Automated driving is advancing all the time, but there’s still a critical missing ingredient: trust. Host Jennifer Strong meets engineers building a new language of communication between automated vehicles and their human occupants, a crucial missing piece in the push toward a driverless future.
- Dr. Richard Corey and Dr. Nicholas Giudice, founders of the VEMI Lab at the University of Maine
- Ryan Powell, UX Design & Research at Waymo.
- Rashed Haq, VP of Robotics at Cruise
This episode was reported and produced by Jennifer Strong,Tanya Basu, Emma Cillekens and Tate Ryan-Mosley. We had help from Karen Hao and Benji Rosen. We’re edited by Michael Reilly and Gideon Lichfield.
Full episode transcript:
Jennifer Strong: Driverless cars have been conjured up in films and talked about for decades. The first attempts to make them were about a century ago. And sure, they’re no longer as unbelievable as in “Herbie The Love Bug from the 1960s” where this self driving car has a mind of its own:
[Herbie clip]: I did not bring you here. It’s this nasty little car.
Jennifer Strong: But while the artificial intelligence driving the technology has come leaps and bounds—and we now have some self driving taxis in the U-S and China—the fact is communication between the car and the human inside it hasn’t really come that far since the days of Herbie driving off without it’s driver’s consent.
And what’s more, the beeps and bings meant to tell us something about what the car is doing don’t intuitively tell us what’s going on.
To be fair, that isn't just an issue with automated vehicles—it pops up in most cars with computerized elements. My car pings for my passenger to put on a seatbelt when all that’s there is a handbag.
Despite programmers best efforts there’s still a fundamental problem: cars and people don’t speak the same language.
This all matters because this isn’t a video game or phone trying to interpret its mathematical language of 1s and 0s in ways that make sense in human language. This is a moving thing carrying our bodies from place a to place b. When things go wrong, they can go really wrong.
I’m Jennifer Strong and this episode, we dive into some tricky questions about usability and trust.
Imagine you’re about to leave the house—you’re scrambling to get to a meeting with some clients. You want to talk to your boss beforehand, but you’ve got to also get your kids to school. You and the kiddos struggle out the door and head to the car, coffee and toast in hand, backpacks and water bottles in tow. It would be nice to set an autopilot. But would you trust it? Can you really imagine letting go of the steering wheel?
And how do you know which scenario is less safe? The inevitable multitasking and kid management while driving? Or handing over some control to a machine?
The technology we have now, such as Tesla’s autopilot, isn’t advanced enough to really be autonomous. People in the field refer to it as automated driving—it requires a human to monitor what the car is doing—we know at least five people have died in crashes with automated driving when someone behind the wheel wasn’t paying close attention.
But distracted drivers are dangerous in every kind of car, and the US still sees some 30 to 40 thousand people die in crashes each year.
Jennifer Strong: Taking a walk in my neighborhood, I’ll often see my best friend on her stoop with her thirteen year old child. It’s a typical urban scene... except for this next part.
Hannah: You go into controls... this won’t load…
Jennifer Strong: Hannah is using an app on her mom’s iPhone to pull their electric car into a parking space where they can plug it in.
Hannah: Well, this is a usability thing… it won’t load.
Camelle: That happens multiple times when we first start to deploy the summon feature…
Jennifer Strong: Kids are growing up with apps and it’s often more intuitive to them how these things work than it is for their parents. They also aren’t as easily frustrated by new tech.
Hannah: Here we go. Okay. So, it shows you a picture of the car. You just push whichever way you want to go and hold it, and it should move. But it didn’t… [giggle]...
Jennifer Strong: Car maker Tesla has an autopilot feature that allows drivers to summon cars in parking lots to come pick them up, among a host of other things.
Oh! There it goes.
Hannah: So the lights are turning on and you can see the steering wheel move inside the car… and also it thinks our bike rack is another car, so we can’t reverse it with this.
Jennifer Strong: Autopilot also offers a more advanced version of cruise control… though it’s not perfect either.
Camelle: But I have come to appreciate that it will come to a stop or slow down based on the person in front of you. If you can't go into the passing lane. However, you still have to be alert because it will stop dead on a fast moving highway. For reasons… no reason. Not that I can tell. So, I'm very leery. It doesn't allow me to relax much. Just lets my foot take a break.
