A ROBOT barista deployed at a Punggol hawker centre stylishly pours coffee from one jug to another, as though it were a seasoned kopitiam uncle. Then it misses its aim, and steaming coffee gushes onto the table and floor. An onlooker dramatically cries: “Aiyoooh!”

This scene of a robot gone wrong was captured in a video and uploaded to Facebook some months ago, with the wry caption: “This is the reason why robot baristas cannot replace human beings”. It went viral, with several commentators agreeing with the caption.

The Punggol bot could have actually avoided the mishap had it detected that the jug on the table was askew, which was why it missed its aim. In its place, even a child might have been able to spot the mistake, set the jug straight and pour the coffee smoothly.

This incident speaks to a big paradox: Humans have built machines with the intelligence of even a PhD holder. But we struggle to reverse-engineer the intuitive abilities that a child has, to simply move one’s hands and manipulate objects.

This phenomenon is called Moravec’s paradox, named after the influential roboticist Hans Moravec. As he observed in 1988, it is easy to make computers pass exams or play checkers, but “difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility”.

Some reckon that this is why the development of robotics has lagged that of artificial intelligence (AI). Ask ChatGPT to crack a maths problem in mere seconds? No sweat. Get a robot to pour coffee into a tilted jug? Aiyoooh.

A NEWSLETTER FOR YOU

Friday, 2 pm

Lifestyle

Our picks of the latest dining, travel and leisure options to treat yourself.

Sign Up

The paradox could also explain why innovations such as self-driving cars have progressed more slowly than expected.

Earlier this month, driverless taxi companies, including Waymo and Cruise, gained the green light to ferry passengers all over San Francisco without restrictions. And yet numerous reports have surfaced of chaotic behaviour by the RoboCabs, from holding up traffic and blocking emergency vehicles, to even getting stuck in wet concrete.

Remember when Lyft’s co-founder said most of its rides would be driverless by 2021? We’re still miles away from that future.

Learning from the littlest ones

How do we tackle Moravec’s paradox? The answer: emulate how children learn. I gained this insight from a speech by UC Berkeley professor Jitendra Malik, at a recent data and AI conference in San Francisco.

In 2021, Malik and a team of researchers at UC Berkeley, Carnegie Mellon and Facebook (now Meta) managed to hit a breakthrough: They trained a dog-like robot to figure out how to adapt to and walk in difficult terrain it had never encountered before, in real time.

Be it hiking trails, sand, a pile of planks or even an oil-covered slope, the dog-bot could conquer it all and even recover its balance after nearly falling.

The research team’s solution, called “rapid motor adaptation” (RMA), combines a base policy – an algorithm that helps the robot move – together with an “adaptation module”. The module allows the robot to teach itself about its surroundings, based on information from its own body movements.

“For example, if a robot senses that its feet are extending farther, it may surmise that the surface it is on is soft and will adapt its next movements accordingly,” explains a blog post on the UC Berkeley website.

It struck me that this isn’t too different from how a baby learns to walk – via trials and feedback from its own movements. Indeed, the learning process of children can be a huge source of inspiration for robot builders.

Sharing his research at the conference, Malik cited how the father of computer science Alan Turing wrote in 1950: “Instead of trying to produce a program to simulate the adult mind, why not rather try to produce one which simulates the child’s?”

After all, children begin to accumulate “training data” right from the crib, through experimentation and learning, Malik explained. “When your toddler is being difficult and throwing food out, you should say, she’s actually a scientist in a crib; she’s conducting an experiment from which she is building models of the world around us.”

AI systems, including ChatGPT, have had the benefit of training on vast amounts of data available online, Malik noted. But obtaining the data to teach a robot movement and perception is so much more complicated.

“Think of what’s the data needed for sensorimotor training? You need to know what are… all my muscles commands and what are my neural activations. Hey, that’s pretty personal. I’m not uploading that on the Web,” he said, highlighting the need for “new clever ways” of solving this problem.

RMA is one novel method that researchers have come up with. To build better robots, more attention needs to be paid to how exactly children learn and ways that can be reverse-engineered.

The key to building robust robotaxis – or even kopi robots – just might lie in the brains of our littlest ones.

Moravec’s paradox is just one of many weird and wonderful phenomena out there that can teach us plenty about the world we live in. Every month, this column will go off tangent from the news and look into more curiosities in various fields, from finance and economics to science and psychology, or even beyond.