Shamini Bundell
Welcome back to the Nature Podcast, this week: how to test and train safer AI drivers...
Nick Petrić Howe
...and what researchers can learn from the immune system of bats. I'm Nick Petrić Howe.
Shamini Bundell
And I'm Shamini Bundell.
<Music>
Shamini Bundell
Self-driving cars, or autonomous vehicles, have been a dream of engineers and people who hate driving for a while now. But they've been hitting a roadblock in their development, and that is training them to react safely in dangerous situations. Now, though, there's a method that could help make autonomous cars safer. Reporter Anne Chisa is here with the story.
Anne Chisa
Autonomous vehicles have been a hot topic for the past decade or so. In theory, by removing the driver from the equation, they could increase safety by never getting distracted like a human might. Also, it would free people up to do things other than driving. But after many years of research, we still don't truly have an autonomous vehicle. That is a vehicle that can drive on autopilot from a starting point to a predetermined destination, where the human doesn't need to be there at all, and is driven instead by artificial intelligence. And part of the reason has been questions about safety — they just haven't matched humans or exceeded them when it comes to driving safely. To make these vehicles safe, we need to teach the AI how to handle risky situations that occur on the road. But to do that, we need to expose them to dangerous situations in the first place. And therein lies the problem.
Colin Paterson
One of the problems is that you can drive many, many millions of miles and never see anything that's of consequence.
Anne Chisa
That's Colin Paterson, an expert on autonomous vehicles from the University of York.
Colin Paterson
So the reality is that if you're looking for dangerous situations, most people will drive for a full year and never see anything close to an accident in their own driving. So that's a pretty inefficient way of finding out how you deal with dangerous situations.
Anne Chisa
Now, autonomous vehicles can handle many critical road situations, but there are certain scenarios known as corner cases that they struggle with. These cases are rare, those once-in-a-million-miles sort of an event, and they require extensive testing time and cost to ensure safety is reached. But this week in Nature, there's a paper that shows a new approach to turn the corner on these corner cases, as Henry Liu from the University of Michigan, and one of the authors of the new paper explains.
Henry Liu
We develop what we call dense learning approach. What it is, is we train an AI agent. And so that AI agent can see these safety critical situations very, very often.
Anne Chisa
If this talk of agents is making you think of spy movies — that's not the case here. It actually refers to AI training the AI using real world data collected from the United States and Germany. Henry and the team were able to train a bunch of AIs to be bad drivers. And so they can train the original AI in a simulated environment where there are lots of different dangerous situations created by the bad AI drivers. In fact, the bad AI drivers were specifically trained to create those rare corner cases, those safety critical times that the AI needs to take over the wheel quickly. The testing of autonomous vehicles happens not only in a simulated world, but a physical test track called Mcity.
Henry Liu
So we have a closed test track we call Mcity, which in this case is a miniature city. It has all the city infrastructure in it, and it's dedicated for autonomous vehicle testing. And so we develop what we call a mixed reality approach. We run the exact copy of the Mcity in simulation as well and we have a physical infrastructure, and then we combine these two together.
Anne Chisa
The neat thing about the combination of testing the autonomous vehicles in a simulated world, and a city, is that it clears up the path of those roadblocks mentioned earlier regarding the efficiency of testing to improve safety.
Henry Liu
We also see a lot of these safety critical cases — they learn from that — so they know what type of situation might lead to safety critical situation. And once this lead to safety critical situation, how they handle that.
Anne Chisa
Henry believes that this method could help see autonomous vehicles actually driving around in the future.
Henry Liu
Oh, I think the technology we're developing this paper has the potential to really improve the safety performance of autonomous vehicles. It's actually not only improved framework for autonomous vehicles, but also for the training of autonomous vehicle as well. It's critical because it's important not only to the developers, it's also important to the consumers. And we want to make sure vehicles are safe before they can be deployed on the road.
Anne Chisa
Colin, who you heard from earlier, was impressed by the new work, but thinks there's still a lot more to be done in order to get truly autonomous vehicles. Part of the ongoing problem is that there are lots of different kinds of roads in the world.
Colin Paterson
In the paper, what's happened is that they come up with one or two environments in which they've been able to test it. But that doesn't necessarily mean that if we were to come to a city in the UK, which has many, many roundabouts that we can say it will still be safe. Or if we were to go to Iceland, and we were to test on particular road surfaces that we don't see in the rest of the world, does it still work?
