For years, engineers have wrestled with a fundamental challenge: How do you teach robots to function naturally in a world designed for people?
A startup called Mecka AI believes the answer lies in observing humans. Rather than relying primarily on simulations or motion-capture studios, the company collects data from people performing everyday tasks and uses it to help train robots for the physical world.
The approach has attracted significant investor interest. Mecka AI, which operates from Toronto and New York City, recently announced it has raised $60 million to expand what it describes as the “data and deployment layer” for robotics and physical AI.
The funding reflects a broader surge of investment in the data infrastructure needed to power the next generation of robots. In March, San Francisco-based Encord raised $60 million in a Series C funding round to expand its platform for managing and organizing the massive volumes of video, sensor and other data used to train physical AI systems.
While Mecka focuses on collecting demonstrations of human behavior, Encord provides the tools that help AI developers curate, annotate and manage the data used to train robots, autonomous vehicles and other machines that operate in the physical world.
Humans perform countless actions each day that remain difficult for robots, from opening a grocery bag and folding laundry to sorting utensils or picking up objects from a cluttered countertop. While some of those limitations are tied to hardware, Mecka argues that robots can improve dramatically if they are exposed to enough examples of how people naturally complete such tasks.
To gather that information, participants wear lightweight devices such as smart glasses or smartphones mounted on head straps. The system records what they see and how they move while working in settings that include homes, kitchens, chemistry labs, metal fabrication shops and leather-working facilities.
The company contributes to EgoVerse, a dataset containing more than 1,300 hours of demonstrations from more than 2,000 people performing nearly 2,000 different tasks. The dataset is maintained through a consortium that includes Georgia Tech, Stanford University, UC San Diego, ETH Zurich, Meta, Scale AI and individual contributors.
The result is a vast library of human behavior that can be transformed into training data for robots.
Researchers found that robots performed better when trained on larger amounts of human data, particularly when the demonstrations closely matched the tasks the machines were expected to perform. The findings suggest robots can learn important concepts simply by observing how people interact with objects and navigate their surroundings.
“It’s a fundamentally different approach from traditional motion capture,” said Josh Gao, Mecka AI’s co-founder and chief executive officer.
Traditional motion-capture systems often require specialized studios, cameras and sensors to track movement in carefully controlled environments. While effective, they can be costly and may fail to capture the unpredictability of everyday life.
Mecka instead records tasks where they naturally occur. Whether someone is preparing food in a kitchen, folding laundry in a bedroom or organizing tools in a workshop, the resulting data reflects the kinds of situations robots will eventually encounter. Exposure to a wide variety of people, work styles and settings also helps robots adapt to unfamiliar situations.
The strategy mirrors the path taken by today’s most advanced AI systems. Large language models such as ChatGPT learned from enormous amounts of text collected across the internet. Robotics companies now face a similar challenge, finding enough real-world data to teach machines how to interact with the physical world.
Collecting that information directly from robots can be slow and expensive because each demonstration requires hardware, supervision and controlled testing conditions. Human demonstrations, by contrast, can be gathered far more efficiently and at much greater scale.
Examples on the EgoVerse website show how those demonstrations can be translated into robotic actions. Videos feature humanoid robots and robotic arms performing tasks such as bagging groceries and placing objects into bowls after training on human-generated data.
The approach addresses one of robotics’ biggest obstacles. Robots often perform well in laboratories, where conditions are tightly controlled. The real world is far less predictable. Lighting changes, objects move, workspaces become cluttered and people complete the same task in different ways.
As physical AI moves from research labs into real-world deployment, the demand for high-quality training data is growing rapidly. Companies such as Mecka and Encord are addressing different parts of that challenge, one focused on collecting human demonstrations and the other on organizing the vast amounts of data needed to train intelligent machines.
Since launching last year, Mecka says it has worked with leading robotics laboratories and major technology companies and has become one of the world’s largest providers of training data for physical AI systems.
To get more training data, some companies are enlisting temp agencies such as Instawork, which recruits people to wear body or head-mounted cameras, usually their own phones, to record their everyday tasks, including what they do at work.
The boom for such data prompted Instawork to recently launch Instawork Robotics Lab on April 16, 2026, to connect robotics companies with its vast network of hourly workers.
Instawork Robotics Lab has a simple query for potential candidates. Applicants are asked to fill out their name, the company they work for, and contact information. There is another section titled “Robotics Role” in which the applicant has to check one of four boxes, from just wanting to learn more, to “I’m interested in earning money by collecting robotics data or performing robot field operations.”
