Socially Assistive Robots, Part 2
21:32 Share

From the JAMA Network, this is the JAMA Medical News Podcast, discussing timely topics in clinical medicine, biomedical sciences, public health, and health policy, featured in the Medical News section of JAMA. Hi, very nice to meet you. My name is Blossom. Welcome to my home, the USC Interaction Lab. Here, my robot friends and I want to try our best to help people in the world. Can I ask what's your name?

About 25 years ago, Dr. Maya Matarik and colleagues at the University of Southern California decided to merge the field of assistive robotics with social robotics. Since then, Dr. Matarik and other researchers have created socially assistive robots for a variety of healthcare applications.

This includes helping children with autism learn social and cognitive skills, encouraging patients who have experienced a stroke adhere to their exercise routines, and supporting people with anxiety and depression. Today, Dr. Matarik is the Patrick Soon-Chung Chaired and Distinguished Professor of Computer Science, with appointments in neuroscience and pediatrics at USC, and co-director of the USC Robotics Research Lab.

Advancements in artificial intelligence and natural language processing are helping to move the field she pioneered forward, allowing conversations between socially assistive robots and humans to feel far more natural. In fact, recent research strides have earned Matarik's team a U.S. National Institutes of Health grant for a larger clinical trial to study how support from robots could make cognitive behavioral therapy more effective for college students with anxiety.

I first interviewed Dr. Matarik in 2017. Back when every word a socially assistive robot spoke still had to be predicted and pre-programmed. I'm Jennifer Abassi. I recently caught up with Dr. Matarik, who is also principal scientist at Google DeepMind. We discussed the promise of socially intelligent robots, like the 3D-printed Blossom robot you just heard.

Dr. Matarik, thank you so much for joining me today. The last time we spoke was in 2017, and actually it was you, me, and the same producer, Daniel Morrow. So it's a pleasure getting back together now, eight years later, it's hard to believe. It's wonderful to be back. Thank you so much for this opportunity to continue the conversation. So when we spoke in 2017, the field of socially assistive robotics that you pioneered was 15 years old.

Now it's almost 25 years old. Where did things sort of stand then? And how far have they come now? What has happened since is just, I think, a huge acceleration along two dimensions. So one dimension is AI, right? So AI has exploded. It is fundamentally enabling to socially assistive robotics, as it is to so many things. The other dimension is that with mounting evidence of AI

efficacy in smaller studies. Now there is more appetite and interest to support larger studies. So for example, now we actually have a randomized controlled trial supported by NIH to really validate the efficacy in this particular case in a study involving socially-assisted robotics for anxiety, but there are other studies as well. So this is really encouraging along both of these dimensions, and I think there's a lot of room to grow.

So let's talk about AI a little bit. I imagine that generative AI has had a huge impact on your research or will have a huge impact on your work. Tell us about that. Let's think back to 2017 when we last talked.

Back then, we had socially-assisted robots that could talk to people, and people would talk back, and there would be nice dialogue, except that the only way that we could really make those robots talk and understand was by having what we called a dialogue tree. So literally, think of it as like...

Every single thing that the robot said had to be pre-programmed in advance. So you had to think about what could the user possibly say and how might the robot respond and how can it have various options on that response that wouldn't get boring and yet were consistent. And it was just, oh, it was incredibly difficult. The field of natural language processing, NLP, as we all call it, it was not in its infancy by any means, but it's gone through the status quo now. Interesting.

If you think about the fundamental goal of socially-assisted robotics, it is to engage the user, understand the user, and support the user. And so how can you understand and support and engage the user if everything has to be pre-programmed in a sense? It's not that there wasn't machine learning. There was, but dialogue was very fixed. And then now that we have large language models, suddenly robots can talk

fluidly, generatively, non-repetitively, engagingly. We can prompt them. Before we thought, "Oh, if we knew that the user likes baseball, then we can look up some stuff on the web about baseball, and then maybe we can weave it into the conversation."

