TECHNO WORLD

Kids learn better with a friend. They're more enthusiastic and understand more if they dig into a subject with a companion. But what if that companion is artificial?

Researchers at the University of Wisconsin-Madison have built a robot, named Minnie, to serve as a reading buddy to middle school kids, and Minnie's new friends grew more excited about books and more attached to the robot over two weeks of reading together.

"After one interaction, the kids were generally telling us that, sure, it was nice to have someone to read with," says Joseph Michaelis, a UW-Madison graduate student studying educational psychology. "But by the end of two weeks, they're talking about how the robot was funny and silly and afraid, and how they'd come home looking forward to seeing it again."

Michaelis and computer sciences professor Bilge Mutlu published their work with Minnie on Wednesday (Aug. 22) in the journal Science Robotics.

Research shows thatsocoal learing—pairing up with a peer to complete math problems or read a chapter in a textbook—is a powerful way to help students develop skills and interests, according to Michaelis. It improves students' comprehension by distributing the cognitive workload and reinforcing understanding through dialogue.

"Most interesting to me is that we know social learning strengthens interest and motivation," says Michaelis, who taught high school science before returning to graduate school. "A lot of kids who don't like reading, in particular, point out that it's an isolated activity, and people just sort of accept that isolation. But it can be demotivating and harder to learn and understand in that situation."

Michaelis and Mutlu believe companion robots will soon be a fixture in homes, and they wondered if those robots could serve as social learning companions for kids.

They designed a two-week reading program including 25 books representing a range of reading skill and story complexity, and programmed Minnie to be an interested listener. The children in the study read aloud to the robot, which could track their progress in the book and react to the story—every few pages or so, especially during important moments in the plot—with one of hundreds of preprogrammed comments

The goal is to try to make it as genuinely conversational as possible. If you were reading a book to me, and I was surprised, I'd say something like, "Wow, I didn't see that coming!" Michaelis says. "When a scary part of the book happens, the robot says, 'Oh, wow, I'm really scared.' It reacts like it would if it had a real personality."

With simple, oversized black eyes on a relatively featureless white globe of a head, Minnie can react and cajole and summarize and appear thoughtful. It even starts by following an initial introductory read by recommending (via an algorithm) a good book from its library of titles, which include the familiar Harry Potter and Goosebumps series and genre classics like "A Wrinkle in Time."

"That match is crucial. If you're trying to make a social connection with someone, and they say, 'You'd love this book,' and they're totally wrong? That ruins that credibility," Michaelis says. "Most kids said the robot did a good job suggesting books to them."

The connection grew from there. The number of children who told researchers the robot has a personality or emotions increased more than fourfold over the two weeks readers spent with Minnie. The number reporting they were motivated to read also spiked—and surpassed a control group following a paper-based version of the reading program. And kids who read with Minnie said they felt like they understood and remembered more about the shared books.

Mutlu, who studies human-robot interaction, said the illusion of personality in a robot is so fragile, it often breaks down in a matter of minutes.

"We're getting more used to interacting with things that talk to us now, with Siri and Alexa and others. When they are behaving in a way that kind of reveals their technical flaws or limitations, you realize immediately you're not having an actual conversation," Mutlu says. "The biggest shock in our study is that two weeks later, the kids are still relating to the character—rather than saying, 'This is stupid. I'm not talking to your robot anymore.'"

He attributes that success to the care taken in developing Minnie's speech repertoire, and also to a sweet spot robots enjoy among the age group in the study—a group that may be emotionally sophisticated enough to make connections, but not wary or snarky enough to push the envelope.

"The children are a match to the robot's capabilities, and vice versa," Mutlu says. "This robot supports an engaging reading activity, but it's not a social companion to the extent that you could have an open conversation with it. If you had a much more capable robot, that picture might change."

Not that other family members were left out. Minnie's homestays got to be comfy visits, with the robot and humans sprawled on the floor or propped up on beds with pillows.

"I had families send me pictures of themselves dressing up the robot. I'd get the box back, and it would have sorts of extra books in it that were just from the kids' shelves, just thrown in there. All kinds of random paper, markers, crayons," Michaelis says. "Your experiments don't get to live like that—to get that authenticity—in a lab."

Michaelis and Mutlu, whose work is supported by the National Science Foundation, see Minnie helping to spur otherwise reluctant students on in all sorts of academic tasks, and have already begun testing a version of the robot that shares in science studies. They hope to try out periods of interaction even longer than two weeks—an eternity for a robot interaction, but a blink of an eye in the development of a child.

They expect social learning functions could be part of a companion robot shared by a whole family, but acknowledge that future as fraught with design concerns over how to craft an engaging personality for users at different ages, protect family members' privacy and maintain the hard-won trust of a child that will also see therobot having private interactions with parents.

"This idea is in its infancy. But now we know if you really, carefully design this, it can actually sustain interaction and heighten kids' emotional experience with reading," Mutlu says. "That's a huge achievement."

Showing prejudice towards others does not require a high level of cognitive ability and could easily be exhibited by artificially intelligent machines, new research has suggested

Computer science and psychology experts from Cardiff University and MIT have shown that groups of autonomous machines could demonstrate prejudice by simply identifying, copying and learning this behaviour from one another.

It may seem that prejudice is a human-specific phenomenon that requires human cognition to form an opinion of, or to stereotype, a certain person or group.

Though some types of computer algorithms have already exhibited prejudice, such as racism and sexism, based on learning from public records and other data generated by humans, this new work demonstrates the possibility of AI evolving prejudicial groups on their own

The new findings, which have been published in the journal Scientific Reports, are based on computer simulations of how similarly prejudiced individuals, or virtual agents, can form a group and interact with each other.

