The first time I paid attention to the debate over collecting data for artificial intelligence, it was through a video of a barber at work. He appeared to be wearing camera equipment as he cut hair, turning an ordinary workday into material that could help train an AI system.

The comments were split. Some argued that machines need examples of how people perform real tasks, and that a skilled barber has something to teach them. Others were less convinced. Why help train a technology that might one day make your own skills less valuable?

Beneath the disagreement was a more difficult question: can workers benefit from teaching machines their skills without making their own livelihoods more vulnerable?

To understand what was at stake, I looked into Verlet Robotics, a Kigali-based startup collecting human movement and robot-control data for developers building robots that can perform tasks in the physical world.

The company grew out of a problem its founders encountered while working in Yale University's Apollo Lab: researchers had ideas for training robots but lacked enough data to test them at scale.

I wanted to know who was doing the work in Kigali, what they stood to gain and how the company approached the prospect of training machines that could one day compete with human workers.

"We think that this century will not be defined by just chatbots,” said Charlie Humphreys, the company's founder and chief operating officer. "We think that the biggest influence in the 21st century is going to be robots that can interact with the real world.”

ALSO READ: National AI Agency gets new leaders

While large language models can learn from collections of books, websites and other digital materials, robots require a different kind of training data.

They need to understand how objects move, how materials respond to pressure and how physical interactions unfold in real environments.

Verlet was established to help address that gap by collecting human movement data and information generated when robots are operated by people. The company evaluated several potential locations before choosing Kigali as its operational base.

According to Humphreys, the decision was influenced by operating costs, security, the availability of technical talent and Rwanda's business environment.

Rwanda offered an opportunity to develop and test robotic systems before deploying them in larger markets.

How human movements help machines learn

At the centre of Verlet's activities is the collection of data showing how people use their hands to interact with objects.

The company works with businesses and cooperatives, including coffee producers in Rwanda's Northern and Western provinces, to record people performing practical tasks. The information is used to help train vision-language models (VLMs), which connect visual information with language-based understanding.

Humphreys explained that the process is not simply about teaching robots to copy human actions.

For example, recording someone weaving a basket does not automatically teach a robot to make an identical basket. Instead, the information can help an AI model understand how materials bend, move and behave when handled.

A separate process uses robot-control data to train vision-language-action models (VLAs), which connect what a robot sees and understands with the physical actions it needs to perform.

At Verlet's Kigali office, operators control robotic arms while recording their movements. The resulting data is reviewed, labelled and prepared for robotics developers, primarily in the United States and, to a lesser extent, Europe.

Building a local robotics ecosystem

Stefan Marxemeier, the company's managing director for robotic operations, said approximately 400 people are involved in field data collection, while another 100 work from the Kigali office across teleoperation, quality control, software, hardware and administrative functions.

This means hundreds of people in Rwanda are participating in work that supports the development of robotic systems intended for international markets.

Although Verlet's current operations are largely focused on generating training data, its ambitions go beyond supplying information to robotics developers.

Humphreys said the company has plans to begin deploying robots in the United States in 2027.

Its initial focus will be on identifying tasks that robots could perform at a commercially viable cost.

Potential applications include serving coffee and scooping ice cream, activities that require machines to interact with objects and operate reliably in real-world environments.

The company is also considering opportunities to deploy robots with local partner businesses for testing and data collection.

ALSO READ: Africa must build AI, unlock financing and scale health innovation

Opportunities for local communities

Beyond its office operations, Verlet is exploring ways to expand data collection into agricultural, industrial and craft activities across Rwanda.

Humphreys said the company is interested in sectors such as tea processing, where repetitive manual tasks could provide useful information for robotics development.

It has also engaged women at Nyamirambo Women's Center, whose practical skills can contribute to the collection of training data.

According to Humphreys, some participants can roughly double their daily earnings through the company's data collection arrangements.

He said the additional income can help families meet expenses such as food and school supplies.

The company is also exploring partnerships beyond Rwanda, including the possibility of working with skilled craftspeople in Japan.

Humphreys said discussions have included carpenters and specialists involved in maintaining historic temples, whose techniques could provide valuable examples of complex manual work.

Such activities illustrate how skills acquired through years of practical experience could become valuable resources for AI development.

Can Kigali become a hub for physical AI?

Verlet's presence in Rwanda comes as the country seeks to strengthen its position in emerging technologies, including artificial intelligence.

Humphreys pointed to Rwanda's policy environment as an important factor in the company's decision to operate in Kigali, citing efforts to establish frameworks governing AI and data use.

The government has taken steps to strengthen national AI governance, including the establishment of a national AI agency.

The long-term question is how much of the technology can be developed and deployed locally. Verlet has ambitions to build more of its own hardware and expand its operations in Rwanda, although much of its robot equipment is currently imported.

For now, the focus is on producing the data that developers need to move robots from research settings into practical applications.

What does this mean for the future of work?

The rise of physical AI brings a difficult question into focus: what happens when machines become capable of performing the same tasks as the people whose skills helped train them?

For participating workers, data collection can provide an additional source of income and exposure to emerging technology. Humphreys said he expects the business to grow in Rwanda, but argued that large-scale robot deployment to replace human workers across Africa remains a long way off.