Now, you know, all the user has to say is like, oh, I like baseball. And the model will take it away and talk about all kinds of things. So dialogue fundamentally has been put in a place that we never anticipated would happen so quickly. And that's really important for socially assistive robotics because to the extent that our robots talk, if they talk, people immediately assume that they understand, that they're intelligent, that they can offer a lot.

And that was hard before, and now it's not. Give us a sense of some of the exciting research you've been conducting and what type of robots you use for that research. I remember we spoke about the Squash and Stretch back in 2017.

We have leaned in completely in that direction because our goal, and I think this is shared by most people in the field of socially assistive robotics, is to make the technology accessible. And so there's little point in developing a robot that is maybe highly sophisticated and humanoid when it's extremely expensive.

and/or potentially dangerous, which humanoid robot will be inherently dangerous. So we're interested, because we're in the space of health and wellness, we're really interested in robots that will be both very low cost and also inherently safe. And that is what these small, engaging, kind of squash and stretch, almost like animation style robots are highly effective.

And so we've decided to just push in that direction. So we designed these low-cost robots based on an initial design from Cornell University, and they're called Blossoms. And they're really low cost. They're under $500, including the computer and everything. By computer, I mean like a simple computer that drives the actual robot's behavior. Then everything else is done, of course, in the cloud with language models.

The Blossom, I want to tell people what it looks like. So to me, it sort of looks like a sock puppet.

Well, it depends. It has different iterations, but you're absolutely right that what Blossom is fundamentally is a body that is made out of 3D printed components. So you can just print them. They're plastic. You can print them on any 3D printer. And then, you know, there's some basically not even springs, but almost like hair bands that are elastic and motors that drive it. And so then you get to put a skin over it.

The exterior can be made out of cloth. It could be made out of anything that can have enough movement. And we actually found that crocheting or knitting provides the best ability to create something that is soft, that can have shape.

So you can, you know, have a head and a neck and a body. And also it's very personalized. So you could put buttons on it. You can put, you know, we're now actually exploring having fiber optic lights so that robot can communicate through light. And also it can do things like blush or it can light up where you touch it. So there's so many capabilities that you could do that are low cost that work with that kind of an exterior. That's really important because what's the alternative?

Pretty much all robots are made out of metal or plastic. Okay, so that's not warm. That is not soft. That is not easily adjustable and personalizable. And it's very different from what we're after. We're after something that you'll feel it's sort of warm and there for you and you enjoy touching it, although touching it may not be the basic interaction, right? If you're doing, if it's helping you do mindfulness and breathing, you can put your hands on it and it can breathe.

It's hard to imagine how you could enjoy that from a plastic or a metal robot. Yeah, they don't look the way people probably imagine when they hear the word robot. These are not humanoid-looking robots. They don't look like Rosie from the Jetsons either.

But I will say something about humanoids. There's a huge push right now in terms of both industry and academic research into humanoids because of the age of large multimodal foundation models. But I don't think that's the right direction for these kinds of robots that we're talking about, at least not in the near term. So what we want is we want these machines that will be low cost and soft and safe and

and also engaging. And for that, you don't need it to be a humanoid. You don't want it to be a humanoid. It just, it needs to be lifelike. And lifelike doesn't even have to be zoomorphic, right? It doesn't even have to look like an animal. One of the studies that you are conducting is on using large language model powered socially assistive robots to deliver a cognitive behavioral therapy.

So that's so interesting. Can you tell us about how that works and what you've learned? Oh, I would love to talk about this. This is our most exciting development, and it is very exciting for the following reasons. So we got motivated back when language models first came out to explore this question that a very large number of people were talking about how transgressions

Chat bots, now powered by language models, are going to democratize mental health support because the idea was that now we would have apps that are using AI to talk to people about anxiety and depression. And because it is generally not as easy to get a therapist and there could be a long wait or there might not be enough insurance coverage, etc. And that combined with the rates of anxiety and depression.

created a real need. And so people optimistically were saying, oh, well, these AI-powered chatbots are going to solve mental health. And this is still very much a belief that people optimistically are holding. And so we thought, well, we already know that socially-assisted robotics are really effective at engaging people and making them feel good. So we thought, let's run a study in which we compare a chatbot that's powered by an LLM

and then using that same LLM to power a socially-assisted robots.