Camelle: And when it wants to talk to me about its safety measures I wish it would speak just plainly. Rather than beep beep… trying to scroll through… what the [bleep] does that mean? I don't know. [laughing]
Hannah: When we do software updates, it gives us an outline of what is done. But like my mom said, it's not in plain English. It's very hard to understand. So you really just have to figure out what's changed as you go, which can be dangerous. If it's something major, you don't know how to turn on the heat or turn on autopilot. And also I know as I've started riding in the front seat, I've become kind of the co pilot and there's no obvious way to adjust the font size. My mom can't read the time or the navigation. She has very good eyesight, but it's still hard for her. And I have to take my eyes off the road. If I were driving, I probably wouldn't be able to do it either. So, it's not very user friendly.
Jennifer Strong: There’s no denying we’ve got real problems in this relationship—it’s just a question of how they get addressed, as researchers who focus on that at the University of Maine are all too aware.
Richard Corey: There is very little discussion from the human side. And, and I will be the first to say this... Good! Because number one, I want to make sure the car stays on the road. That's, that's our number one thing. Make sure the engineering is right and they stay on the road…but we're going to get to a point real soon here where we're going to have to get in them, and we've got to figure out what that means from the human side, what do we need?
Jennifer Strong: Richard Corey is the director of the VEMI Lab at the University of Maine, which focuses on human machine interactions.
Richard Corey: We started to look deeper into it, realized that the notification on a Tesla, which everyone sort of considers to be the leading in autonomous vehicles right now, the notification to let you know that the Tesla no longer has control is four beeps. And we thought that doesn't seem like enough information [laugh] considering the neuro processing and, and stuff that we go through when we're driving a car and being aware of things around us and...
Nicholas Giudice: Dovetailing from that we realized that…
Jennifer Strong: And that’s the lab’s co-founder, Nicholas Giudice.
Nicholas Giudice: While people are also not really thinking about the human side of things and looking at all the human factors, there has been some research, mostly survey research that's pretty consistently shown that people kind of like the idea of autonomous vehicles. And then when you say, well would you like to go out and get into one? We have one right outside of the lab. They're like, Oh no, absolutely not. I mean that's that they show this kind of fear of the technology, fear of how it works. Fear of not knowing, kind of losing control. Cause driving gives you this sense of, you know, you have agency over the process, you have control over what's happening. And when you get into an autonomous vehicle, people seem to be concerned not only about that, they don't have control, but they don't know why the vehicle’s doing what it's doing.
Jennifer Strong: And so, to help understand how this trust relationship might function, they’re studying how humans do it.
Nicholas Giudice: When humans trust each other, it's because they've learned to work together. They learned to understand what they need from each other. And so this idea of human vehicle collaboration, it's kind of this underlying framework to our research. Partly we’re getting at trust by saying if we can build a collaborative process where the human can understand how to interact with and what the vehicle needs and the vehicle understands more about what the human is doing and, and what they need. That process, that collaboration process, builds trust. And so we are essentially building trust through better communication.
Jennifer Strong: Trust isn’t tangible. So how do scientists, focused on user experience, measure it?
Nicholas Giudice: Are they relaxed? We can look at this from physiological measures. So you can look at heart rate or uh, skin conduction. Look at how their attention changes, and so kind of getting at trust through these different types of tools we think is an innovative way to do this. Because so far, this has been looked at through surveys, right? So you have people take these surveys and that's a good way to look at aspects of trust, but it really isn't nearly as direct. And so when we combine the qualitative and quantitative and behavioral and physiological measures, we're really getting a much bigger, kind of more robust way of looking at trust from different levels.
Jennifer Strong: Their in-person testing has been held up by the coronavirus. But essentially what they hoped to do is this:
Richard Corey: We were going to use an autonomous vehicle over the course of a year in a public place. And we were going to start to look at what can we put in there. Can we put in cameras in there? Can we have people in there to do observational studies? Can we put sensors in the seat to find out, I don't know, how much people are wiggling around if they're nervous or something. Any and all data that we could collect to start the process of figuring out who’s comfortable and who isn’t.
Jennifer Strong: Researchers want to track whether people become more comfortable and trusting as cars evolve, and eventually gain the ability to explain what they’re doing and why.
But even when you can measure trust, it’s not created equal. And it’s not given by everyone in the same way.
Nicholas Giudice: People are more willing in some ways, especially someone like myself that's congenitally blind, I'm more willing to take some of the unknown risks that people are concerned about with autonomous vehicles. Cause I'm like, yes! This is going to allow me to get wherever I want when I want. You know, and you could have bus routes, you can have paratransit, but then you're tied to very specific schedules. This is an economically viable and just truly freeing way of, of being able to get around. And so that connection to independence, that connection to quality of life is tied to trust, to put part of the trust is also just knowing that it is even, they're trusting that it's something that can work because right now, so many of the systems that are out there don't work.