Anne Chisa
There were also some assumptions made by Henry and the team to build the AI models. But some of these assumptions may be missing things that are key when it comes to driving in the real world.
Colin Paterson
Things like the grip on the tyres is not considered. Yet if you talk to an automotive engineer, they'll say the engineering behind something like a tyre is crucial to how the vehicle manoeuveres on the road.
Anne Chisa
Altogether though Colin and Henry believed that this new method of training autonomous vehicles could really help improve their safety. There's still more work to be done, of course, but it's possible that a future where driving is also a time to catch up on emails is a little bit closer.
Henry Liu
Automated vehicle has the potential to improve traffic safety in the mobility. And so then you can utilise that time in the vehicle to do some more productive work.
Shamini Bundell
That was Henry Liu from the University of Michigan in the US. You also heard from Colin Paterson from the University of York in the UK. To find out more about training autonomous cars, we'll put a link to the paper and to Collins News and Views article in the show notes. But that's not all. If you want to see self-driving cars being put to the test. I've been making a video all about it, which involves lots of virtual vehicles swerving dangerously into the path of the actual car on a special test track in Michigan. Again, we'll put a link to that video in the show notes.
Nick Petrić Howe
Coming up. We'll be hearing about researchers renewed interest in bats and what it's revealing about these flying mammals. Right now, though, it's time for the research highlights, with Dan Fox.
<Music>
Dan Fox
Bird flu may soon become bird flu in name only, as it appears to be adapting to wild mammals. Researchers sequenced the genomes of bird flu strains isolated from 40 infected foxes, skunks and mink from across Canada. The isolated viruses, all variants of the H5N1 strain, fell into four categories. One closely related to strains circulating in Europe, and three that seemed to have evolved in North America. Many of the strains contained mutations that favour viral replication in mammals, and all the analyzed strains contained mutations that increased the virus's ability to bind to human upper respiratory tract cells in lab experiments. Although the alterations that confer the strongest binding ability were absent. Whether these mutations will cause the virus to jump to humans remains to be seen, but the authors suggest health authorities should monitor wild animals continuously to track the spread of these and future strains. Read that research in full in Emerging Microbes and Infections.
<Music>
Dan Fox
Bad news for fans of upbeat headlines. A study has now shown that news headlines containing negative language are more likely to be clicked on than those containing positive words. A team of researchers analysed more than 100,000 variations on headlines finding that a reader was 2.3% more likely to click on a headline of average length when it contained an negative word — such as wrong, bad, or awful — than when it contained no emotional words. In other awful news, a reader was 1% less likely to click on a headline of average length for each positive word it contained. The effects of negative language and headlines was strongest for political and economic news, which suggests that people are most likely to consume this type of news when it is negative. You'd be wrong, bad and awful to miss that research, read it in full in Nature Human Behaviour.
<Music>
Nick Petrić Howe
If you cast your mind back to 2020, when the COVID-19 pandemic was first rearing its ugly head, you probably remember that one thing that people were quite keen to figure out is what animals could have passed SARS-CoV-2 to humans. One of the candidates was bats. That's because these flying mammals are known to carry a range of coronaviruses. They also carry over serious viruses like SARS, and Ebola. And unlike humans, bats seem to tolerate all of these viruses pretty well. This combination of factors has meant that of late, a lot of researchers have been delving into the world of bats, breeding them, transporting them and studying their immune systems. Smriti Mallapaty, Senior reporter for the Asia Pacific Bureau here at Nature, has been writing a feature all about this renewed interest and what it means for bat research. She joins me now. Smriti, hi.
Smriti Mallapaty
Thank you for having me. I'm happy to be here.
Nick Petrić Howe
Well, thank you so much for being here. And so the first thing I wanted to talk to you about is that there's been this kind of renewed interest in bats for maybe quite obvious reasons. And so I was wondering, what are the sort of key questions that researchers are trying to answer about these flying mammals?
Smriti Mallapaty
Yeah, so bats are really diverse, they're more than 1450 species of Bat. But different bats are reservoirs for different families or different types of viruses. And there's a really high diversity of coronaviruses in certain species of bats, especially Rhinolophid bats. I think the biggest question that I've found from speaking with researchers, especially as it relates to the immune response, is how bats can live with all of these viruses that are so deadly to people but not deadly to bats — I think that's the biggest question — is, what is it about bats that allows them to do that? Is it something unique about bats? Or is it just a relationship that some of these bat species have with some of these viruses? And then what is it that we can learn from the bat immune system that we could then translate into potential therapies or treatments for people.
