Josh Gao's company turns people into sensors. Mecka AI, where Gao is founder and chief executive, gives people baseball caps fitted with cameras and gloves that sense force to wear at home and at work. It collects what they record, processes it and licenses the result to labs that build AI for robots. Gao described how that works on the Training Data podcast, in an episode published October 20, 2026. The host was Sonya Huang of Sequoia Capital, which announced an investment in Mecka on October 7.
Huang laid out the problem at the start. Language models learned from huge amounts of free text on the internet, but robotics, she said, is "fundamentally data constrained." No comparable stockpile shows machines how to move through the world and handle objects. Gao called Mecka "the data and deployment company" for physical AI. It creates, processes and licenses much of the data used to train robotics foundation models, the large general-purpose models that labs hope can control many kinds of robots across many tasks. He split robot data into three ingredients: "pixels," "actions" and "logic."
He described the company's job in three parts. Mecka builds hardware that captures people's movements. It runs the operations that collect data from individual contributors and businesses. And it handles the post-processing and "own[s] the quality gates" that turn raw recordings into usable datasets. Mecka's website says much the same thing: recordings become useful training material through processing and evaluation, and piling up video does not do that on its own.
Huang asked whether paying to collect all this data could ever work economically without the internet's head start. Gao said robotics is "not starting from zero," because it can build on vision-language models and world models trained earlier. He also contrasted robotics with language models. For language models, pretraining data (the broad material a model learns from before it is tuned for specific jobs) was mostly gathered by the labs themselves with web crawlers, and it "didn't become a large market." In robotics, he said, pretraining data is "quite a large market."
Why film from the wearer's eyes
Neither founder came from robotics. Gao had built back-of-house order management and other software for cloud kitchens, restaurants that exist only for delivery. In his fourth year of university he skipped class to install kiosks for restaurants in rural Rochester, New York. His co-founder built a stablecoin that was eventually sold to Coinbase. When they entered robotics, around late 2024 by Gao's account, they contacted hundreds of people in the field, read every paper they could find and visited robot labs.
One of those papers was Ego4D, a research dataset of first-person video. Gao remembered it as about 4,000 hours of human demonstration data. Looking through the actual footage, he said, showed how broad it was, down to everyday activities such as bowling. The Ego4D paper roughly confirms his figure: 3,670 hours recorded by 931 camera wearers in 74 locations across nine countries, mostly of unscripted daily activity.
That kind of footage is called egocentric data, Gao explained: video of people doing tasks, filmed from their own point of view. Robots also see the world from their own viewpoint, so the idea was to show them what it looks like to interact with the world in the first person. At the time, he said, most robot data came from teleoperation, where a person remote-controls a robot through a task while its movements are recorded, and that was "very hard to scale." Humanoid robots were starting to take off, and "Humans and humanoids look the same," he said. "Why couldn't we turn humans into this fleet style learning that Tesla has" to help ground robots in the physics of the world?
From a phone app to "Waymo-fying people"
Huang asked whether this meant mounting an iPhone on someone's forehead and letting them go about their day. Gao said Mecka's earliest pilot was a mobile app that got "Chinese aunties and uncles" to record themselves cooking, cleaning and doing other tasks. The company then tried UMI data. The Universal Manipulation Interface is a research design in which people do tasks holding handheld grippers fitted with fisheye cameras, so the recorded motions can later be carried over to robot arms. Eventually Mecka moved to hardware of its own.
Gao said the app scaled quickly and had many virtues, but it did not capture enough "rich information about the world." His answer was what he called "Waymo-fying people." Waymo's driverless cars carry overlapping cameras, lidar and radar; its sixth-generation system has 13 cameras, four lidar units and six radar units. Mecka's version is a baseball cap fitted with cameras and sensors, plus a central hub that connects accessories such as tactile gloves and external cameras. The goal, Gao said, is to capture how people interact with the world day to day and "hopefully be able to clone that behavior."
The gloves are about force. Huang pointed out that self-driving cars don't much need to know about force, but a robot handling things has to know how hard it is squeezing a door handle or an egg, and video doesn't really capture that. Gao agreed. The usual substitute today is a wrist camera that shows when a hand touches an object, which he called "kind of a hacky way" to work out how many newtons it takes to pull open a chair or twist a bottle cap. He expects force data to be "incredibly important." But "there is nothing off the shelf that you can purchase" that a person can wear and that reliably measures everyday force in a form ready for robotics models, he said. So Mecka designed its own gloves, which took three or four product cycles.