Jennifer Strong: The same goes for those who are aging, or have problems that may prevent them from driving.
So, what can we learn from watching how people trust someone behind the wheel—and how do you replicate that relationship when it comes to humans and driverless cars?
Richard Corey: Us as humans can physically tell if somebody's nervous.
Jennifer Strong: Richard Corey.
Richard Corey: They're gripping the dashboard, nails into it, that type of thing, and they’re nervous and then you can adjust how you're driving. You can even say, Hey, it's okay. I mean, I was once driving in a really bad snow storm and everybody in the car was freaking out. And I was like, look, you know, it's all right. I still got control of this. And it's the little human things that you do that connects. What can the car or computer, whatever it is say to you that says, Hey, it's okay. This is why I'm doing this. We're going to pull over to the right lane because it's really hard and we need to slow down a little bit, but we're not really seeing any of that existing technology happening today. And that as humans, we need.
Jennifer Strong: And his colleague Nicholas Giudice says this is how communication might work...
Nicholas Giudice: There may be an avatar, just things that you can actually start to think of the, the, the vehicle in terms of having human-like properties that it isn't just this black box automation that you're just... in.
Jennifer Strong: The longstanding front runner in this race to build autonomous vehicles is a company called Waymo, which used to be known as the Google self-driving car project.
It became the world’s first robo-taxi in 2018, after putting a fleet of self driving cabs on the road in Phoenix, Arizona. And despite having to suspend in-person rides during the global pandemic, their cars still travel 20 million virtual miles a day in simulators.
It’s also among a handful of companies now testing self driving trucks in Texas.
At Waymo, communication looks like TV screens, sound effects… and human-like voices...
Ryan Powell: And so we know today that when you're a passenger in a car that's driven by a human there's a lot of communication that happens between you and that driver that helps establish trust.
Jennifer Strong: Ryan Powell heads up the user experience research and design team.
Ryan Powell: ...so it can be direct communication. Like you might ask a driver which route are we taking or why aren't you moving the light is green but it can also be indirect too. Like you could notice that that human driver might reposition his hands on the steering wheel right before he's ready to make a left turn. But we knew from a very early point that we needed some method of communication in our vehicles to build trust so that riders could feel safe and secure and informed…to that end we spend a lot of time thinking about the visual interface in our car. we try to visualize what the car can see, you know, how do we communicate the car's intent in what situations do riders need additional information, what's the right level of information.
Jennifer Strong: Because, on the other hand, they’re also trying not to overload riders with too much information.
Ryan Powell: So what you hear when you interact with a product or service should receive just as much design consideration as what you see. So when we think about our user experience, that's why we've decided to design not only for what you can see, but also what you can hear as well.
And so this is what we kind of refer to as sort of our welcome track that you would hear once the doors open and as you're kind of making your way into the vehicle, but voice feedback allows us to convey more complex information and also to personalize messages as well. So this is an example of what, when the car starts to move, this is what a rider would hear. So desert breeze park, of course, is the destination that the rider put in there. And we're able, again, to personalize that, that voice by, um, taking the destination that they entered into the app and just reinforced to the rider that we know where the destination is, where we're headed as we, as the car to actually move.
Jennifer Strong: Something I’ve wondered, is what happens if you fall asleep in a driverless taxi? How does the “car” —or the “helper” as Waymo dubs it—wake you up?
Ryan Powell: We give you a heads up when, um, when you are about a minute away from your, your, destination and the idea there is to it in a way, if you've kind of nodded off or dozed off, that's where we use a voice. So a voice comes on and lets you know that you're arriving in about a minute. And the other thing that it serves as a, as a reminder to kind of collect your things, if you've sort of put something on the other seats and so that you don't leave that, that behind, cause again, you don't have that human driver to kind of look around and say like, Hey, you forgot your jacket. Or, you know, you forgot your, your, um, your phone or your bag. And so we do use voice in that case. And we also think, um, you know, that that would help somebody to wake up if they had sort of dozed off while they were on the ride.
Jennifer: A lot of research went into choosing a voice for the autonomous vehicle that riders would trust.
Ryan Powell: We often get asked to like why a female voice versus a male voice, but really the decision for us wasn't about gender. It was more about finding the best sounding voice that matched those principles. And so in the end it just happened to be a female voice for us you know, a voice that sounds competent and thoughtful and friendly, you know, again kind of goes a long way to helping people have a positive experience. And that's very much what we were going for when we thought about the voice that you hear in the car.
Jennifer Strong: And what’s more, Waymo produces all its sounds in a particular key, hoping to influence how we feel inside the car.
