Sonya Huang invests in companies building robot foundation models. These are large general-purpose models meant to control many kinds of robots. She has done some arithmetic that worries her. A car has two or three degrees of freedom, meaning the independent ways a machine can move. A robot has far more. Josh Gao, founder and CEO of the robot-data company Mecka AI, supplied the number: "Even dozens in the hand."
Add the need for force data and the harder physical world, Huang said, and "you might need billions, trillions, quadrillions of hours" of training data. "Is there even enough GDP in the world to support the cost to bring that in?"
The exchange came on Training Data, the podcast Huang hosts for Sequoia Capital, in an episode published October 20, 2026. Huang noted that a gap that size also looks like a lot of revenue for robot data factories like Gao's. His answer was that robots do not need to close the whole gap before they become useful.
How far behind robotics is
Gao said the field has "finally entered a world where we are in the millions of hours of data." He contrasted that with Ego4D, a well-known collection of first-person video that he put at about 4,000 hours. The Ego4D paper, in its March 2022 revision, reports 3,670 hours recorded by 931 camera wearers across 74 locations in nine countries, mostly of unscripted daily activity.
Gao used self-driving cars to show how far there is still to go. In his account, Tesla spent more than a decade solving for two degrees of freedom, acceleration and direction, with millions of cars on the road, and has "a lot more than a few million hours of video data at this point." Measured that way, he said, robotics is "still many orders of magnitudes away from where we have to be, ultimately."
Gao's answer: it doesn't have to be binary
Gao rejected the idea that robotics will be solved all at once. It "doesn't have to be so much binary in that, oh, one day we've unlocked robotics or not," he said. "It is going to be a slow creep up into the world."
He described two steps. First, the results teams already see, with models generalizing to new situations and scaling laws (the pattern of models improving predictably as they get more data) at work, give companies "the courage to scale up" their data collection. Broad data of that kind gives a robot some ability to cope with unfamiliar environments. Second comes post-training, a later stage that tunes an already trained model for one job, using data from the exact setting where the robot will work. "Maybe I can collect data exactly from this room, exactly in this podcast demonstration," Gao said, to close the gap to what real deployments often require: "four nines of reliability," or success 99.99% of the time.
Huang summed it up as a steadily growing set of things robots can do, rather than a single jump from zero to one. Gao agreed: "we will unlock a lot of tasks along the way."
Why homes demand the most care
Huang pointed out an awkward mismatch. Home footage is the most plentiful robot data because anyone can film it, yet the home looks like one of the hardest places to put a robot. Gao agreed that homes are where the field will have to be "most diligent on by far." Testing and evaluation methods in robotics have only just started to mature, he said, and they have to make sure, for example, "that the robot doesn't fall on the baby."
He expects the first tasks to be constrained and goal-oriented, with standard operating procedures that already define what doing the job means. Clear rules let developers "hill climb" a task, improving it step by step against a measurable target. Huang offered laundry folding as an example. Gao said "we spent about a year and a half plus trying to solve" laundry folding and that it "will absolutely get solved in the near term." He gave a slightly longer horizon, "the next year or two," for pick-and-place and packaging work on assembly lines, manipulation jobs that traditional robotics could not solve.
Gao also argued that useful robots are already deployed, and that people "underestimate how diverse robotics can be." Humanoids will be "a very popular terminal state," he said, but many other robot shapes are in use today. He pointed to hotels in China and Southeast Asia, where robots clean, provide services and carry deliveries from the front desk to a guest's room once an Uber Eats order arrives. Such robots are widely deployed around the world, he said, and are only starting to spread in the US and North America. Uber Eats runs its own robot delivery in the US, though it works differently. The company's October 2025 merchant guide describes sidewalk robots that restaurant staff load and customers unlock through the app, with the robot handling the trip between them.
Gao expects some of the first robots to be industrial cleaning machines and other "thematic robotics" built for one kind of job. From there, he said, robots will step up through early manipulation tasks "all the way up until we get to that humanoid solving out of the box, any task in the future."
When the flywheel starts
Huang asked when robotics would get something like the flywheel behind Tesla's Full Self-Driving software, a loop in which deployed machines produce the data that improves the next model. Gao thinks that is "a while away." It will begin as commercial tasks get solved and robots in the field start producing in-deployment data, meaning failure cases and the corrections that fix them. How long that takes depends on how many types of tasks and settings remain unsolved, he said.
By then, Gao expects Mecka to look very different. The company now specializes in egocentric data, video and sensor readings recorded from a person's own point of view as they work. He said Mecka is "not fully subscribed to only egocentric data" and sees value in many kinds of data products, including what he called "Michelin star data" and edge cases the robots have not yet met. "If we believe that the world is incredibly vast and rich in detail, there's going to be no shortage of things to teach," he said. He also sees Mecka becoming a platform that collects domain data inside a customer's environment, runs the post-training and gives businesses "white glove" support as they adopt robots.
A new kind of systems integrator
Huang compared the question with how language models have spread through the economy. In that world, she said, there are vertical application companies, model makers scaling up their own sales efforts, and "armies of people" at service firms helping enterprises adopt AI. She asked what the physical-world version would be.
Gao said he believes in both approaches: vertical companies that build everything for one task, "all the way from hardware to models to data," and a horizontal ecosystem with many kinds of robot hardware and open models. He predicted "a modern robotics version of a systems integrator," the kind of firm that installs and connects equipment for a business. In the traditional version, he said, installation is "incredibly competitive" and margins are tight. The new version would install a robot and then make it "more data-driven and productive over time." The milestone he set: "can you actually have the bot you installed last week do something that it didn't think was possible a few days ago even?" When Huang asked whether she should therefore invest in both general and vertical robots, Gao said yes.
Favorite places to start
Asked for his favorite vertical categories, Gao chose industrial robots, where traditional robotics is already common and the tasks are constrained. Huang suggested data-center maintenance robots, and Gao agreed, because there "you just don't have to expect a baby to appear." He also likes hotel robots that could manipulate objects well enough to hand off a delivery.
Huang mentioned Gecko Robotics, which uses robots to help maintain ships. In a March 2023 announcement, Gecko said its wall-climbing robots take ultrasonic measurements for U.S. Navy maintenance planning. The company reported that on one Navy asset its robots captured more than 4.2 million data points, compared with 100 from conventional methods.
Gao's longest-range ambition is work no human can do. A joke at Mecka's all-hands meetings, he said, is that if the company one day helps deploy a robot that mines an asteroid, "we've probably done a pretty good job." He called it a job "that humans can't do, that we have no idea how to do just yet. But it seems inevitable."