Inside a Robot Data Factory: How Mecka Sensorizes Humans to Train Robots

Josh Gao describes Mecka as the data and deployment company for physical AI: it creates, processes and licenses data used to train robotics foundation models. Host Sonya Huang points out that, unlike LLMs, robotics is fundamentally data constrained. Gao says the data comes down to three things, pixels, actions and logic, and the clean, curated version of it does not exist at scale.

Asked whether collecting it all is even economically feasible, Gao says robotics still benefits from the internet's legacy, since it can build on VLMs and world models. He notes that pre-training data is already a sizable market in robotics, including data for teaching video models more physically grounded worlds. Whether the volume is reachable depends on which commercial use case each level of data unlocks, starting with simple pick-and-place tasks.

Gao and his co-founder came from outside robotics. He previously built software for cloud kitchens, and his co-founder built a stablecoin that was sold to Coinbase. They chose robotics because they wanted something exciting that would make the future incredible. In late 2024 they read every paper and visited labs, and they noticed that Ego4D offered only about 4,000 hours of human demonstration data, some of it about everyday activities like bowling.

Egocentric data is first-person video of humans doing tasks. Gao says the team bet on it because teleoperation was hard to scale, and humans and humanoids look alike. The idea was to turn people into something like Tesla's fleet learning, with everyone contributing to a model's grasp of physics. The first pilot was a mobile app in which "aunties and uncles" filmed themselves cooking and cleaning.

Huang notes that video does not capture how hard someone squeezes a door handle or lifts an egg. Gao says wrist cameras that show contact between hand and object are a hacky substitute, and that force data will be very important. Mecka is building gloves that record how much force daily tasks take.

On the operations side, there is a direct-to-consumer route in which people sign up through a mobile app and receive a sequence of tasks, and a B2B route that recruits businesses across the taxonomy of physical work, from bakeries to PCB lines. Mecka tracks a heat map of jobs and industries to see where its data is thin. Gao jokes that the heat map is why there are cameras around the podcast studio.

The harder part is everything behind collection. Nothing off the shelf reliably measures the force a person applies, so Mecka runs several hardware product cycles and then distributes the devices to thousands of homes and businesses. It has set up 100-gigabit upload lines in many of the places it operates to handle hundreds of petabytes, and then needs computer vision and strong quality gates to make the data useful.

Quality is a moving target, Gao says, because the real world keeps producing new edge cases, so heuristics and escalation engines must change quickly. Fraud is underestimated: one contributor submitted YouTube footage of people working at McDonald's as his own. The answer is to build the gates, go to where collection happens, and design incentives that reward good contributions.

Huang asks whether a competitor could undercut Mecka by going straight to data owners. Gao replies that the payment to contributors is not the largest factor in making data useful end to end. Differentiation comes from verifiably high-quality data at scale, using the right hardware, operations and software.

He treats each type of data as a product line, or SKU, that will change over time. Today's egocentric footage from an everyday restaurant could give way to data from a Michelin-star kitchen, in the same way that a goal might be to learn what it means to be Federer rather than an average tennis player.

On scale, Gao says the field is now in the millions of hours, a big step up from Ego4D's 4,000 hours two years ago. Tesla, however, spent more than a decade on two degrees of freedom with millions of cars, so robotics is still orders of magnitude from where it needs to be. Huang points out that a robot hand alone has dozens of degrees of freedom, and wonders if there is enough GDP to pay for it.

Gao answers the GDP worry by rejecting the binary view. Robotics will creep into the world as commercial use cases unlock, and evidence of generalization and scaling laws gives people the courage to invest more. With that, companies can also collect in-domain, in-task data to close the last gap, which can mean four nines of reliability in real environments.

Huang notes that home data is the most plentiful, yet the home is probably the hardest place to deploy. Gao agrees that the home demands the most diligence, since benchmarks must ensure, for instance, that a robot does not fall on a baby. He expects early successes in goal-oriented tasks with existing procedures, and says laundry folding, which his team spent over a year and a half on, will be solved soon, with assembly-line pick-and-place within a year or two.

He argues that useful robots can already be deployed because people underestimate how varied robotics is. Service, cleaning and delivery robots already work in hotels in China and Southeast Asia, and industrial cleaning robots may be among the first to spread in North America. The path then steps up through early manipulation tasks toward a humanoid that can do anything out of the box.

Asked when an FSD-style flywheel could appear, Gao says it is a while away, but each solved commercial task starts to generate in-deployment data such as failure cases and corrections. When that happens, Mecka will look very different. It is not committed only to egocentric data and expects many SKUs of the data soup, including edge cases the robots have not met yet.

He also sees Mecka helping with deployment by collecting domain data, post-training models and supporting enterprises. For diffusion he expects both vertical, task-specific companies and horizontal ecosystems, with something like a modern systems integrator. Instead of a tight-margin installer, such a company would install robots that become more capable over time, letting the bot installed last week do something that seemed impossible a few days earlier.

Asked for favorite vertical categories, Gao names industrial robots, because their tasks are constrained and no baby will suddenly appear, and also hotel delivery and handoff robots. He jokes that if Mecka ever helps deploy a robot that mines an asteroid, it will have done a good job, since that is a task humans cannot do and that does not exist yet.

Mecka's hardware started with the mobile app but quickly moved to what Gao calls Waymo-fying people: baseball caps with cameras and sensors, and a central hub that connects accessories such as tactile gloves and external cameras. He also describes EgoVerse, a consortium of researchers who believe in scaling egocentric data. Mecka contributed data, post-processing, quality gates and platform tooling, in the hope of lifting the whole industry.

On the training mix, Gao says that even a year ago people argued that only UMI data or only egocentric data would work, but the field is now converging on a recipe of roughly a million hours of human data, tens of thousands of hours of UMI data, and a few hours of in-domain teleoperation. He does not subscribe to one religion, because tactile data has not entered the mix yet and simulation will too.

To the objection that watching tennis will not let you beat Federer, Gao points to AlphaGo: first behavior cloning from experts, then self-play that produced Move 37. For robotics, the rules of the game are the entire physical world, and no world simulator can recreate that yet. Egocentric data therefore grounds basic physics and helps build such a simulator, and he expects self-play in both the real world and simulation.

He likes omni-model approaches that accept any observation, force, pixels or even smell, and output actions or voice. He thinks AGI labs have the courage to scale imperfect data, and that robot data in LLM pre-training mixes explains why frontier models do well at robotics. But specialized models may reach four nines of reliability on specific tasks.

Closing on his science-fiction roots, Gao describes the Kawasaki Corleo, a robot panther-dog-horse he has hung on the wall of Mecka's New York conference room. He loves how absurd it is and the philosophy of The Hitchhiker's Guide to the Galaxy, and hopes a robot-filled world will include eccentric entrepreneurs building wacky form factors that do jobs no one has imagined.

Asked how many robots there will be in a decade, Gao says many times more than there are people. Huang suggests 100 billion, and he agrees, adding that panther-dog-horses would add at least another order of magnitude. The episode ends with Gao giving Huang a Corleo figurine.