In a game of give and take, each individual makes a decision as to whether they donate to somebody inside of their own group or in a different group, based on an individual's reputation as well as their own donating strategy, which includes their levels of prejudice towards outsiders.

As the game unfolds and a supercomputer racks up thousands of simulations, each individual begins to learn new strategies by copying others either within their own group or the entire . population

Co-author of the study Professor Roger Whitaker, from Cardiff University's Crime and Security Research Institute and the School of Computer Science and Informatics, said: "By running these simulations thousands and thousands of times over, we begin to get an understanding of how prejudice evolves and the conditions that promote or impede it.

"Our simulations show that prejudice is a powerful force of nature and through evolution, it can easily become incentivised in virtual populations, to the detriment of wider connectivity with others. Protection from prejudicial groups can inadvertently lead to individuals forming further prejudicial groups, resulting in a fractured population. Such widespread prejudice is hard to reverse."

The findings involve individuals updating their prejudice levels by preferentially copying those that gain a higher short term payoff, meaning that these decisions do not necessarily require advanced cognitive abilities.

"It is feasible that automachnical machine with the ability to identify with discrimination and copy others could in future be susceptible to prejudicial phenomena that we see in the human population," Professor Whitaker continued.

"Many of the AI developments that we are seeing involve autonomy and self-control, meaning that the behaviour of devices is also influenced by others around them. Vehicles and the Internet of Things are two recent examples. Our study gives a theoretical insight where simulated agents periodically call upon others for some kind of resource."

A further interesting finding from the study was that under particular conditions, which include more distinct subpopulations being present within a population, it was more difficult for prejudice to take hold.

"With a greater number of subpopulations, alliances of non-prejudicial groups can cooperate without being exploited. This also diminishes their status as a minority, reducing the susceptibility to prejudice taking hold. However, this also requires circumstances where agents have a higher disposition towards interacting outside of their group," Professor Whitaker concluded

Humans have long been masters of dexterity, a skill that can largely be credited to the help of our eyes. Robots, meanwhile, are still catching up. Certainly there's been some progress: for decades robots in controlled environments like assembly lines have been able to pick up the same object over and over again

More recently, breakthroughs in computer vision have enabled robots to make basic distinctions between objects, but even then, they don't truly understand objects' shapes, so there's little they can do after a quick pick-up.

In a new paper, researchers from MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL), say that they've made a key development in this area of work: a system that lets robots inspect random objects, and visually understand them enough to accomplish specific tasks without ever having seen them before.

The system, dubbed "Dense Object Nets" (DON), looks at objects as collections of points that serve as "visual roadmaps" of sorts. This approach lets robots better understand and manipulate items, and, most importantly, allows them to even pick up a specific object among a clutter of similar objects—a valuable skill for the kinds of machines that companies like Amazon and Walmart use in their warehouses.

For example, someone might use DON to get a robot to grab onto a specific spot on an object—say, the tongue of a shoe. From that, it can look at a shoe it has never seen before, and successfully grab its tongue.

"Many approaches to manipulation can't identify specific parts of an object across the many orientations that object may encounter," says Ph.D. student Lucas Manuelli, who wrote a new paper about the system with lead author and fellow Ph.D. student Pete Florence, alongside MIT professor Russ Tedrake. "For example, existing algorithms would be unable to grasp a mug by its handle, especially if the mug could be in multiple orientations, like upright, or on its side."

The team views potential applications not just in manufacturing settings, but also in homes. Imagine giving the system an image of a tidy house, and letting it clean while you're at work, or using an image of dishes so that the system puts your plates away while you're on vacation.

What's also noteworthy is that none of the data was actually labeled by humans; rather, the system is "self-supervised," so it doesn't require any human annotations

Two common approaches to robot grasping involve either task-specific learning, or creating a general grasping algorithm. These techniques both have obstacles: task-specific methods are difficult to generalize to other tasks, and general grasping doesn't get specific enough to deal with the nuances of particular tasks, like putting objects in specific spots.

The DON system, however, essentially creates a series of coordinates on a given object, which serve as a kind of "visual roadmap" of the objects, to give the robot a better understanding of what it needs to grasp, and where.

The team trained the system to look at objects as a series of points that make up a larger coordinate system. It can then map different points together to visualize an object's 3-D shape, similar to how panoramic photos are stitched together from multiple photos. After training, if a person specifies a point on a object, the robot can take a photo of that object, and identify and match points to be able to then pick up the object at that specified point.

This is different from systems like UC-Berkeley's DexNet, which can grasp many different items, but can't satisfy a specific request. Imagine an infant at 18-months old, who doesn't understand which toy you want it to play with but can still grab lots of items, versus a four-year old who can respond to "go grab your truck by the red end of it."

In one set of tests done on a soft caterpillar toy, a Kuka robotic arm powered by DON could grasp the toy's right ear from a range of different configurations. This showed that, among other things, the system has the ability to distinguish left from right on symmetrical objects.

When testing on a bin of different baseball hats, DON could pick out a specific target hat despite all of the hats having very similar designs—and having never seen pictures of the hats in training data before.

"In factories robots often need complex part feeders to work reliably," says Manuelli. "But a system like this that can understand objects' orientations could just take a picture and be able to grasp and adjust the object accordingly."

In the future, the team hopes to improve the system to a place where it can perform specific tasks with a deeper understanding of the corresponding objects, like learning how to grasp an object and move it with the ultimate goal of say, cleaning a desk.

The team will present their paper on the system next month at the Conference on Robot Learning in Zürich, Switzerland