Why robots need to learn from electricians
·Julian Enciso·4 min de lectura
The demo and the job site
Robots can now fold laundry, sort parcels and tidy a kitchen in videos that collect millions of views. Put the same robot in a half finished apartment and ask it to pull a cable through conduit, strip it and land it in a junction box, and it has almost nothing to go on.
The hardware is improving fast. What holds robots back on a job site is data: the models that drive them learn from examples of skilled hands doing real work, and for the trades those examples barely exist.
This post is about why electrical work is such a hard and valuable target for physical AI, and why we think the people who already do that work every day should be the ones teaching the machines.
Why trade work is hard for robots
Electrical work is hard for robots because the environment and the hand movements are both complex. In its August 2026 robotics report, Prosus Ventures frames task difficulty as environment complexity multiplied by action complexity, and places construction, repair and maintenance in its long term group.
Start with the environment. No two job sites look alike: an apartment from 1905 and a new office floor differ in wall materials, existing installations, light and clutter. Often another trade has left something in the way.
Then the material. Researchers group cables, ropes and wires as deformable linear objects, which demand high precision and behave in ways that are hard to model. Pulling cable through conduit, dressing it into a box and stripping the end means handling something that changes shape with every touch.
Then the tools. Strippers, crimpers, screwdrivers and voltage testers all need the right force at the right angle, and a skilled electrician judges the result by feel as much as by sight. Much of this happens in tight spots like distribution boards and ceiling voids, with the hands half hidden.
Finally, the cost of a mistake. A loose terminal can start a fire years later, so the reliability bar sits far above that of a robot that sometimes drops a towel.
Why the data barely exists
Robots learn manipulation from examples, and for trade work there are almost none. The Prosus report puts the core problem plainly: robotics has no internet scale library pairing what a robot senses with the action it should take.
The usual sources each fall short for the trades. Teleoperation, where a person steers a robot remotely, gives clean data but scales slowly and tends to carry over the operator's own awkwardness. Internet video is plentiful, but reading precise hand actions from footage is still unreliable, and very few electricians film a full working day with the camera on their hands. Simulation offers endless practice, yet materials like cable are hard to model and the gap to real physics still leaves a long tail of edge cases.
Large first person datasets are finally appearing, but they mostly show factory lines, kitchens and household chores. Build AI's open Egocentric-1M, released in April 2026 with roughly a million hours, comes from factory work. That's valuable for pretraining. It still won't teach a robot how to land a cable in a junction box behind a kitchen cabinet.
Why tradespeople are the best teachers
The best training data for this work comes from the people who already do it all day. A trained electrician carries knowledge that never makes it into a manual: how to feed cable without kinking it, when a crimp is good, how to work around a messy existing installation, which shortcuts are safe and which are not.
Their routine also has a structure a model can learn. In Germany every electrician learns the five safety rules before touching an installation: isolate, secure against reconnection, verify that no voltage is present, earth and short circuit, and cover adjacent live parts. Recordings that show those steps in order, on real sites, are worth far more than a polished demo.
Experts also show recovery. They fumble a screw, reroute a cable, back out a damaged terminal and try again. Robotics data teams increasingly name failure coverage as one of the things today's datasets lack, and real work by skilled people is full of small corrections.
There is an economic side too. In 2025, building electrics had the largest skilled worker gap of any trade occupation in Germany, and for more than seven in ten open positions there was no matching job seeker in the whole country (KOFA, June 2026). Firms that cannot hire will welcome machines that take on parts of the job. The electricians who teach those machines should be paid for what they know.
What we are doing at Klavax
Klavax records real electrical work, with the consent of the people doing it, and turns it into training data for robotics teams. We start with electricians in Europe, on actual job sites rather than staged setups.
Our first rigs use a head mounted stereo camera with an inertial sensor, so every recording carries depth and head motion along with the video. Next we add handheld gripper rigs that capture the action itself, which we explain in our post on head cameras and grippers.
Every partner is paid for each accepted hour of footage, and every recording comes with documented consent and usage rights. We are running our first recordings now and will share what we learn here.
If you build robots that need to work in real buildings, or you run an electrical firm and want to know what recording with us looks like, get in touch.