Nick Petrić Howe
And so with this renewed interest, people are starting to do a lot more work with bats. And one thing you wrote about in your Feature was this quest to make stem cells from bats. These are a bit of a workhorse in biology, but they're tricky to make from bats. But they could help us better understand how viruses and bats coevolved. So right at the start of the pandemic, there was this team in New York trying to make these stem cells, but they didn't have the right sort of bats, so they called up some researchers in Spain.
Smriti Mallapaty
Yeah, so they called this team in Spain. And these researchers in Spain went out to an abandoned tunnel in a dam. And they went out in the evening, they trapped two bats. Then they drove six hours from the south of Spain to Madrid. And there were very few planes flying at the time, but they managed to get one of two flights that FedEx was flying from Europe to the United States. And they arrived at the FedEx warehouse. And just as they got there, they ended up dissecting all the bats and putting different tissue into different tubes, wrapping it all up. They had like an hour in the parking lot of the FedEx warehouse to do all this, put it into an icebox. And they almost missed their flight because as they reached the FedEx counter, the woman was like, you're too late, the plane is about to leave. And there was a bit of drama there. And they finally found the manager and the manager kind of understood the importance of getting these cells to New York and then he said, don't worry, I'll sort it out and they managed to get the samples onto the plane finally, and the way the researcher described it, it was a really emotional moment because he'd spent months getting permits to get these bats, so he was very relieved. And so the package arrived in New York about 26 hours later, there was one stopover. And in New York, the team in New York, we're waiting for the samples and a lot of the cells had died. But they managed to still recover some of the cells and from the skin cells, they managed to develop these induced pluripotent stem cells that are quite a useful tool for studying bat biology. And from that you can develop different bat tissue.
Nick Petrić Howe
Crucial work right at the start of the pandemic. And one of the other things that you're writing about any feature as well, is this team in Singapore who have been kind of breeding bats to sort of help understand them a bit better. Can you tell me a little bit about this?
Smriti Mallapaty
Yeah, so one really important tool that researchers rely on are animal models, so live animals. And one way you could study bats, as you could, you know, go out into the wild and trap bats. But another way is to also develop colonies where you have a consistent source of bats, and you can do reproducible research. Unlike the typical lab mouse, bats, you know, are pregnant for longer and have fewer pups. And so it requires a little bit more investment to breed bats. And this team in Singapore led by Lin-fa Wang, who has been studying bat immunology for many years, they've established the first bat colony in Asia, so I went to visit that colony. And the bats are really important because they help them study bat biology in more detail. And so they have more than 100 bats there now, and it is really quite surprising how small they are. And they showed me around, they gave the bats a little health check and little bits of juice out of a pipette. They're really cute too, they have these large eyes and pointy ears and furry faces, but these bats are the ones that they're using to study the immune response of bats as well. As you know, bats are interesting for many other reasons. They're also studying their metabolism, bats have low incidence of cancer, they live relatively long lives for their size. So they're studying many different aspects of that biology and one of them is immunology.
Nick Petrić Howe
And so what is all this effort that's going into this field revealing about bats? Do we understand anything more about how they're tolerating all these viruses?
Smriti Mallapaty
So a lot of what we know about bats is about the innate immune response in bats. So the response that's more of an immediate response, but not the adaptive immune response. And so now that you know, there are more colonies that are coming out, there are, you know, bat genomes being sequenced to high quality, there are these new tissue models like the stem cells, researchers are starting to be able to understand a little bit more about the adaptive immune response in bats. And the other thing that they're starting to be able to do is — we talk about bats, but since there are, you know, more than 1400 species — they're starting to really characterise the differences in different species. And you know, the immune response might not be the same in all of these species, and how different is the response in all these bats? And how much of what we're learning is something that can be generalised to bats? Or is it really just very specific to the one species of bat and the viruses that it hosts? And yeah, researchers kind of hope that now in the next few years with all of these new tools that are coming out, and the money and the interest that we'll really be able to get in more detail answers about what is it about the bat immune system that makes them able to tolerate all these deadly viruses.
Nick Petrić Howe
Smriti Mallapaty there. To find out more about bats be sure to check out Smriti's Feature article in the show notes.
Shamini Bundell
And that is all for this week. If you want to keep in touch with us you can, we're on Twitter as @naturepodcast, or you can email us podcast@nature.com. I'm Shamini Bundell.
Nick Petrić Howe
And I'm Nick Petrić Howe. Thanks for listening.