Inside Mecka's Robot Data Factory: How Human Demonstrations Could Train the Next Generation of Robots

The conversation opened with an idea from science fiction: once robotics and AI have fully generalized and solved all tasks humans can do, the final frontier will be organic, or biological. The host pointed to Asimov, Ender's Game, and Hitchhiker's Guide to the Galaxy, in which the planet at the end is revealed to be a giant organic computer trying to understand the meaning of life.

The host introduced Josh Gao as founder and CEO of Mecka, a company building a data and deployment business for robotics. The host argued that robotics is fundamentally data-constrained, unlike language models, which benefited from vast free human data on the internet.

Gao explained that Mecka creates, processes, and licenses much of the data used to train robotics foundation models. He said robotics lacks the cold start that language models had with the internet. Robot data, in his description, consists of three things: pixels, actions, and logic.

Because this curated, processed, clean data is scarce, Mecka's work includes building different kinds of hardware, such as sensorized wearables for humans, running operations to collect data, partnering with individual contributors, businesses, and tasks across environments, and post-processing data while owning the quality gates that make datasets useful for teaching robots about the world.

Asked about the phrase 'teaching robots about carbon life' on the company website, Gao said he made the site himself and that the phrase draws on the same science fiction idea of a biological final frontier. He noted that the human brain runs on just 20 watts and that humans are self-healing.

Along the journey of getting robots to generalize fully, Gao said, the company will slowly show them what carbon life is made of.

Gao described an unconventional path. Before Mecka, he built software for cloud kitchens, including order management and back-of-house tools, and he once skipped class to install kiosks in rural New York restaurants. He said his co-founder had built a stablecoin that was ultimately sold to Coinbase.

When the team entered robotics, Gao said they conducted a full discovery process. They contacted hundreds of people in the field, read every paper they could find, and visited robot labs. By late 2024, when GPT was taking off, he said it seemed obvious that data and compute would become very important across the industry.

He pointed to the Ego4D dataset, which contains roughly 4,000 hours of human demonstration data. Examining it, the team found daily-life activities such as bowling, which suggested broad use cases at that scale. Gao said coming in with a fresh perspective let the team reason from first principles and iterate closely with people who were much smarter and held different views of the future.

Gao defined egocentric data as video of humans demonstrating and performing tasks from a first-person point of view. The core idea was that egocentric video could close much of the vision gap between human data and robotics data, showing robots what it is like to interact with the world from a first-person perspective.

When the team explored the field, most robot data then came from teleoperation, in which humans control robots to perform demonstrations. Gao said this was very hard to scale. Drawing on scaling principles seen in language models, the team concluded that data would need to scale far more significantly. Their early intuition was that humans and robots share many similarities, and that humanoid robots were just beginning to emerge, so humans could contribute to fleet-style learning in a way similar to Tesla's fleet approach.

Gao said Mecka uses several approaches to sourcing data. Its direct-consumer mobile app lets people anywhere in the world sign up, register their profession, and receive a streamed sequence of tasks, such as cleaning a bedroom or cooking breakfast. He described this as a lightweight way to gather diverse pretraining data.

A second approach is business-to-business recruiting, because much valuable work happens beyond the home. Gao listed examples across a wide range of physical human work, including bakeries and PCB manufacturing lines.

Gao said the data engine is far more sophisticated and painful than outsiders might expect. Many of the tools needed did not exist until Mecka designed and produced its own hardware. He gave tactile gloves as an example: no off-the-shelf glove reliably measures how much force a person applies in daily work.

Operations also require substantial distribution and processing infrastructure. Gao said the company operates in several locations to sustain hundreds of petabytes of data, and must process that data, maintain strong quality gates, and apply computer vision to make it useful.

To decide who to recruit, Mecka tracks a heat map across jobs, industries, and dimensions of diversity. When the map shows a gap, such as too few podcasters, the company goes after those contributors.

Asked whether Mecka can earn a durable margin, if a competitor approaches a contributor for the same data, Gao said the contributor payment is not the largest factor in making data useful end to end. Even if that baseline becomes more competitive, he said, companies can differentiate by verifiably delivering high-quality data at very large scale using the right hardware, operations, and software.

