Inside the Robot Data Factory: Mecka's Josh Gao on Sensorizing Humans to Train Robots

Gao & Huang (Sequoia Training Data) — inside the robot data factory

▶ Watch on YouTube · Sequoia Capital · 45:12

Robotics never got the internet’s gift of free training data, so someone has to manufacture it. Josh Gao, founder and CEO of Mecka AI, runs one of the largest robot data factories, and he explains what that actually involves: custom baseball caps and tactile gloves because nothing off the shelf captures force, 100-gigabit upload lines to move hundreds of petabytes, and custom quality gates. Josh lays out the training recipe the industry is converging on (roughly a million hours of egocentric human data, tens of thousands of hours of robot data, and a few hours of in-domain teleop) and why tactile and simulation data still have to enter the “data pot of soup.”

TL;DR

A 45-minute episode of Training Data, Sequoia Capital’s podcast, published 20 October 2026. Sonya Huang interviews Josh Gao, founder and CEO of Mecka, which calls itself “the data and deployment company for physical AI”: it creates, processes and licenses data for robotics foundation models. It is the wiki’s first source on where robot training data comes from. Five claims:

1. Robotics had no internet to train on. Robot data is “pixels… actions, and… logic,” and there was no free corpus of it. Unlike LLM pre-training data, robotics pre-training data is becoming “quite a large market.”
2. Sensorize humans. Mecka’s bet is egocentric data: first-person video of people doing tasks, scaled like Tesla’s fleet. It started with a phone app recording “aunties and uncles” cooking and cleaning; it now makes camera caps, a sensor hub and tactile gloves, because “there is nothing off-the-shelf” that measures the force people apply.
3. It is an operation, not a website. Contributors come through a consumer app and through recruited businesses across “this incredible taxonomy of all physical human work.” Mecka installed 100-gigabit upload lines to move “hundreds of petabytes,” and its quality gates catch fraud, such as people submitting YouTube “POV working at McDonald’s” videos as their own.
4. The training mix has converged, for now: roughly “a million hours of human data,” “tens of thousands of hours” of UMI data from handheld demonstration grippers, and “just a couple hours” of in-domain teleoperation, with tactile and simulation data still to enter “the data pot of soup.”
5. No single breakthrough. The field is in “millions of hours” of data, against Tesla’s decade on “two degrees of freedom”; a robot hand has dozens. Progress will come as “a slow creep up” of commercial tasks, constrained industrial work first and the home last.

Key claims

Why egocentric data

Gao and his co-founder came from payments (a restaurant point-of-sale system; a stablecoin sold to Coinbase) and entered robotics in late 2024 with “a fresh lens.” The largest academic first-person dataset then, Ego4D, held about 4,000 hours, much of it everyday life such as bowling. Teleoperation, people driving robots to record demonstrations, was “very hard to scale.” Their intuition: “Humans and humanoids look the same. Why couldn’t we turn humans into this fleet-style learning that Tesla has?” Video alone misses force (“how much force in Newtons it takes to pull open a chair or twist a bottle cap”), which is why Mecka builds gloves.

Running a data factory

A heat map of jobs and industries tells Mecka where its data is thin, and it recruits contributors to fill the gaps. Quality is “a moving target”: the real world keeps producing edge cases, so it needs heuristics that catch problems in processing and the “nimbleness” to change them weekly. Fraud is “an underestimated one across data industries”, because Mecka creates data that does not yet exist rather than filtering what does; “you also have to be there where collection is happening” and design incentives for good contributions. On margins as an intermediary, contributor pay is “not the largest factor”: verifiable quality at scale is, and the products will move up-market, from “the Chinese restaurant from the side of the road” to Michelin-star kitchens, from an average tennis player to “what it means to be Federer.”

The data mix, and what will close the gap

A year earlier, people argued for one data type only. Now “all of this is essentially a data pot of soup”: about a million hours of human egocentric data, tens of thousands of hours of UMI data, and a few hours of in-domain teleoperation that closes the last gap to “four nines of reliability” in a specific setting. Tactile data has not yet entered the mix and simulation will. On AlphaGo’s move 37: Go had its rules as a complete simulator, but robotics “ultimately… [doesn’t] have a world simulator” for the physical world, so egocentric data both grounds basic physics and helps build that simulator. World models matter because they let a model “dream what it means to drop a cup and have it shatter without actually doing it.” He expects omni models that take force, pixels and other senses as input.

Huang asks whether general frontier models will “steamroll” robotics, given reports that a recent frontier model generalises well to robot tasks. Gao: “It’s actually quite expected,” since robot data is now in their pre-training mixes; specialised models may still reach four nines on particular tasks first.

Deployment

Useful robots exist today in more form factors than humanoids, such as delivery robots in hotels in China and South-East Asia and industrial cleaning robots. Laundry folding “will absolutely get solved in the near term,” and pick-and-place and packaging on assembly lines “within the next year or two.” The home is the hardest place to deploy, “making sure that the robot doesn’t fall on the baby,” although home data is the easiest to collect. He expects the deployment model to be “a modern robotics version of a systems integrator”: one that installs a robot and then makes it more capable over time with in-domain data. Asked for a number, he predicts robots “many orders of magnitude more” than people “in a decade plus.”

Dynamic-capabilities reading

• strategic-renewal/business-model — training data as a manufactured, licensed product, and the systems integrator that improves the robot it installed.
• digital-transforming/navigating-innovation-ecosystems — Mecka contributed data, processing and tooling to EgoVerse, a research consortium on egocentric data, to “lift up everybody in this wave” because “we are so far behind as an industry”; its contributors are businesses across many industries.

How it connects

• industrial-ai-agents — robots enter constrained industrial tasks first, through a new kind of systems integrator.
• foundation-models — robotics foundation models are limited by data, not compute or algorithms, and general models are picking up robot skills from robot data in their pre-training mix.
• Hines-Pierce (McKinsey Podcast) — the case that value moves to execution in the physical world.
• Luo (Null Labs) — the synthetic alternative to collected data, from defence.

What was actually ingested

The full episode (45:12, length_seconds: 2712) from auto-generated captions; the company’s name and technical terms are corrected in quotes. The names of the EgoVerse researchers are garbled in the captions and left out. No stills: a studio podcast.

Linked entities and concepts

• Sequoia Capital — publisher; the host is a Sequoia partner. Updated in this ingest.
• industrial-ai-agents, foundation-models — updated in this ingest.
• dynamic-capabilities — the two cells tagged above.

Dangling (single-source mention, deferred): Josh Gao, Mecka, Sonya Huang (named on several Sequoia sources, never author:), EgoVerse.

Scope and reliability

A founder describing his own company on an investor’s podcast; whether Sequoia is an investor in Mecka is not stated. The numbers (millions of hours, hundreds of petabytes, the converged mix) are his and are not documented in the episode; the mix he describes is the industry’s current recipe as he sees it, not a published result. The forecasts (laundry folding soon, robots outnumbering people in a decade) are his. The value for this wiki is an operational account of physical-AI data collection, which the corpus has lacked.