The Life Scientific: Helen Hastie

0:00This BBC podcast is supported by ads outside the UK.

0:30at Whole Foods Market. Apple Vacations, where your story starts. The July Savings Event from Apple Vacations is here. Book by July 23rd and save up to$400 off your flight and hotel package to Mexico, the Caribbean, Central America, Hawaii, or Europe. From all-inclusive escapes to bucket list adventures, Apple Vacations makes it easy to create memories with the people who matter most. Start today at applevacations.com or contact your trusted travel advisor. applications where your story starts

1:03hello let's start with a quick test if i ask you to conjure up an image of a useful and reliable robot helper what springs to mind some of you might be thinking of those industrial robotic arms used in car assembly lines or perhaps an explosive ordnance bot the remote controlled devices used by bomb disposal teams. Others might be focusing on fictional robots, R2-D2 from Star Wars perhaps. Not just a tool, but a machine that can actually interact with humans. One that is communicative, responsive and even trustworthy. In fact, those sorts of relationships are the focus for today's guest. Helen Hastie is a professor of human-robot interaction and head of the School of Informatics at the University of Edinburgh.

1:50Her mission is to develop robots that don't just think, but connect. Systems that will competently perform their tasks, but which can also hold a natural conversation and explain themselves clearly to their human colleagues. Although, as we'll hear, that's not quite as straightforward as Star Wars made it seem. Helen's career has taken her from developing early dialogue systems, the ancestors of today's generative AI, to working on sophisticated bots that can serve coffee with a side of small talk, teach struggling kids with empathy, or provide calm and confident decisions as triage nurses. She's also driven some of the UK's flagship robotics initiatives, not least as co-lead of the National Robotarium.

2:34Ultimately, her ethos is robotics for all, shaping a future where robots aren't just in our world, they're part of it. Ideally, a welcome one. Professor Helen Hastie, welcome to The Life Scientific.

2:46Helen Hastie:Thanks for having me. Now, Helen, a big part of your work is getting people to trust and welcome robots in everyday settings. What would you say is the biggest misconception people have that might put them off having a robot colleague? So we see in the media and in sci-fi movies these robots that are highly functioning, that can do many, many jobs and are social and cognitively quite intelligent. Now, we are far from that. We do have some robots that can perform very, very well on specific tasks, but these tend to be very narrow tasks. So we are quite a long way from these all-functioning, intelligent robots.

3:25I suppose a lot of us hear the word robot and do think of those humanoid versions from sci-fi movies that can do everything, including take over the world, of course. But I gather there are actually quite a few issues with making robots in human form. Sure.

3:41Helen Hastie:So robots that generally look like humans, not necessarily those with legs, but in some kind of human shape, can be really useful for certain settings. but we have to be really careful. There's a phenomenon called the uncanny valley and as robots become closer to looking like humans, particularly in terms of detailed looks, this can actually be quite unsettling. So it's good that robots look a bit like us but not too much like us. Okay, I mean so is there a sweet spot of how human a robot should look if it's going to be accepted? Yes, most definitely and it's very dependent on the task. So robots in the home, Maybe it's good they look a bit like us.

4:21Helen Hastie:But robots working in factories or warehouses or stacking shelves, it's actually better if they look like they know what they're doing in terms of the function that they're designed to do. So having wheels, for example, rather than legs. Exactly, yes. Thinking about trustworthiness in robots does make me wonder about the current concerns around over-reliance on large language models, like ChatGPT, for example, having influence over vulnerable youngsters or isolated individuals. Is there an element of risk to making robots more trustworthy, as in the more we trust them, the greater the risk of manipulation?

4:57Helen Hastie:So it's very important that we instill the appropriate amount of trust. So it's important not to overtrust robots. And it's also important not to undertrust them. For example, a robot that's used in surgery that is highly accurate, maybe more so than a human. It's really important that we do adopt these robots if they are capable of performing properly. So in some cases, we should trust them because they've been designed to do something better than we can. But it's also important not to overtrust. So you shouldn't overtrust an autopilot in a ship, for example. It's important that you always have the human eye on it.

