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

Josh Gao describes Mecka AI as a data and deployment company for robotics foundation models. Mecka creates, processes, quality-checks, and licenses physical-world training data consisting of pixels, actions, and logic. Unlike language models, robotics did not receive trillions of freely available, clean internet tokens that describe how to manipulate the world. Mecka therefore builds sensing hardware, recruits individual contributors and businesses across environments, collects demonstrations, post-processes them, and owns the quality gates that make the resulting datasets usable. Sonya frames this as a potentially foundational position in a coming robotics “golden age.”

Asked whether collecting and paying for robot data is economically feasible, Josh argues that robotics is not starting from zero. Video models, language models, world-model research, and web-scale pretraining have already supplied useful infrastructure and learned representations. Robotics data is becoming a substantial pretraining market, including video that teaches models to generate physics-grounded realities. Josh says the economics depend on which commercial use cases each scale of data unlocks: simple pick-and-place tasks may arrive first, while the long tail of physical work could require many orders of magnitude more data than exists today.

Josh explains the whimsical language on Mecka’s website—“teaching robots about carbon life”—as a science-fiction reference. In stories by Asimov, in Ender’s Game, and in The Hitchhiker’s Guide to the Galaxy, the ultimate frontier after generalized intelligence is biological or organic understanding; in the latter, a planet becomes a giant organic computer trying to understand the question behind life. Josh connects that idea to human capabilities: the brain runs on roughly 20 watts and the body self-heals. Teaching robots to generalize may therefore require progressively showing them what carbon-based life is made of.

Josh and his co-founder came to robotics from unconventional backgrounds. Josh had been building a Toast-like operating system for cloud kitchens, including order management and physical deployment; during university he skipped class to install kiosks in rural Rochester, New York, sometimes staying overnight to replace restaurant front-of-house systems. His co-founder built a stablecoin that was eventually sold to Coinbase and had ranked among the top thousand players in four competitive games, including League of Legends. Both are Canadian and share a strong science-fiction interest, so they wanted their next company to address a difficult problem with an exciting future rather than remain in payments.

Josh defines egocentric data as first-person video of humans performing tasks and explains why Mecka initially bet on it. Robot cameras experience the world from a point of view, so human first-person footage can help close the visual gap between ordinary human activity and robot perception. At the time, much robotics data came from teleoperation—people controlling robots to demonstrate tasks—which was difficult to scale. Mecka reasoned that humans share many physical regularities and that a distributed fleet of people could contribute to learning the world’s physics, analogous in spirit to Tesla’s vehicles contributing driving data.

Mecka’s earliest pilot was a mobile app that recruited Chinese “aunties and uncles” to record themselves cooking, cleaning, and doing daily tasks. The company also tried UMI data and other approaches before moving toward custom hardware. Sonya asks whether video alone can capture force—such as how hard someone squeezes an egg or pulls a door handle—and Josh agrees that force data will be important. Risk cameras can provide a rough contact-time substitute, but Mecka is developing tactile gloves that measure the force used in ordinary tasks. Its broader approach is to “sensorize” humans with cameras, wearables, and other devices.

Mecka collects data through both consumer mobile apps and a B2B recruitment engine covering homes and businesses, from bakeries and PCB manufacturing lines to other forms of physical work. Much of the required equipment did not exist off the shelf, so Mecka designs and manufactures hardware, iterates through several product cycles, distributes it to thousands of homes and businesses, and processes the resulting uploads. Josh says some collection sites needed 100-gigabyte upload lines to handle hundreds of petabytes. Quality systems must detect edge cases and fraud; one contributor submitted YouTube-style POV videos of supposedly working at McDonald’s rather than genuine work.

Josh says contributor payments are not necessarily the largest cost in producing useful robotics data, even though competition for contributors could increase. Mecka’s defensibility instead comes from making data verifiably high quality at scale through specialized hardware, operations, software, and quality controls. The company tracks a heat map of jobs, industries, and diversity gaps; if it lacks enough data from a profession such as podcasting, it recruits specifically for that category. Josh expects the product catalog to evolve from broad egocentric data toward more specialized datasets—such as Michelin-star cooking or Federer-level tennis—rather than treating all human demonstrations as interchangeable.

Josh estimates that robotics has finally reached millions of data hours, a major increase from the roughly 4,000 hours in the Ego4D dataset, but remains far behind autonomous driving. Tesla spent more than a decade collecting data while solving mainly acceleration and direction; robots may have dozens of degrees of freedom and may also need force data. Josh rejects a binary “robotics solved or not” framing: generalization can unlock increasingly narrow commercial tasks along the way. Home data is plentiful, but homes are among the hardest environments because of safety and edge cases. He expects constrained packaging, pick-and-place, assembly-line manipulation, cleaning, and hotel-service robots to arrive first.

Josh says the emerging recipe resembles roughly a million hours of human data, including egocentric data; tens of thousands of hours of robot or UMI data; and a smaller amount of last-mile teleoperation and in-domain demonstrations, sometimes only a few hours. He rejects choosing one “religion”: tactile data, simulation, internet video, human demonstrations, robot trajectories, and teleoperation will all matter. He compares future robotic self-play to AlphaGo’s progression from behavior cloning to discovering move 37, but notes that robotics lacks a simulator whose rules capture the entire physical world. He expects hybrid self-play in simulation and reality, with multimodal models eventually using force, pixels, perhaps smell, and producing actions or voice.

Josh expects an FSD-like robotics flywheel only after commercial robots are deployed in enough constrained tasks to generate failures, corrections, and in-domain data. Mecka could then collect deployment data, post-train models, and help enterprises integrate robots into their operations. He anticipates both vertical companies that build complete task-specific systems and a broader ecosystem of varied hardware and open models, connected by a modern systems integrator that installs robots and makes them increasingly data-driven. Industrial robots and data-center maintenance are attractive early categories; hotel delivery and manipulation are also promising. The long-term aspiration is a robot capable of mining an asteroid.

Asked to imagine the world Mecka is building toward, Josh returns to absurd science-fiction form factors. He points to Kawasaki’s Corleo, a robot that looks like a panther-dog-horse, and jokes that a promotional image of a woman and child riding it represents him and his mother. He believes future robotics will include many strange, specialized shapes rather than only humanoids. Josh predicts that the world could contain more than 100 billion robots within a decade—many orders of magnitude more than people—and jokes that there should be at least another order of magnitude of panther-dog horses. The conversation ends with Sonya receiving a Corleo image as a gift.