Cognitive behavioral therapy is a really effective way to deal with anxiety. It's a really effective way to mitigate a lot of challenges, including insomnia, weight loss, anxiety. It's just very broadly used. So we wanted to look at it specifically for anxiety. Now, one of the issues with cognitive behavioral therapy is that it takes practice. Basically, it requires you to reframe your thoughts, to think about, you know,

Is there a different way to think about this? What might I do about it? So there's a set of sort of prescribed ways that you need to go through it and you need to reframe how you think. And it's usually done by writing it down in a notebook. So as you could guess, people don't like to do this because nobody likes to practice. So people don't do it. So cognitive behavioral therapy, effective, but people don't do it either because it's not accessible or because they don't, even when it is accessible, they don't do all the steps because it's really hard. It's emotionally hard. It takes time, et cetera.

So we thought, oh, maybe the robot will just make people better at adhering to the regimen. And so we ran the study in the college dorms at USC, University of Southern California, our home institution that has been a great place to do all of this research over the last 25 years.

And it was really interesting what we found. It's been the driver for what we've been doing since. So we found that both the robot and the chatbot was engaging for college students. They interacted with both of them. This was a two-week study, so a short study, short, small study. We didn't find that over the two-week period, they particularly spent more time with the robot than the chatbot. But, and this is a big but,

When we tested their psychiatric distress, and for this we used a proper assessment, a clinical assessment, before and after the sessions, we found that only the students who interacted with the robot had a significant reduction in their psychiatric distress. The students who were interacting with the chatbot, they liked it, but their psychiatric distress was not lowered. So there was no clinical benefit.

Then the next thing we did was we actually looked at the transcripts of what the students were saying to the language model and what the language model was saying back. And we use these measures of alignment. So synchrony and alignment are things that are used to actually evaluate how well a therapist is

is interacting with a patient or a client. So it's a standard measure from psychiatry and psychology, and we applied it to language models. Turns out actually significantly, the alignment with the chatbot was very poor.

It's really interesting that at a surface level, people are very, very willing to engage with language models. I mean, it's wonderful dialogue, right? You feel like you're talking to a human. It's always available. You could prompt it to always be friendly, to have whatever persona you want. But is it really helping? And our preliminary data showed that it was not actually clinically helping.

And so that's why we submitted this as a randomized controlled trial and got support from NIH, from the National Institutes of Health, to actually run this on many, many, many more students and to do it properly. And so we're very excited about having this opportunity. It's a certain step of maturity of the field, and it's just really wonderful to be in this place.

So what is it about the robot that's different from the chatbot? Isn't that the billion-dollar question, right? And so there have been many studies over the last 20 years in the field of human-robot interaction more broadly that have compared screen-based agents, like virtual humans or just even things like animation, with physical robots. And there is hands-down evidence

a ton of literature that shows that interactions in the physical world with a physically embodied creature, like a robot, are much more effective.

And by effective, I mean there are many studies that have shown more engagement, more enjoyment, statistically significantly more enjoyment, more engagement, more learning, learning retention over longer periods of time. So all of the measures that you would be interested in are significantly better when we're interacting with a physically co-present environment.

creature, if you will, as opposed to a screen-based one. And so why would that be? Well, you know, fundamentally everyone would say that that is because we are social creatures and we have evolved to be in the physical world with other social creatures. And, you know, it works with dogs, it works with pets, and it works with humans.