5:32Yes, indeed. So, Helen Hastie, do you remember where this fascination with AI and robots first started for you? Were you particularly into tech or sci-fi as a kid?

5:42Helen Hastie:Well, I grew up in a house that had early computers. My father was an early adopter. We had Commodore PC, for example, in the house. And he was always putting together computers. So I think from an early age, I was surrounded by tech. You grew up in Aberystwyth in Wales. Tell me a bit about your family. My father was a lecturer in international relations and my mother in classics. And I'm the youngest of four. Most of us have PhDs in the family, to put it like that. But across a wide range of disciplines, clearly. I gather you felt rather lost at school because you love both science and languages.

6:21And it didn't feel like there was a way to follow both paths. But then a light bulb moment. Tell me about discovering linguistics.

6:29Helen Hastie:So when you're 16, you have to decide what three A levels you're going to do. And I wanted to do all the languages and all the sciences. And I think my father could see a bit of a conflict inside me. So he bought me this encyclopedia of language. This really caught my attention. And linguistics being the science of language really kind of found a sweet spot with me. So off you went to study AI and linguistics at Edinburgh, followed by a master's in computational linguistics at Georgetown University in the United States. What drew you to America? So I'm a half American, so I wanted to study out there.

7:06Helen Hastie:and I was given a fellowship for, believe this or not, Welsh female studying science at master's level in America. Excellent. Very niche. I'm not sure there was many of us. So I got that fellowship and off I went to Georgetown. Right, right. You returned to Edinburgh to do your PhD in speech and language technologies. Then you decided to take a year out. I married Stuart in the year 2000 and we took a year-long honeymoon, which we were very fortunate to do, travelling around the world, which included going overland in a truck across Africa and then Asia and Australia. And then we stopped off in the States.

7:43Helen Hastie:Having a degree in technology, I thought I'd be quite quick to get a job. But unfortunately, that was the time of the dot-com bubble burst. So everything was basically shuttered up on the west coast of the US. So then we moved over to the east coast. And that's when I got a research position at AT &T Research Labs. One of your biggest projects there was called How May I Help You? One of the first truly interactive voice response systems. First of all, what was it designed to do? So IVRs, as we call them, were pretty terrible back then. Basically, most of them would allow you to say things like say one for billing, say two for technical support.

8:23Helen Hastie:So not a very rich interaction, not much more benefit than pressing the button on the phone. So how may I help you was different in that it was an open question. And so the system had to understand a whole range of different responses and then channel the customer to the appropriate department. What I was doing was I was working with Marilyn Walker and we were looking at how we evaluate these systems. I mean, that advance was quite revolutionary at the time. An ancestor, if you like, to the voice assistants we use today, you know, like Siri and Alexa and so on. but doing this 25 years ago I mean what sort of difficulties did you run into?

9:02Helen Hastie:So we called them spoken dialogue systems and we would divide them into different parts so you would have the speech recognizer understanding the words and then you need to convert those words into meaning and decide what to say next and then what the words would be and then turn that into a computer voice. Nowadays it's all one big end-to-end system and it's trained on much data, large compute. So it's tackled in a different way. But back then, we had to divide into each component and each component had their own challenges. And I mean, we should say this sort of voice assistant technology is quite different from today's all singing, all dancing, large language models.

9:39Correct. Can you explain how?

9:40Helen Hastie:We didn't have the data back then. We were using language models, but they weren't large. Okay. Well, a couple of years later, Helen, and you joined Lockheed Martin, also in New Jersey, on the East Coast, and you developed an onboard AI assistant for ships called SUSIE. Tell me about that. At Lockheed Martin, it was all about applications and putting these systems where they would be used and adopted. So one of these was on a ship, and it was using the intercom, and it was sensorised. So you could call up SUSIE on the intercom, and it would tell you, for example, the starboard engine temperature.

