‘Self-aware’ robots can learn complex tasks by watching humans. Is that a good thing? : NPR

Imagine a robot that can wash your clothes, make your bed, cook your dinner, or stock the dairy section at your local grocery store. Humans have long been able to teach robots to perform individual tasks—but despite billions of dollars invested in robotics, instructing them on these more sophisticated tasks has been an elusive goal.

Now, a team of scientists in Switzerland has made progress in their quest to invent assistive robots that can act on complex instructions from humans. This development raises questions about whether such technology might someday learn not only to help humans, but also be capable of harming them.

Invention of the Personal Barista

For years, robotics scientist Sthirpragya Gupta has been dreaming about the things his robot could do. “I personally want a robot to make coffee for me,” says Gupta. He and his colleagues at the École Polytechnique Fédérale de Lausanne – an engineering school in Switzerland – stay up late in their laboratory in the Swiss Alps. “There is a lot of coffee consumption there,” says Gupta.

“If I could just say, ‘A little bit of sugar, a little bit more cream,’ things like that,” he says. “That would be a dream come true.”

One problem that robotics scientists and engineers like Gupta have long grappled with is that robots cannot perform tasks beyond those for which they have been specifically programmed. Gupta uses a tennis example to explain this point. He explains that robots may be able to learn to hit backhand shots. They can backhand that ball perfectly again and again and again and again. But if circumstances change – say, their opponent moves forward, or the lighting changes – everything falls apart. Human beings do not have any problem in adjusting to such changes. However, teaching a robot to adapt is much more difficult.

“It is very difficult to transfer this behavior from humans to robots,” says Gupta.

Until now, he hopes. Gupta and his colleagues have published a paper in the journal science robotics Demonstrating a new way to teach robots using machine learning, a type of artificial intelligence. This approach relies on kinetic intelligence – a robot’s inherent awareness of how its own body can move safely in space.

In a video demonstrating its technology, a one-armed robot attached to a base watches as a human trainer throws a ball into a small container. The robots then pick up the ball and mimic the instructor’s behavior, adjusting to its position and adjusting to their non-human body. The robots are then able to transfer these skills and knowledge to other robots.

“This could be a turning point,” says Robert Platt, who studies engineering and robotics at Northeastern University. Platt, who called the work a “breakthrough,” said there is not widespread consensus in the field of robotics about the path to creating effective robots through machine learning — but most agree that the problem these researchers are tackling is an important one. “More people may do this in the future,” he said.

Platt was hesitant to predict any specific timeline for robots becoming widespread household accessories. “We are at a point of very rapid change,” he says, referring to generative AI chatbots like ChatGPT or Cloud that have been widely adopted, he says, “One of the reasons I hesitate to make predictions – look what happened with large language models.” “We were so far away and then suddenly – we weren’t.”

A fine line between self-awareness and consciousness

If a robot can improve itself and teach others, does that make it self-aware?

“This robot looks like it’s capable of doing some pretty impressive things like learning,” says Susan Schneider, who studies artificial intelligence at Florida Atlantic University. “But this does not mean that something has a fully developed consciousness or internal awareness, in the sense that biological beings have.”

Schneider points out that a key difference between robots and humans is feeling. “Consciousness is the felt quality of experience,” she says. “When you sip your morning espresso, when you look at the beauty of a sunset, when you have a headache, it feels like something inside is you.”

But this lack of consciousness raises new questions about morality. “This immediately rings alarm bells in the mind of any AI security researcher,” says Schneider. She says later versions of such technology could potentially be weaponized against humans.

Researchers have taken care to include safety protocols to ensure that robots are not able to hurt people. However, they also acknowledge that future developments in this technology will require guardrails. “I think really soon we should have a regulatory framework governing who operates robots and how,” says Gupta.

Susan Schneider says humans are at a turning point with robotics. “It’s a very exciting time and we don’t know where it’s going,” she says.