Gao said the products will scale and mature, and that egocentric podcast data could be one SKU among many. He also expects the types of data needed to change, for example moving from roadside Chinese restaurants to Michelin-starred kitchens.

The host raised the investor's concern: a car has two or three degrees of freedom, while a robot hand can have dozens. Robots may also need force data and must operate in a far harder environment. The host estimated that billions, trillions, or even quadrillions of hours might be needed and asked whether the world's GDP could support that cost.

The host noted that roads already contain a few million hours of video data, which Gao's framing implied is still many orders of magnitude short of what robots need.

Gao responded that progress will come through unlocking commercial use cases rather than a single binary breakthrough. He said that as generalization and scaling laws give companies the courage to scale, they can also collect in-domain, in-task data to close gaps in post-training, for example achieving four nines of reliability in specific real-world environments.

Gao argued that people underestimate how diverse robotics is. Humanoids will be popular, but many other form factors already exist. He cited hotels in China and Southeast Asia where service, cleaning, and delivery robots handle tasks from the front desk to hotel room doors, noting that these robots are widely deployed there and only beginning to gain popularity in the US and North America.

He expects early deployments to include industrial cleaning robots and then to progress to early manipulation tasks. Asked when a Tesla FSD-like flywheel might arrive, he said the company is still far from it, but that a slow transition will occur as commercial tasks are solved and deployment data is collected.

Gao said he believes in both vertical and horizontal approaches. Vertical companies are highly task-specific, spanning hardware, models, and data. The horizontal ecosystem will include a wide range of hardware and different types of open models.

He compared the future to a modern systems integrator. In traditional settings, integrators install a product in a competitive, tight-margin business. Now, he said, an installed robot can become more data-driven and productive over time, and a bot installed last week might do something previously thought impossible.

Asked for investing advice, Gao said industrial bots will be the best early category. These robots already include much traditional robotics and operate in constrained tasks, such as data center maintenance. The environment is far more controlled than the open world, where a baby might appear on the floor.

He also named hotel delivery, where robots could hand items from the front desk to rooms. He added a joke that if Mecka helped deploy a robot that mines an asteroid, the company would have done a good job.

Gao said the industry is converging on a data recipe. A year ago, many argued that only UMI data or only egocentric data would work. He now sees all data sources as ingredients in one pot.

He described the current recipe as roughly a million hours of human data across egocentric and other modalities, tens of thousands of hours of robot data, and last-mile teleoperation and in-distribution data gathered in just a few hours of demonstration. He noted that tactile data has not yet entered the mix, but would be needed for models to learn how much force to apply to a door handle, and that simulation will also become part of the mix.

The host raised a frequent objection: watching millions of hours of tennis would still not teach someone to beat Roger Federer. Gao said he likes the AlphaGo analogy. DeepMind's early work involved large-scale behavior cloning of experts, followed by self-play under the game's rules, which led to the surprising move 37.

In robotics, Gao said, the rules are the entire physical world, which has rich detail that no current world simulator can recreate. Egocentric data helps ground basic physics and behavior cloning and contributes to what a world simulator might look like. He expects robots may eventually play tennis like Roger Federer, probably through a hybrid of simulation and real-world learning.

Asked whether scaling the current data mix will reach the promised land of a robot that folds clothes or cooks a meal, Gao said that scaling the ratio by roughly an order of magnitude would solve many tasks along the way. Early successes are likely to occur in more constrained settings, such as business operations with specific standard procedures.

Full real-world complexity, such as weather, will require adding other modalities, especially tactile data. Gao described a move toward omni-models that accept any input, including force, pixels, and possibly smell, and output actions, voice, or controls. He said smell-sensing approaches are still very early.

Asked about recent foundation models performing well on robotics, Gao said it is quite expected once you consider the robot data now entering training mixes. He closed by imagining a world full of robots, including unusual form factors such as the Kawasaki Corleo, a robot panther-horse hanging in Mecka's New York office.

Gao predicted that in a decade or more there would be many orders of magnitude more robots than people, around 100 billion or more.