10:18Helen Hastie:Right. So the important thing here was the equipment was in situ and it may be of use. OK, well, having a named AI assistant like Susie helping you control a vehicle does have slight overtones of hell from 2001 at Space Odyssey. I'm afraid I can't do that, Dave. That's my best hell voice. So do you have any reassuring words for people concerned that we're hurtling towards that sort of AI dystopia? So in critical situations we put in certain guardrails It's important that we test them properly But the technology is advancing so rapidly That we really have to think about this now And we have to think about it from inception of design of the AI Or design of the robot So that everything we do we embed responsible research And we think about what it's going to do when it's actually deployed Right.

11:13Well, Helen, in 2007, you and your husband Stuart decided it was time to come back to Scotland. You get a research fellowship at the University of Edinburgh, followed by one at nearby Heriot-Watt University, where in 2011, you were asked to head up a major EU project called Parlance. What was that about?

11:34Helen Hastie:Parlance was a project with the University of Cambridge, Geneva, Yahoo. and what we were looking at was a voice search application. So you would talk to an intelligent agent, a bit like you might do with a well-informed colleague, and you would have a discussion and get the information through conversation rather than just typing in words in the search box. And my particular area was looking at what the computer should say and how it should say it. It was a system for giving restaurant recommendations. conversations so we wanted to inform the user but not overload them with too much information so this is more along the lines of agentic AI that people are talking about now so this is AI that will do actions in the real world like make a booking for you and of course essentially this was the early days of generative AI so just to clarify this is AI that can scan whatever data it has access to and create new content from it it must have been pretty exciting to be at the forefront of this shift in computer science?

12:38Helen Hastie:It's been fascinating. It's changing so rapidly that it's quite hard to keep up. And universities, I think we have to find a way to advance the technology and the science within our means. And so we're looking at things like how we can train models on less data more efficiently. So there are areas that we can research, but the adoption of AI is just, I've just been blown away. Is there a concern that academia, the researchers who aren't working in the field for commercial gain, for example, are losing the race against, you know, the big players, you know, the Googles and the Microsofts and so on?

13:17Helen Hastie:So there are different research directions that we can go in that are going to be important for the future. It's true we don't have access to hundreds of millions of pounds for training really large models running on very large systems. But we do, for example, at Edinburgh University, We have the UK's national computer and we do have access to data. But what's going to be important going forward is changing the paradigm of AI. So, for example, learning how to learn rather than just throwing more data, more compute at these models. Right, right. Now we come to the point in your career, Helen, where your focus shifts from dialogue systems to human robot interactions.

14:00You were made the principal investigator on a massive project involving several UK universities looking at trust in autonomous systems. We spoke briefly about trustworthy machines earlier. Why is this so important?

14:14Helen Hastie:There's a general definition of trust being the willingness to be vulnerable to another entity's actions. And this is particularly important for robots, increasing so if you can't actually see the robot. So an example of this is underwater robotics. They're very expensive robots. They go very deep. They don't communicate very well. So this idea that you have to trust the robot to be autonomous and to complete its task is just an example of why human robot trust is really important. Right. So we looked at the relationship between humans and trust, and we investigated whether certain aspects of human-human interaction can map over to human-robot trust.

14:59At the time, Helen, you were at Heriot-Watt, so I really have to tell you this story. I made a documentary for the BBC a few years ago called The Joy of AI, where I visited Heriot-Watt. I was introduced to this chatbot called Alana, one of the leading versions, leading chatbots at the time. But the conversation wasn't always free flowing. This is we're talking about seven years ago or something like that. And I told Alana one of my favourite sci-fi authors was Philip K. Dick. And she told me off for being inappropriate. So there were those sorts of issues. I guess the conversational side has come a long way since then.

15:36Helen Hastie:Most definitely. and also with these large language models what they are good at is failing gracefully so rather than computer says no they are able to respond in an appropriate manner. Well Helen during your time at Heriot-Watt you were involved with developing and trialling various projects we'll come to them in a moment but first let's talk about how robots are put through their paces. In 2020 you became academic co-lead of the National Robotarium, which is a joint initiative launched between Harriet Watt and the University of Edinburgh. Tell me about that facility. So the Robotarium looks at translating research into real world applications.

