Robotics experts are urging caution in the face of growing claims that humanoid robots will soon be able to perform complex tasks like surgery, factory work, or household chores. While large language models (LLMs) have led to rapid advances in AI chatbots used as assistants and customer service representatives, robotics researchers say that transferring this success to physical robots faces significant hurdles.

Ken Goldberg, a professor at the University of California, Berkeley, outlined these challenges in two papers published August 27 in Science Robotics. He described a “100,000-year data gap” between the amount of information available to train language models and what is needed for robots to gain comparable real-world skills. Goldberg explained that while vast amounts of text on the internet enabled LLMs’ fluency, there is far less equivalent data for training robots—and even more would be required because robotic tasks are more complex than processing language.

“We’re all very familiar with ChatGPT and all the amazing things it’s doing for vision and language, but most researchers are very nervous about the analogy that most people have, which is that now that we’ve solved all these problems, we’re ready to solve [humanoid robots], and it’s going to happen next year,” said Goldberg. “I’m not saying it’s not going to happen, but I’m saying it’s not going to happen in the next two years, or five years or even 10 years. We’re just trying to reset expectations so that it doesn’t create a bubble that could lead to a big backlash.”

Goldberg identified dexterity—the ability for robots to manipulate objects—as a particularly tough challenge. “It’s a paradox — we call it Moravec’s paradox — because humans do this effortlessly... But the fact is that picking up a glass requires that you have a very good perception of where the glass is in space... It turns out that’s still extremely difficult,” he said.

Attempts to gather training data from sources like online videos or by simulating robot movements have met with limited success when applied to practical tasks requiring fine motor skills. Teleoperation—where humans remotely control robots—is another method being used but yields only small increments of new data relative to human labor hours invested.

Goldberg noted an ongoing debate within robotics over whether progress will come primarily from collecting ever-larger datasets or from continued reliance on traditional engineering principles. “Most roboticists still believe in what I call good old-fashioned engineering... But there is a new dogma that claims that robots don’t need any of those old tools and methods. They say that data is all we need,” he said.

He suggested hybrid approaches might be key: “This is a way to bootstrap the data collection process. For example, you could get a robot to perform a task well enough that people will buy it, and then collect data as it works.”

Companies such as Waymo and Ambi Robotics are already using deployed systems—self-driving cars and warehouse sorting robots—to gradually improve performance through real-world experience.

Despite advances in deep learning enabling some progress—for instance allowing some robots to grasp and move objects more easily—Goldberg remains skeptical about near-term displacement of blue-collar jobs by humanoid machines. “To my mind as a roboticist, the blue-collar jobs, the trades, are very safe. I don’t think we’re going to see robots doing those jobs for a long time,” he said.

He added some white-collar roles involving routine paperwork may become automated sooner but highlighted limitations in areas requiring empathy or sensitive communication: “Many companies want to replace customer service jobs with robots... but the one thing a computer can’t say to you is ‘I know how you feel.’ Another example is radiologists... But do you want a robot to inform you that you have cancer?”

Goldberg concluded: “The fear that robots will run amok and steal our jobs has been around for centuries, but I’m confident that humans have many good years ahead — and most researchers agree.”

Deep learning techniques continue helping some types of industrial robots become better at handling objects during sorting or assembly tasks; however widespread adoption across varied industries remains distant due largely to challenges outlined by experts like Goldberg.