework in which a robot learned a dexterous manipulation task from only one RGB-D recording of a human performing it. The human demonstration was used to define the task objective, while reinforcement learning in simulation helped the robot discover movements suitable for its own physical body. [2]
That distinction is extremely important for maintenance.
Imagine an experienced technician demonstrating how to connect a diagnostic cable.
The robot does not need to reproduce every movement of the technician’s shoulder, elbow, and wrist.
It needs to understand the objective:
“Locate the correct connector. Approach it safely. Align the plug. Control the insertion force. Confirm successful engagement.” How the robot moves its own body to achieve that goal may be completely different.
Research is moving in that direction rapidly. TThe University of Maryland’s HumanEgo project, announced in June 2026, demonstrated that robots could learn manipulation skills from first-person videos of humans performing tasks. The system required neither robot demonstrations nor robot-specific training data; instead, it represented the spatial relationship between hands and objects to transfer task-relevant behavior across the human–robot embodiment gap. The authors reported an average success rate of 92.5% across four real-world manipulation tasks using approximately 30 minutes of human video per task, with zero-shot transfer to new robot configurations, cameras, and environments. [3]
This is where the idea becomes particularly interesting for maintenance.
How can a Robot Learn From a Technician?
Imagine an experienced technician wearing a camera while performing a well-defined inspection or service task. The recording could capture not only the official procedure but also the physical context: where the technician looks, which component receives attention first, how the tool approaches the equipment, and how the sequence unfolds.
Of course, video alone is not yet enough for most complex maintenance work. The robot may also need depth information, force data, successful robot demonstrations, simulation, or teleoperated examples – as we are going to see next -. But the principle is already real: human experience can become part of the training data used to teach robots physical skills.
And this brings us to an interesting possibility.
For decades, maintenance departments have worried about losing knowledge when experienced technicians retire or leave.
Humanoid robotics introduces a new question:
Could part of that practical knowledge eventually be captured not only in documents and training sessions, but also in demonstrations from which a machine learns how to act?
That does not mean replacing the experienced technician.
At least not in the way the headlines often suggest.
It may mean that the technician becomes something new: the person who teaches the new maintenance colleague how the job is really done.
The next question is even more practical.
What if watching is not enough? What if the robot needs an experienced technician to perform the task through its own body?
That is where teleoperation comes in.
Learning by Doing: When the Technician Controls the Robot
Watching a technician can provide valuable information. But some maintenance tasks involve details that are difficult to capture from observation alone.
How much force was needed?
How did the technician react when the component resisted movement?
How did they adjust the tool when the first approach failed?
In these cases, another training method is becoming increasingly important: teleoperation.
Instead of watching the robot work, the technician temporarily becomes its brain.
Teleoperation uses interfaces such as VR headsets, motion-capture systems, cameras, or specialised controllers to allow a human operator to control a robot’s movements. While the robot performs the physical task, its sensors and control system can record the resulting observations, actions, joint states, hand trajectories, object interactions, and task outcomes. The important point is that the operation does not necessarily end when the task is completed: the recorded interaction can become training data for learning from demonstration.
This is not merely a future concept. In 2024, at the 8th Conference on Robot Learning (CoRL 2024), Cheng et al. demonstrated Open-TeleVision, an immersive teleoperation system in which an operator’s arm and hand movements were mirrored by humanoid robots. The paper was subsequently published in the Proceedings of Machine Learning Research in 2025 [4].The authors used demonstrations collected with the system to train imitation-learning policies for longer-horizon, precise tasks, including can sorting, can insertion, folding, and unloading, and deployed the resulting policies on real robots.
In same conference Tairan He et al. presented OmniH2O, which extends this idea to whole-body humanoid control. The system allows a human to teleoperate a full-sized humanoid using interfaces including VR, verbal instruction, and RGB cameras. The researchers used teleoperated demonstrations for skill learning and developed a sim-to-real pipeline to move from direct human control toward autonomous execution of whole-body tasks.
The Teleoperation idea is already moving beyond research laboratories.
