Can $2 Billion Startup Generalist AI Solve the Robot Training Conundrum?

Jack of all trades or a master of one? For now, robots should probably just learn how to dust the house without accidentally smashing the flat-screen TV.

But accomplishing that is hardly automatic. Last week, Nvidia- and Jeff Bezos-backed robotics software startup Generalist AI raised $400 million (at a nifty $2 billion valuation) to help build its AI-for-robots system. Their trick for training the robot army of tomorrow? An army of humans today.

Training AI chatbots required downloading all of humanity’s printed history. Training AI robots to interact in the physical world requires a similarly gargantuan set of training data. The only problem? The dataset for physical-world interaction and navigation is theoretically much more complex than language. Oh, and it doesn’t really exist.

For Generalist AI, the solution is to make the data itself. The company is doing so by way of what it calls “grippers,” a sort of dummy-prototype version of its flagship robotic hands that real-life humans can wear like gloves and puppeteer. It has sent thousands of grippers to humans around the world, who are wearing the gloves while accomplishing simple tasks like tying zip ties and opening bags of chips; the goal is to create 500,000 hours worth of real-world data, the company told Bloomberg last week.

It’s just one of many solutions robotics companies have recently devised to solve the data bottleneck:

Vibe Shift: The schemes don’t end there. Last week, New York City-based company Shift went viral with a promotional video promising to send humans to clean your messy apartment for free … so long as it can record the whole process and send the video to its German-based parent company, MicroAGI. Which begs the question: Will MicroAGI eventually clean our apartments for free, too?