Robot Training Data: Data Farms, Chores and ABB

A robot that folds your laundry has to learn from somewhere, and in 2026 that somewhere is mostly people. Robot training data now comes from factory workers with cameras on their heads and from freelancers in motion capture suits who wash the same plate two hundred times, while other people steer robot arms by remote control. This post follows that data from the first master and slave manipulators to today's data farms in India, China and Europe. It also looks at where ABB's robots and software fit, and what all this recording means for the home robots people may one day buy.

Visitors film an AgiBot X2 humanoid at its stand during MWC 2026 in Barcelona. Photo: JJxFile / Wikimedia Commons (CC BY 4.0)

From hot cell manipulators to recorded robot programs

The idea of a human hand driving a machine hand is older than the industrial robot. In the late 1940s Raymond Goertz and his team at Argonne National Laboratory built mechanical master and slave manipulators so that chemists could handle radioactive material from behind thick shielded windows. The operator moved a handle on the safe side and a pair of tongs copied the motion inside the hot cell. Nothing was recorded: the skill stayed in the operator's arm, and the machine passed each motion through the wall.

A row of master-slave manipulators above the shielded windows of the Alpha-Gamma Hot Cell Facility at Argonne National Laboratory. Photo: Argonne National Laboratory / Wikimedia Commons (attribution)

Recording came with the first industrial robots. The Unimate, which started work at a General Motors plant in New Jersey in 1961, was taught by an operator who stepped the arm through each position and stored the points in memory. The robot then replayed them all day. Later controllers added teach pendants and lead-through teaching, where a person moves the arm by hand. Each program was one demonstration of one task. It worked well for welding a car body, where every part arrives in the same place, but it could not cope with a messy kitchen.

A Unimate robot pours coffee at the Biltmore Hotel, Los Angeles, 1967, replaying motions stored in its memory. Photo: Frank Q. Brown, Los Angeles Times / Wikimedia Commons (CC BY 4.0)

The next step was capturing the human body itself. In the late 1980s VPL Research sold the DataGlove and a full DataSuit that measured finger bends and limb angles for virtual reality. Film studios later used optical motion capture to animate digital characters. Robot labs borrowed both tools. Around 2022 Google reported that its RT-1 model had learned from about 130,000 robot episodes gathered by 13 robots over 17 months. In 2023 the Open X-Embodiment project pooled more than a million episodes from 22 robot types into one shared set.

A VPL DataSuit, the full body sensor suit VPL Research sold for virtual reality. Photo: Dave Pape / Wikimedia Commons (public domain)

Data farms, head cameras and robot trainers in 2026

This year the hunt for data turned into a job market. On 14 October NPR reported from Karur in southern India, where Shoba Kirubakaran packs aprons at a textile factory with a GoPro strapped to her head. The footage is what developers call egocentric data, seen from the worker's own eyes, and it is sold to train robots. She earns about $90 a month and said nobody had told her the video was meant for machines that might one day do her job. Business Insider and The New York Times ran similar stories the same week.

Workers at a textile workshop in India, the kind of workplace where head cameras now record jobs for robot developers. Photo: Fabrics for Freedom / Wikimedia Commons (CC BY 2.0)

In China the work looks different. The South China Morning Post reported on 14 October that freelance robot trainers wear motion capture gear while they fold clothes and wash dishes, sometimes repeating the same movement up to 200 times a day, for between 4,000 and 6,000 yuan a month. In Europe, The Robot Report wrote on 23 October that the Italian humanoid startup Generative Bionics plans three Physical AI Data Factories. In the first, a 5,000 square metre site in Italy due by late 2027, humanoids will perform industrial tasks while the runs are rebuilt in simulation on AMD computing hardware.

A performer in a marker suit is tracked live by an optical motion capture system at SIGGRAPH 2023. Photo: Ma1ted / Wikimedia Commons (CC BY-SA 4.0)

ABB comes at the same problem from the factory side. In October 2026 it showed an AI Robot Trainer add-on for OmniCore EyeMotion at the CIIF fair in Shanghai, a tool that lets non-experts train vision models for picking and handling. ABB says it can cut commissioning by up to 90 percent compared with custom built systems, a figure nobody has tested independently. On the arm side, GoFa CRB 15000 and YuMi cobots already support lead-through teaching, and engineers build and test the rest of a cell in RobotStudio before any real data is recorded.

Lead-through teaching: moving an ABB GoFa arm by hand. Photo: ABB

What all this data means for home robots

For homes, the bet is that enough recorded chores will let one general model handle many kitchens instead of one program per task. Teaching a single robot by demonstration has one person show one arm one job. The data farm approach needs thousands of hours of different hands, rooms and objects. Startups also send in remote operators when a robot gets stuck, and those sessions go back into the training set too, as our post on teleoperated home robots explains.

A Tesla Optimus humanoid at a Tesla event in October 2024, one of the robots home chore data is meant for. Photo: Steve Jurvetson / Wikimedia Commons (CC BY 2.0)

Factories will likely mix the two worlds. A cobot such as GoFa, which senses contact in its joints and can be moved by hand, could serve as a recording station where a worker guides a task a few dozen times and a model learns from it. ABB has not announced a data collection product like this, so for now it is only a possibility. Simulation will carry much of the load, because a digital twin can vary lighting, part positions and mistakes far faster and more cheaply than a person in a motion capture suit.

Using an AR headset to check a GoFa cell before it is built. Photo: ABB

Most of the open problems involve people. Workers in Karur and in Chinese apartments are being paid little to record skills that may later replace them, often without being told what the footage is for. Consent, fair pay and data privacy will matter even more when robots learn inside people's homes. Safety is the other question, since a model trained on video still has to obey force limits and tools such as ABB SafeMove near people. Skills will shift as well, and robot trainer may end up a job title as common as robot programmer.

ABB's GoFa cobot is guided by hand and stops when it senses contact. Photo: ABB

Innovation

AI solutions for effortless ABB robot control.

Automation

Robotics

ceojohntran@chatwithrobot.net

+84905311611

© 2025. All rights reserved.

qtran1215@gmail.com