And we attribute a different level of agency and engagement to physically embodied agents than we do to screen-based agents or even chatbots. And there's another line of evidence along this, which is that as we observe more social isolation, not just due to the pandemic, but in general, as people are spending less time physically together, and we see this in younger generations in particular, as they're spending more time on their devices and interacting via devices, but not physically co-present,

It isn't that they don't have social connections, but they feel more isolated. So you get more loneliness, you'd get more isolation. And it is hypothesized that all these effects are due to the fact that this is a very different way of interacting than how we're evolved to interact. And so while we can do it, we're very adaptive. Humans are extremely adaptive.

But it doesn't mean that what we're adapting to is actually good for us. And if we can add to people's social connections a little low-cost robot that bridges the gap until they're talking to a real human in the real world in a way that might be helpful, well, then that's worth exploring. I've thought about our conversation a lot over the years, and in particular, I thought about it during the pandemic, the COVID pandemic.

I wrote stories that were really challenging about the older adult population in nursing homes. And at the time of lockdowns, the difficulties they were facing in being separated from people, separated from their families and their caregivers. And medical directors of nursing homes told me that people were dying because of this.

Because of the isolation. And I thought about how, man, if we had those robots, those little robots, maybe that would have helped. I'm sure you've thought about that at the time. So do you think that would have made a difference?

I do, of course. I'm heavily biased, but I just know it. And it did break my heart while it was happening as well, because so much good can be done with some small interventions. I can't believe how much of a lost opportunity there was with the elderly, because they really, really missed out. The elderly were just alone. And just the health price tag of loneliness is huge. And it's good that it's finally being recognized.

So I don't want to be understood as saying that I think we'll just give robots to old people and then it'll be fine and we can forget about them because that is not at all the idea. People need to be with other people. But it's remarkable what people have shown in the field of research. And this is not my work. So I really, really, I think it's important to give credit to all of the brave work that people have been doing where, for example, they've taken robots into nursing homes and in particular into memory care units where, you know,

The residents there, they have various forms of dementia, often very debilitating. And yet, introducing something that is just engaging and interesting enough will end up being a social catalyst. So you can't fix dementia, but you can at least improve the quality of life and possibly longevity.

Where do you see things going? Take us maybe five years out. When will we have affordable tabletop robots in the home that can keep people company and bridge that time gap between when they can see people? I'm so glad you're asking me this now. And then let's talk again in five years, because right now we're at a really interesting point. We have the science. We know that this is important and effective for human health.

We have the technology because not only do we now have better materials and people are discovering soft, like soft robotics is a whole new field that's now been really emerging. And we have, of course, these language models and foundation models. So we have all the technology, but we don't have investor interest. By we, I mean the field and humanity necessarily.

more significantly. Because pretty much all of the investors currently are investing into humanoids. These like, you know, human-like, human-sized, highly dangerous, not ready for prime time robots. So much can happen, but will it? Now, it is also possible, and my fingers are crossed for this,

that there might be one or two brave companies that will do something that will show such benefits on the market. And I say this as someone who has no startup and has no vested interest in this personally, but I do think it's a loss because, you know, what about the next pandemic? What about, you know, the next reason why we need so much support? What about the aging population?

You know, what about infirmities? What about the loneliness epidemic? We could be addressing all of this. I'm not saying fixing it. I'm not being naive and Pollyannish about it. But I think we can have a positive impact. Thank you so much, Dr. Macharik. I hope we can make a date to talk again, maybe in five years, and catch up and see where we are then. You're on. I look forward to it. Thank you.

That's it for this episode of JAMA Medical News. I'm Jennifer Abassi. Thanks to my guest, Dr. Maya Matarik. We'll also include a link to a video of the Blossom robot and Dr. Matarik in their lab. You can find that link in the show notes. This episode was produced by Daniel Morrow at the JAMA Network. To follow this and other JAMA Network podcasts, please visit us online at jamanetworkaudio.com. Thanks for listening and join us again next time.

This content is protected by copyright by the American Medical Association with all rights reserved, including those for text and data mining, AI training, and similar technologies.