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

Summary · 4 min read

Josh Gao is the founder and CEO of Mecka AI, a data and deployment company for robotics. Mecka creates, processes and licenses training data, and builds sensor hardware and collection operations for robotics models.

Josh Gao says robotics cannot rely on free internet-scale action data, so Mecka builds sensors, collection operations and quality controls to create usable training datasets. He argues that robotics will advance by unlocking commercial tasks incrementally, with constrained industrial settings likely to come before homes and broadly capable humanoids. Gao describes an emerging training mix of roughly a million hours of human data, tens of thousands of hours of UMI data and a few hours of in-domain teleoperation, with tactile and simulation data still to be added. He expects deployment feedback and multimodal models to extend capabilities, rather than a single GPT-3-style breakthrough.

Minute by minute

1. 1:32

   Robotics data must pay its way. Gao says Mecka creates, processes and licenses data for robotics foundation models, filling a gap left by the lack of free internet-scale demonstrations. He describes the needed data as pixels, actions and logic, and says commercial use cases can make collection economically viable as capabilities expand from simple pick-and-place tasks.

2. 4:16

   Science fiction shaped Gao's path. Gao says he framed Mecka's mission around teaching robots about carbon life because humans are self-healing and the brain runs on 20 watts. Before robotics, he built software for cloud kitchens, while his cofounder built a stablecoin that was sold to Coinbase. He says their interest in science fiction and an exciting future drew them to robotics.

3. 8:29

   Egocentric data bridges human and robot views. Gao defines egocentric data as first-person video of people doing tasks, which he says can help robots learn how humans see and interact with the world. Mecka's early pilot used a mobile app to collect cooking and cleaning demonstrations, before the company moved toward custom hardware and industry partnerships. He says tactile gloves can capture force that video alone misses.

4. 11:48

   Data factories need custom hardware and logistics. Gao says Mecka recruits contributors through consumer apps and businesses across physical-work categories, then designs hardware when off-the-shelf tools fall short. The company built engineering and logistics operations to distribute sensors across thousands of homes and businesses, and set up 100-gigabit upload lines to handle hundreds of petabytes. It then processes the data with computer vision and quality gates.

5. 15:38

   Quality control protects data value. Gao says quality checks must adapt as new real-world edge cases appear, and identifies fraud as a persistent risk: some contributors submitted YouTube POV videos of McDonald's work as if they had recorded them. He says Mecca must inspect collection and design incentives for valid data. He argues its margin depends less on contributor cost than on high-quality data at scale, supported by hardware, operations and software.

6. 19:25

   Commercial tasks can unlock robotics gradually. Gao says the industry has reached millions of hours of data, but remains many orders of magnitude behind Tesla's decade of work on acceleration and direction, only two degrees of freedom. He argues progress need not wait for a single all-capable robot: constrained tasks can unlock deployments and support in-domain training. He says home safety makes homes especially difficult, noting Mecka spent about a year and a half plus on laundry folding, while expecting pick-and-place and packaging tasks on assembly lines within the next year or two.

7. 23:19

   Deployment will grow through integration. Gao says useful robots already operate in settings such as hotels and cleaning, and expects constrained industrial applications to spread before general-purpose humanoids. He says an FSD-like flywheel will require deployed robots to generate failure cases and corrections, but thinks that remains some way off. He sees a role for systems integrators that install robots and use data to expand what businesses can do; he cites asteroid mining as an aspirational task.

8. 29:50

   Robot training needs a mixed data recipe. Gao says Mecka's custom baseball caps, sensor hub, external cameras and tactile gloves aim to capture how people interact with the world; Mecka also contributed data, action trajectories, quality gates and tooling to EgoVerse. He describes a current recipe of roughly a million hours of human data, tens of thousands of hours of UMI data and a couple of hours of teleoperation and in-domain demonstrations. He says tactile and simulation data still need to enter the mix.

9. 33:48

   Robotics needs simulation and self-play. Gao compares robot learning to AlphaGo: behavior cloning can teach human play, but self-play under the game's rules enabled moves people had not anticipated. He says robotics lacks a simulator that can recreate the physical world's detail, so egocentric data can ground basic physics and help build one. He sees world models as a way to imagine outcomes such as a cup shattering and potentially map video predictions to action trajectories.

10. 36:44

    Scaling will unlock tasks before generality. Gao says scaling today's data mix by an order of magnitude could solve many constrained business tasks, but broader capabilities will require more modalities, including tactile sensing and possibly smell. He says multimodal foundation models can generalize across domains, while specialized robotics models may reach four nines of capability and reliability on defined tasks. He calls recent foundation-model abilities in robotics an expected result of adding robotics data to training.

11. 40:45

    Robots could outnumber people. Gao says robotics may produce unconventional forms such as the Kawasaki Corleo, and imagines robots tackling tasks humans cannot, including asteroid mining. Asked for an estimate a decade out, he agrees there could be 100 billion robots, more than the number of people.

Published Oct 20, 2026.