In Beijing, large humanoid training facilities are being built specifically to collect this kind of physical training data. At one facility, operators guide robots through repetitive tasks such as picking up a key, aligning it with a lock, inserting it, and turning it. The motion data is recorded, processed, and used to train improved AI models before those models are sent back to the robots. [5]
China’s largest humanoid training centre, in Beijing’s Shijingshan district, has also replicated industrial and other real-world environments at full scale. Its robots are trained using VR and motion-capture equipment for tasks including sorting, packaging, inspection, carrying, and delivery. The centre reported that robots had learned more than 20 basic or “atomic” skills. [6]
This gives us a much more practical picture of how a maintenance department could eventually train a humanoid.
Imagine teaching it to collect vibration data from a pump.
At the beginning, the technician may teleoperate the robot:
- Walk to the pump.
- Identify the correct measurement point.
- Take the instrument.
- Position it correctly.
- Maintain contact for the required measurement.
- Move to the next point.
Every successful demonstration becomes an example but equally important are the corrections.
Perhaps the robot approaches from the wrong angle, The instrument does not make proper contact, A cable obstructs the path. Then, The technician adjusts the movement.
That adjustment is valuable data because it shows the system not only what normally works, but also how an experienced operator reacts when reality differs from expectation.
This could create a gradual path toward robot autonomy:
Step 1: The technician controls the robot.
The humanoid performs the task, but the human makes every important decision.
Step 2: The robot repeats demonstrated parts.
Simple actions become partially autonomous while the technician remains in control.
Step 3: The robot performs the routine task under supervision.
The technician intervenes when something unexpected happens.
Step 4: The robot earns greater autonomy.
Only after repeated, reliable performance should more responsibility be transferred to the machine.
This progression may be particularly suitable for maintenance.
We already use a similar process when training a new technician.
- Watch me.
- Do it with me.
- Do it while I watch.
- Now do it yourself.
The difference is that, with a humanoid, every demonstration can potentially become reusable training data. That leads to an interesting possibility for the future maintenance team. An experienced technician may not only transfer knowledge to younger colleagues. They may also spend part of their time teaching a robot through its own body. And that raises the next question: If a humanoid can practice a task repeatedly under human control, could we let it practice thousands or even millions of variations before it ever touches a real production asset?
That takes us from the shop floor into a world that does not physically exist at all:
Learning in Simulation—and Facing Reality Later
The third method is where the contrast with maintenance becomes especially important: a robot can practice millions of times in a digital world, but eventually it must deal with the oil, wear, tolerances, obstructions, and surprises of the real one.
Learning in a World That Does Not Exist
There is one obvious problem with teaching a humanoid through repeated practice.
You cannot allow it to learn every lesson on a real production asset.
Imagine training a new humanoid to open a machine door, replace a component, or manipulate a diagnostic connector. During the learning process, it may approach from the wrong angle, apply too much force, lose its balance, or collide with the equipment.
That may be acceptable in a laboratory.
It is not acceptable beside a running production line.
This is why simulation has become a major part of humanoid training.
Before the robot performs the task in the real world, it can practice inside a digital one.
The idea is not simply to animate a robot on a screen. Modern simulation attempts to model the robot’s body, joints, contact forces, gravity, objects, cameras, and sometimes the visual appearance of the environment. The robot can then attempt the same task again and again without damaging itself or anything around it.
And this is already happening at a serious scale.
NVIDIA, for example, has used synthetic data generation for humanoid training. Its Isaac GR00T data-generation approach produced more than 750,000 synthetic trajectories in 11 hours—equivalent, according to NVIDIA, to about 6,500 hours of human demonstration data. The synthetic data was then combined with real robot data, with NVIDIA reporting a 40% improvement in performance compared with training on the real data alone. [7]
For a maintenance task, this creates an intriguing possibility.
Imagine training a humanoid robot to inspect an industrial pump. Inside a simulation, the robot could practise the task again and again:
- walking toward the equipment;
- identifying the correct pump;
- approaching it from different directions;
- avoiding obstacles;
- locating inspection points;
- reaching for them with an instrument; and
- maintaining its balance while manipulating the tool.
Once the basic behaviour is learned, the simulated environment can be changed in countless ways. The pump can be moved slightly. The lighting can be altered. An object can be placed on the floor. The appearance of the equipment can be changed. The inspection point can be repositioned. Sensor noise can be introduced.
Through this process, the robot can experience thousands of variations—conditions that would be impractical, expensive, or even dangerous to recreate repeatedly on a real shop floor.
However, this leads to one of the central challenges in robotics: the sim-to-real gap.