16:17Helen Hastie:And to do that, you really have to have a testing environment. So there's part for autonomous robotics, large hangar for flying drones, testing autonomous vehicles. And then there's a living lab, which is essentially a flat that's censored up where we can put robots and invite people to come and interact with them in a kind of more natural, comfortable setting. It does sound to me a bit like a spy kit testing facility, like Q's lab in the Bond movies. But the first test I want to ask about is one you conducted in a real world environment. Tell me about your robo barista. So what we wanted to do is put a robot in the wild.

16:58Helen Hastie:So this is where the field is really trying to push forward, moving away from very controlled settings to just leaving the robot and seeing how people get on with it. So we installed a robot in our common room at the university. We attached it to a very good Italian coffee machine so that people would be motivated to interact and work with it. And what we were looking to investigate was whether people's attitudes to robots would change over time. What did this robot look like then? It's essentially a head and it has a back-propagated light with a face that can have various facial expressions and has a computer voice.

17:37And what does it do? I mean, it must have arms, presumably, if it's going to make you coffee.

17:40Helen Hastie:Well, no, it was attached to a coffee machine, but it would take your order. Oh, I see. It would take your order, it would chat to you. But these robots are really good at trying to understand social cues and how people react to robots' social cues because it does have quite an expressive facial expression. But it wasn't just the facial expressions. It could also engage in small talk. In fact, I think we have a short clip of your robot barista interacting with a customer. Hi, I'm Alex. What can I get to you today, buddy? Can I have an espresso, please? Can you please confirm that you would like me to make you espresso?

18:14Yes, please. My superpower can detect when someone needs an extra coffee shot. Would you like me to try it? No. Let me just brew the drink for you. I'll be back in just a minute. So what's going on inside that robot head when it's interacting with people? How does it decide on the best opening gambit or a suitable response for a particular customer?

18:38Helen Hastie:So each customer had a loyalty card, so that was a way that we could keep track of them. Very good. And we would change the interaction periodically so that they would feel like it was more natural, varied, like a human barista. So what did you learn from that trial? We measured people's attitude across six weeks and luckily we didn't put them off. So there's this scale called the negative attitude to robots. So people, their attitudes basically didn't change. There is something called a novelty effect in robotics where people, first of all, get really excited and use robots and then they kind of go off it and they can't be bothered afterwards.

19:12Helen Hastie:afterwards. But we didn't see that, which is good. If the coffee's good. I was going to say that might have been the fact. And also trustworthiness and likability and usability are closely linked. Moving on to some quite different contexts, you also worked on an empathic robot tutor and a robot triage nurse, both of which I imagine involve quite different design requirements. Yes, our empathic robotic tutor was very complicated. It could understand and sense a student's frustration and it changed the way it taught based on how the student was feeling or how it perceived the student was doing. We found that the empathic robot created more motivated learners.

19:58And you say, you know, it could sense whether a student is frustrated or not. Something to do with picking up the tone of voice or the volume if a student starts shouting at it.

20:08Helen Hastie:This is mostly the visual indicators of kind of frustration and also the volume of their speech as well. I gather you trialled this robot tutor at home on your own kids. Yes, my kids did tend to be guinea pigs, willing guinea pigs, I would like to say. So they would pilot the systems quite frequently. They didn't mind, but for this particular application, it did mean that they had to learn quite a bit more geography, which they weren't really keen on. They were doing this as a favour for mum rather than because they thought it might be useful to them. Correct. And tell me about the triage nurse robot.

20:40Helen Hastie:This was an experiment run during COVID, so we weren't actually in situ running the experiment. But we were looking at a robot that would be, for example, at the reception of a hospital and then would ask about symptoms and then tell people whether to go home or to take a seat. And we found that transparency was really important. So the robot had to explain the reasoning behind its decision and that was more accepting and trusting to the participants. And these three very different robots, you know, the coffee maker, the tutor, the nurse, are great examples of how different designs and interfaces will affect how people interact with and even trust them.

21:20For instance, the movie robots I mentioned earlier, cute sidekicks like WALL-E or R2-D2, you wouldn't necessarily feel confident about those slightly rather hapless robots handing out drugs in a hospital, for example.