AI chatbots have advanced rapidly over the past few years, so much so that people are now using them as personal assistants, customer service representatives and even therapists.
The large language models (LLMs) that power these chatbots were created using machine learning algorithms trained on the vast troves of text data found on the internet. And their success has many tech leaders, including Elon Musk and NVIDIA CEO Jensen Huang, claiming that a similar approach will yield humanoid robots capable of performing surgery, replacing factory workers or serving as in-home butlers within a few short years.
But robotics experts disagree, says UC Berkeley roboticist Ken Goldberg.
In two new papers published online today (Aug. 27) in the journal Science Robotics, Goldberg describes how what he calls the "100,000-year data gap" will prevent robots from gaining real-world skills as quickly as AI chatbots are gaining language fluency. In the second article, leading roboticists from MIT, Georgia Tech and ETH-Zurich summarize the heated debate among roboticists over whether the future of the field lies in collecting more data to train humanoid robots or relying on "good old-fashioned engineering" to program robots to complete real-world tasks.
Below, UC Berkeley News spoke with Goldberg about the "humanoid hype," the emerging paradigm shift in the robotics field and whether AI really is on the cusp of taking everyone's jobs.
UC Berkeley News: Recently, tech leaders like Elon Musk have made claims about the future of humanoid robots, such as that robots will outshine human surgeons within the next five years. Do you agree with these claims?
Goldberg: No; I agree that robots are advancing quickly but not that quickly. I think of it as hype because it's so far ahead of the robotic capabilities that researchers in the field are familiar with.
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.
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.
What are the limitations that will prevent us from having humanoid robots performing surgery or serving as personal butlers in the near future? What do they still really struggle with?
The big one is dexterity, the ability to manipulate objects. Things like being able to pick up a wine glass or change a light bulb. No robot can do that.
It's a paradox - we call it Moravec's paradox - because humans do this effortlessly, and so we think that robots should be able to do it, too. AI systems can play complex games like chess and Go better than humans, so it's understandable that people think, "Well, why can't they just pick up a glass?" It seems much easier than playing Go. But the fact is that picking up a glass requires that you have a very good perception of where the glass is in space, move your fingertips to that exact location and close your fingertips appropriately around the object. It turns out that's still extremely difficult.
In your new paper, you discuss what you call the 100,000-year "data gap." What is the data gap, and how does it contribute to this disparity between the language abilities of AI chatbots and the real-world dexterity of humanoid robots?
To calculate this data gap, I looked at how much text data exists on the internet and calculated how long it would take a human to sit down and read it all. I found it would take about 100,000 years. That's the amount of text used to train LLMs.
We don't have anywhere near that amount of data to train robots, and 100,000 years is just the amount of text that we have to train language models. We believe that training robots is much more complex, so we'll need much more data.
Some people think we can get the data from videos of humans - for instance, from YouTube - but looking at pictures of humans doing things doesn't tell you the actual detailed motions that the humans are performing, and going from 2D to 3D is generally very hard. So that doesn't solve it.
I believe that robotics is undergoing a paradigm shift.
Ken Goldberg
Another approach is to create data by running simulations of robot motions, and that actually does work pretty well for robots running and performing acrobatics. You can generate lots of data by having robots in simulation do backflips, and in some cases that transfers into real robots.
But for dexterity - where the robot is actually doing something useful, like the tasks of a construction worker, plumber, electrician, kitchen worker or someone in a factory doing things with their hands - that has been very elusive, and simulation doesn't seem to work.
Currently people have been doing this thing called teleoperation, where humans operate a robot like a puppet so it can perform tasks. There are warehouses in China and the U.S. where humans are being paid to do this, but it's very tedious. And every eight hours of work gives you just eight more hours of data. It's going to take a long time to get to 100,000 years.
Do roboticists believe it is possible to advance the field without first creating all this data?
I believe that robotics is undergoing a paradigm shift, which is when science makes a big change - like going from physics to quantum physics - and the change is so massive that the field gets broken into two camps, and they battle it out for years. And we're in the midst of that kind of debate in robotics.
Most roboticists still believe in what I call good old-fashioned engineering, which is pretty much everything that we teach in engineering school: physics, math and models of the environment.
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 to get us to fully functional humanoid robots.
This new wave is very inspiring. There is a lot of money behind it and a lot of younger-generation students and faculty members are in this new camp. Most newspapers, Elon Musk, Jensen Huang and many investors are completely sold on the new wave, but in the research field there's a raging debate between the old and new approaches to building robots.
What do you see as the way forward?
I've been advocating that engineering, math and science are still important because they allow us to get these robots functional so that they can collect the data that we need.
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.
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.
Ken Goldberg
Waymo, Google's self-driving car company, is doing that. They're collecting data every day from real robot cars and their cars are getting better and better over time.
That's also the story behind Ambi Robotics, which makes robots that sort packages. As they work in real warehouses, they collect data and improve over time.
In the past, there was a lot of fear that robotic automation would steal blue-collar factory jobs, and we've seen that happen to some extent. But with the rise of chatbots, now the discussion has shifted to the possibility of LLMs taking over white-collar jobs and creative professions. How do you think AI and robots will impact what jobs are available in the future?
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.
But there are certain jobs - those that involve routinely filling out forms, such as intake at a hospital - that will be more automated.
One example that's very subtle is customer service. When you have a problem, like your flight got canceled, and you call the airline and a robot answers, you just get more frustrated. 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. Some claim that AI can read X-rays better than human doctors. But do you want a robot to inform you that you have cancer?
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.
This interview has been edited for length and clarity.
