Studying urban wildfires, and the challenges of creating tiny AI robots
33:46 Share

This is the Science Podcast for March 28, 2025. I'm Sarah Crespi. First up this week, researchers converge on Los Angeles to study the impact of the urban fires from January this year. Contributing correspondent Warren Cornwall is here to talk about his visit to the Altadena neighborhood, where researchers are collecting pollution and health data. Next on the show, researcher Ming-Zi Chen talks about changing up the physics of computing in order to make drones and robots less power-hungry.

In January of this year, wildfires swept into densely populated areas of Los Angeles, burning thousands of buildings and killing at least 29 people. After the fires comes the cleanup, the rebuilding, and the research. Contributing correspondent Warren Cornwall visited Altadena, California, where researchers are exploring pollution from fires.

Hi, Warren. Welcome back to the podcast. Hi, Sarah. Great to be here. I'm really glad to have you back. This is a super interesting story. What was your visit like to Los Angeles? Like you were there in, you know, a month or two after the fires? Yeah, that's right. I was there a little bit more than a month after the fires. And, you know, where I was staying, you wouldn't have even known that there had been a fire if you hadn't been following the news.

It's not like the air across LA was still smelling like smoke. But as you drive into the Altadena neighborhood, you start to see signs of it, which means that you start to see an occasional building that has been burned. And then sort of the closer that you get to the San Gabriel Mountains, which the Altadena neighborhood runs up against, you start to see larger and larger sections of entire blocks that have burned.

just been devastated. There's really nothing left except the burned out hulks of cars and chimneys. Yeah, I think you said in your story 16,000 structures burned. Yeah, I mean, that's not just in Altadena. The two major fires are in Altadena and the Palisades area. And so those two combined, yeah, a lot of homes. And yet, weirdly enough, you would go past blocks that were just incinerated and then next would be blocks that looked like nothing had happened. So it was very eerie.

You call this an urban wildfire. How is that different from a wildfire or a house fire? A wildfire usually just involves burning organic material. So you're burning trees and bushes and grass and things like that. When you get into the urban setting, you're burning a lot of different things that emit a lot of different chemicals. And then in terms of the difference from a house fire, usually a house fire is confined to a couple of houses.

And so one difference is just the sheer enormity of the smoke and ashes. Of course, the effect on the people living there. And then the other thing is that usually house fires, the fires are more or less contained inside the structure. Whereas when you have these fires that are engulfing entire neighborhoods, you actually wind up with different dynamics within the fire because it's an oxygen rich environment instead of an oxygen poor environment contained inside the house.

And so that actually can trigger different chemical reactions. So you wind up with sort of different chemistry happening in the fires. So is just the size, the scale of this event what...

kind of made all these research teams convene on Los Angeles shortly after the fires or even during the fires? Yeah, well, I mean, certainly the size was striking. You know, the fact that it was affecting so many people meant that the stakes are very high. As these wildland-urban sort of hybrid fires go, this one was particularly big and particularly destructive. And also it happened, you know, in a major...

city in the United States. It happened in a place that's full of universities that conduct a lot of research and have a lot of scientists. It happened just a short distance from Caltech. The Jet Propulsion Laboratory was just up the hill. A number of scientists who worked there live in the Altadena neighborhood. You know, it was not just a massive fire that warranted attention in its own, but it was a place where you had a lot of scientists who were sort of ready and eager to understand what it was.

One of the scenes you describe in your story is this house that looks untouched on the surface, sandwiched between burned down buildings and all these teams of scientists kind of surrounding it and dissecting it. Can you talk a little bit about that? Yeah, that house to me really felt like a microcosm of what was happening more broadly. I showed up at this little modest ranch house in an Altadena neighborhood that was just surrounded by assholes.

ashes, the wrecked remnants of homes. And for whatever reason, this house was still standing. And the wooden fence within an arm's reach of the walls of the house was charred. And yet the house itself showed no obvious signs of having caught fire. And when I showed up there, I met with three scientists from UCLA who were preparing to start measuring the air pollution and other things.

inside and outside of the house. And pretty soon there were a dozen scientists there from three major universities who were just swarming over this house, measuring just about everything you can think of. And so to me, that really captured in a tiny little moment what was going on more broadly in LA, that you just had researchers coming from all over the place with an arsenal of different kinds of instruments to try to figure out what was going on.

Were they looking inside the house to see how much smoke stuck around or if there were still pollutants in there? Yep, that was certainly a big part of it. They were measuring the air. They were studying the dust. They were collecting ash. They were taking water out of the kitchen sink to see if there was any evidence of pollution in the water.

They were collecting soil samples. I mean, they were just sort of testing just about everything. And they're also going to be testing, I don't know about for that particular house, but for a number of houses where they're doing this same thing, they're going to be testing the blood of the people who live in the house. And they're going to be tracking their health for years to see whether you can see traces of the pollution in their bodies. To me, this really highlights how many unknowns there are in this situation.

What pollutants are coming out of the fires, where they go, how long they hang around, how dangerous they might be. Right. And that's why this is such a rich scientific story. I don't want to underscore like, I mean, it's a huge tragedy. And the scientists are well aware of that. They're somber about the effects that it's having, but they're also eager to try to understand it. Let's talk about the air pollution inside houses and then also kind of citywide. You know, how has that been tackled?