This will be the core of our next talk
The Thought Experiment
Where Does This Leave Us?
We now have four complementary ways of teaching a humanoid:
Watch — learn from human demonstrations.
Control — learn while humans operate the robot.
Simulate — practice thousands of variations without risking real equipment. – Detailed Next Lecture
Learn — refine behavior through feedback, success and failure. – Included in next Lecture
None of these methods alone is likely to produce a dependable maintenance colleague.
Together, however, they suggest a realistic training pathway.
The technician demonstrates.
The robot observes.
The operator teaches.
The simulator provides endless practice.
The learning system improves.
The human supervises.
And gradually, the robot earns autonomy.
This brings us to the question that matters most to a maintenance manager:
What should we teach it?
Not every maintenance task requires a humanoid robot. Some tasks are better suited to conventional automation, specialized machines, or human technicians. The challenge is therefore not simply to develop better training technologies. It is to identify the maintenance tasks where a humanoid can provide real value.
That is the focus of the next lecture: moving from how humanoids learn to what they should learn first—and why.
Must-Know Jargon
Learning from Demonstration (LfD): A robot learns a task by observing how a human performs it. The goal is to capture useful behaviour and reproduce the task, rather than manually programming every movement.
Imitation Learning: A form of machine learning in which the robot learns behaviour by copying examples. The challenge is not simply copying human motion, but adapting the demonstrated task to the robot’s own body.
Embodiment Gap: The difference between the human body and the robot’s physical structure. A humanoid may have to achieve the same objective using different joint movements, reach, strength, or balance strategies.
Motion Retargeting: The process of translating human movements into movements that the robot can physically perform. A technician’s arm trajectory, for example, may need to be transformed to fit the robot’s different joints and dimensions.
Teleoperation: A human remotely controls the robot in real time using controllers, motion capture, or virtual reality. The robot performs the physical task while the human provides the decision-making and movement commands.
Demonstration Data: The information collected while a human demonstrates or teleoperates a task. It may include video, robot movements, joint positions, forces, sensor readings, and the outcome of the task.
References
1- Figure Company, Project Go-Big: Internet-Scale Humanoid Pretraining and Direct Human-to-Robot Transfer, September 18, 2025, https://www.figure.ai/news/project-go-big
2- Stanford University, Conference on Robot Learning (CoRL) 2025, Crossing the Human-Robot Embodiment Gap with Sim-to-Real RL using One Human Demonstration,
https://human2sim2robot.github.io
3- University of Maryland Institute for Advanced Computer Studies. “UMD Researchers Enable Robots to Learn from Human Experience.” , June 3, 2026: https://www.umiacs.umd.edu/news-events/news/umd-researchers-enable-robots-learn-human-experience
4- Cheng, X., Li, J., Yang, S., Yang, G., & Wang, X. (2025). Open-TeleVision: Teleoperation with immersive active visual feedback. Proceedings of the 8th Conference on Robot Learning(2024), 270, 2729–2749, https://proceedings.mlr.press/v270/cheng25b.html
5- People’s Daily Online, China’s humanoid robot training centers multiply as sector gains momentum, April 28, 2026, https://en.people.cn/n3/2026/0428/c90000-20451334.html
6- Over 10,000 Square Meters: China’s Largest Humanoid Robot Training Center Opens in Beijing, Source:China Government, 2025-10-09, https://english.beijing.gov.cn/latest/news/202510/t20251009_4216190.html
7- NIVIDIA Developer: Accelerate Generalist Humanoid Robot Development with NVIDIA Isaac GR00T N1, March 18, 2025, https://developer.nvidia.com/blog/accelerate-generalist-humanoid-robot-development-with-nvidia-isaac-gr00t-n1/#gr00t_n1_data_strategy_for_pretraining%C2%A0
8- MaintenanceWorld.com, Ahmed Rezika, An Engineer’s Guide to Robots, August 5th, 2026, https://maintenanceworld.com/2026/08/05/an-engineers-guide-to-robots/
9- MaintenanceWorld.com, Ahmed Rezika, Physical AI on the Shop Floor: Translating Emerging Tech into Maintenance Tasks, July 9th, 2026, https://maintenanceworld.com/2026/07/09/physical-ai-on-the-shop-floor-translating-emerging-tech-into-maintenance-tasks/
Related Articles