Ryan Powell: There is this idea that the key of E is a very sort of brilliant or optimistic key. And so you'll notice that all of our soundscape again, is, is played around this, this key of E. And so I'll, I'll play a sound here in the key of E.
As researchers, that's what we spend a lot of time doing with our riders is we try to uncover those moments where people sort of need additional information to understand what's happening in the vehicle. And again, it kind of goes back to that idea of building trust. So, it could be a simple example where the car, you might just be driving along and the car sort of suddenly slows down and we see this all the time. You kind of, you know, you're not it's, of course it's not a, an anxious moment or an alarming moment, but you, you kind of look up from your phone and you're sort of like wondering like, Oh, what's kind of going on, you know, there's nobody in front of us yet we're kind of driving much slower than we just were. And so that's an example of where we, in that case, we've entered a school zone, but again, because you aren't driving, you probably aren't going to know that offhand. And so we, that's where we are relying on that visual interface or in some cases using sound or voice to let you know what happened.
Jennifer Strong: One of Waymo’s primary competitors is called Cruise. We meet its head of robotics, after the break.
Jennifer Strong: Self driving car company Cruise is part of General Motors. The San Francisco start up builds electric cars and it focuses on ridesharing and delivery in dense urban environments.
Rashed Haq is its vice president of robotics.
Rashad Huck: Self-driving cars have to learn the social behaviors of other drivers or other participants in traffic like pedestrians. And at the same time, they have to conform to human-like social expectations of other drivers and, and the passengers. And they have to do this while attaining superhuman levels of safety.
Jennifer Strong: Part of his job is teaching autonomous vehicles to be more “human,” because there are lots of things we do while driving simply for our own comfort. We favor smoother turns, and we don’t drive in the exact middle of a lane next to parked cars.
He and his team study our habits not just to mimic them in certain scenarios, but also to adapt to them. For example, he says they’re using computer-vision algorithms to analyze the intent of pedestrians from their behaviors.
Rashad Huck: So for instance, if there's somebody standing at a crosswalk, but looking down at their phone, then the AV can infer that either they're unlikely to cross right away or they're distracted and they actually might start walking and not look up.
We have sensors that can see 360 degrees pretty far out, usually more than humans can. And in any given day, we're doing more driving than a human would do in a lifetime. So that gives you some sort of boundaries on how, how fast the algorithms can pick up. Obviously it doesn't have yet to have the judgment of humans.
Jennifer Strong: And before we get there, we need to know how we want to automate away from human driving.
Can we realistically expect people to take control at any moment, when they’ve ceded that control for most of the ride? Airplane engineers realized a long time ago that pilots can't do this safely, so navigation systems always keep them doing some sort of task to stay engaged in flying.
Maybe that means cars should hold back on how much they take the wheel until they can be truly autonomous.
This type of technology requires a new kind of human trust. Though we can’t quantify it, we know that levels of trust people feel increase with transparency, personal experience and safeguards. So, when we drive our own cars, we never really doubt that they will stop when we hit the brakes. As we outsource decision-making to machines, that trust needs to be earned. Without it, more decades will pass, and some of these technologies might never see the light of day.
Next episode: emotion recognition technology hints at a future where our cars, phones, and other devices react to our feelings.
Rana al-Kaliouby: So we build algorithms that can understand your facial expressions, like your smiles or your frowns or eyebrow raises, and map that into an, an understanding of what your emotional and mental state is.
Jennifer Strong: Can we even codify emotions? And should we? Join us for a miniseries on emotion AI, as we explore how to transform smart tools into empathetic ones.
This episode was reported and produced by me and Tanya Basu, with Emma Cillekens and Tate Ryan-Mosely. We had help from Karen Hao and Benji Rosen. We’re edited by Michael Reilly and Gideon Lichfield.
Thanks for listening, I’m Jennifer Strong.
A horrifying new AI app swaps women into porn videos with a click
Deepfake researchers have long feared the day this would arrive.
The therapists using AI to make therapy better
Researchers are learning more about how therapy works by examining the language therapists use with clients. It could lead to more people getting better, and staying better.
DeepMind says its new language model can beat others 25 times its size
RETRO uses an external memory to look up passages of text on the fly, avoiding some of the costs of training a vast neural network
2021 was the year of monster AI models
GPT-3, OpenAI’s program to mimic human language, kicked off a new trend in artificial intelligence for bigger and bigger models. How large will they get, and at what cost?
Get the latest updates from
MIT Technology Review
Discover special offers, top stories, upcoming events, and more.