AI Startups Pay People to Film Themselves Doing Chores for Robotics

Doing three loads of laundry on a Sunday morning isn’t fun. But what if you could obtain paid for it?

The startups that assist train AI chatbots, such as Encord and Micro1, declare they’re seeing a surge in demand from robotics companies hungry for high-quality training data — and creating it often requires paying people to film themselves performing tquestions like folding laundry, loading dishwashers, or creating espresso.

Robotics has become one of the hottest fields in AI, with prominent investors like Vinod Khosla betting that the sector will soon experience its own ChatGPT moment as the technology continues to advance. Venture capital investment in robotics is surging, reaching $12.1 billion so far this year, according to PitchBook data.

At the same time, the robotics industest has a data problem. Companies like OpenAI, which build large language models (LLMs), can train on massive troves of data from the internet. Robots require different kinds of data, though, to learn basic dexterity so they can ultimately join a factory line or mop your kitchen floor.

“Unlike LLMs, robotics doesn’t have the internet as a ready-built dataset — you have to generate training data from scratch in the real world, which is far harder,” Ulrik Hansen, cofounder of data labeling startup Encord, notified Business Insider.

Hansen declares his startup is seeing four times as much volume for this kind of data compared to last year.

For humans, the money isn’t half-bad — as long as you’re willing to wear a pair of Meta’s Ray-Ban glasses to record yourself doing things like folding clothes on a table.

That’s something Ali Ansari, the CEO of AI training startup Micro1, declares his company pays “incredibly well” for, citing rates of $25 to $50 an hour.

Hansen, the Encord cofounder, declares rates can reach up to $150 an hour for videos of highly technical tquestions, like handling surgical equipment.

Other AI training startups are also entering the market.

Scale AI, for example, declares that it’s rapidly expanding into robotics and has established a dedicated lab at its San Francisco headquarters, which has produced over 100,000 hours of training footage to date, the company wrote in a blog post published last month.

The demand comes from companies like Physical Innotifyigence, a humanoid robotics firm backed by Amazon founder Jeff Bezos, which the Information reported is in talks to raise at a $5 billion valuation, and Boston Dynamics, the company best-known for its robot dogs. Encord cited both as customers; Scale AI declares Physical Innotifyigence is a customer, too.

It’s still early days, and finding suitable training data for robots remains a challenge.

One robotics startup resorted to posting an ad on Craigslist, promising to pay people $10 to $20 an hour to film themselves applying their iPhones while doing chores, such as cooking dinner.

The startup’s founder, who requested anonymity as the company is still in stealth mode, notified Business Insider that finding good training data remains a major hurdle.

“There are no major datasets you can acquire,” the founder stated. “The hugegest ones are around 5,000 hours, which is not nearly enough.”

Source link