Recruiting a map of human work
Huang asked how she would join if she wanted to make an extra buck. Gao described two routes. Through Mecka's consumer apps, anyone anywhere can sign up, register their profession and receive a stream of tasks, such as cleaning a bedroom or cooking breakfast and cleaning up afterward. He called this "a very lightweight approach" to getting variety into pretraining data.
The second route goes beyond the home. Mecka built what Gao called a business-to-business engine to recruit companies from "this incredible taxonomy of all physical human work": bakeries, printed-circuit-board manufacturing lines, "everything to podcasting." He conceded that "there's not too much handwork in podcasts."
To decide where to recruit next, Mecka tracks a heat map across jobs, industries and other measures of diversity. When Huang asked whether that meant knocking on the doors of donut bakers, Gao turned it on her. If the map shows too little podcasting, he said, "that's actually why we're here today. This is our secret message. That's why there's cameras everywhere." Huang shot back that he was "never automating me out of this," then added: "No, I'm kidding. Please do."
The operations behind this are "much more painful than folks can expect," Gao said. Mecka ships hardware to thousands of homes and businesses. In many of the places it operates, it has set up dedicated upload lines, 100-gigabit lines according to the episode description, to handle "hundreds of petabytes" of data. A petabyte is a million gigabytes. All of that footage then has to be processed, checked and run through computer vision before it is useful.
Quality, and a McDonald's surprise
Huang said she had heard that many of the hours of robot data that get sent in end up rejected. Gao traced the lesson back to large "robot factories": warehouses where operators repeat the same tasks to scale up teleoperation data. Research from that work, he said, showed a threshold of diversity, a point where more data stops paying off.
With human data, Gao said, quality is "quite a bit of a moving target." Real-world edge cases are "very extreme," and "you will always see something new." Mecka writes heuristics, simple rules that flag problems as data is processed, and builds systems to escalate issues quickly. "We will learn new things about quality and diversity that weren't true a week ago," he said, so the checking system has to adapt fast as well.
Fraud, in Gao's view, is "an underestimated one across data industries." Mecka is creating data that doesn't exist yet rather than filtering what is already online, and that attracts "very creative people." At one point, he said, people tried to submit YouTube videos in the point-of-view style, showing someone working at McDonald's, as their contributions. "Never underestimate how entrepreneurial people can be," Huang said.
Gao called these "painful lessons." Quality gates are necessary but not enough, he said. Mecka also has to be present where collection happens, see how the work is done, and design an incentive model that rewards good contributions.
Buyers describe the same moving target. Jason Ma, co-founder and chief scientist of Dyna Robotics, which buys outside data, said on Latent Space earlier in October that no single company can collect enough robot data alone. He also called data quality "a non-stationary problem": what counts as good data changes as a model improves.
Where the margin comes from
Huang pushed on the business model. A rival could approach her directly and ask for her podcast data, she said, squeezing the margin Mecka earns as a middleman. Gao said contributor pay is "actually not the largest factor" in making data useful from start to finish. That cost will always be there and may become more competitive, he said, but suppliers can stand out by making sure their data is "verifiably high quality at very large scale," with the right hardware, operations and software.
He described data as a set of products, or SKUs (stock-keeping units), that will change over time. Egocentric podcasting footage is just one SKU. Demand may move from a Chinese restaurant by the side of the road to Michelin-star kitchens, he said, and from an average tennis player to something "closer into what it means to be Federer."
Sequoia's investment post, written by Anas Biad and James Flynn, makes a similar case from the investor's side. They argue that a data supplier has to deliver volume, diversity, quality, speed and competitive cost all at once. What makes Mecka worth backing, in their account, is a platform whose capture methods can change as robotics changes.
Sharing data through EgoVerse
Mecka also gives some of its work away. Gao described EgoVerse as a consortium of people who think scaling up human data matters for robotics, and credited its researchers as among the earliest believers in egocentric data. The project's paper is led by Ryan Punamiya and Simar Kareer, with Danfei Xu among its authors. Gao said Mecka contributed a large amount of egocentric data from businesses, industries and homes. It also shared its action-trajectory data, meaning the processed records of how hands and bodies move, along with the computer vision and quality gates behind it, and the platform and web tooling for loading the data.
He explained the giveaway by where the field stands. Gao said he hoped the effort could "lift up everybody" by sharing tooling, data and lessons, "just because we are so far behind as an industry."
The EgoVerse project page lists a release snapshot of 4,003 hours from 2,087 demonstrators. Academic partners include Georgia Tech, Stanford, UC San Diego and ETH Zurich, and industry participants include Mecka, Scale AI and Meta. The paper's experiments trained robots on human and robot demonstrations together, on three robot platforms and tasks such as placing cups on saucers and bagging groceries. Human data generally helped, though grocery bagging got worse on one of the three robots.