One of the teams you talk about had air sensors in place. So the Jet Propulsion Laboratory had sensors that were running and they are using that to reconstruct not what's in a home, but details about the emissions that were blanketing L.A. more broadly. But in terms of in the homes,

people are certainly piggybacking on existing work. So there's a group in Harvard that is collaborating with a private company that sells air quality sensors that people can put in their homes. And so that company has sensors in a thousand homes in LA. And so they'll be able to see before and after what kinds of effects there were. Obviously, these aren't sort of laboratory quality air sensors necessarily, but it's

Then another really important part of the research that's being done is going to be long-term epidemiological work, where they're going to be tracking thousands of people over the next 10 years to see if there is any sort of fingerprint of the fire on people's health. And a lot of that is going to be done with existing studies that have been going on in LA that have enrolled thousands of people to track their health long-term.

And often those have like blood samples and things like that already in place. And so then they can come in later and say, do we see something that changed with the fires? Those are, you know, citizens, residents who are just exposed to the fire. What about firefighters? I know there's been some study of what happens to them when they're fighting wildfires for days at a time. How about fighting urban wildfires?

affect their health differently. If you're a firefighter and you're going into a burning house, chances are you're wearing the equivalent of basically a deep sea diving scuba tank, right? So you're breathing fresh air. But when you're fighting one of these sort of massive urban wildfires that are going on for day after day after day, you're probably not wearing any respiratory protection at all. And you're standing there just

stewing in this enormous cocktail of chemicals. I was talking to one scientist who's working with firefighters and has worked with them for years. And he said that some of the firefighters are equating this with sort of a 9-11 event. You know, when firefighters responded to the 9-11 attacks and the collapse of the World Trade Centers, those responses have been linked to a

all kinds of different diseases. And so firefighters in LA are concerned that they're basically having their own 9-11 in terms of potential health effects. And so they're going to be monitored now and going forward in time? Some of them are, yep. Scientists are enrolling hundreds of them in different studies that will involve tracking them for years.

Wow. What else are people looking at, you know, in the aftermath of this huge fire? One of the things that is being done with all this pollution related data is that scientists are trying to build sophisticated computer models that can be used for future fires so that questions like what is in the smoke can be answered more easily than, you know, spending months having people running around with air quality sensors.

And so that's part of what people are hoping will come out of this long term is that the next time this kind of fire happens, and it will happen, they might be able to give more answers to policymakers and firefighters and residents than is available for people in LA now. Yeah. And this is going to happen more because of climate change?

It's going to happen more, I'm told, for a number of reasons. One is that more and more people are building at the edge of wildlands. And so you just have more times when a wildland fire starts, there's going to be houses nearby that can burn.

And then you compound that with climate change, making a lot of places drier and hotter. Then you can also throw in, depending on where you are, you know, years of fire suppression have made a lot of forests more vulnerable to catching on fire. So, yeah, it's really a constellation of factors that are all leading towards more wildland fires. And it's not just in the U.S. There is a study that came out recently that showed that

a higher incidence of these wildland urban fires, a higher sort of encroachment of human development on wildlands all across the world, with Africa actually being identified as sort of a main hotspot. One thing that's interesting to note about the research that's happening right now is that the largest research initiative that's been launched is this joint effort between a number of universities. And

They're saying it's going to cost, you know, $25 million plus for them to sort of make good on all the research that they want to do. So far, the largest funder and the donation that really sort of kicked off this initiative came from the family foundation of a Palisades resident, Evan Spiegel, who was a co-founder of Snap, which is the company that has the Snapchat app.

messaging app. It's not your usual, you know, federal science funding agency kicks loose a bunch of money.

And I think part of what that has enabled them to do is to move quickly. What was it like for you to report from a disaster zone? I think one of the things that really struck me, and this wasn't necessarily a research thing, but I was just really struck by how the fire sort of transformed this entire neighborhood. And yet simultaneously, there were parts of it that were not obviously touched at all. And

you saw people going about just their daily lives of like washing their cars and doing things like that. And in some ways it's a sign of a return to normalcy, but also, I mean, I had people, you know, scientists who were sort of cringing when they saw people just sort

sort of like out picking up debris with like no mask on at all or just an N95 mask on. And I think there's a sense that even though people might think that the fires are quote over, they're not because the legacy of those fires is continuing in the smoke and the ashes and the house that survived. But do they know when it's safe to go back? I think so.

I think the short answer is no. You know, when I was standing there next to this house that looks untouched more than four weeks after the fire, and one of the lead scientists turns to the homeowner and says, have you moved into your house yet? And the homeowner says no. And the scientist says that was very wise.

So it doesn't mean that the scientist has conclusive, you know, they're just visiting this house for the first time, right? Like she doesn't really know what the data says, but she has seen enough. But the homeowner said, you know, some people just down the street had. They have to sometimes, right? There's nowhere else to go. They're not going to pay for a hotel forever. Right.

Right. But, you know, even when I was standing there, you know, so the University of Texas at Austin had driven this instrument laden van, you know, halfway across the country in order to study what was happening in places like this house. And they had these mass spectrometers that were connected to little tubes that they'd snake through a window into the living room. And these devices are giving sort of a real time picture of what's going on inside the house. And

And, you know, I'm sitting there next to one of the scientists who's going, wow, those levels are really high. Of what? Of benzene, which is, you know, a common product of fires and all kinds of other things. It has been linked to different health problems. So, you know, I was talking to David Allen, who's a chemical engineer at UT Austin, who's sort of helping to oversee this work. And he was saying, yeah, the message is homeowners near these fires don't

need to be very cautious, even six weeks later. I mean, when I was there, people were still in the very early stages of cleanup. So on the lots next to the house where the scientists were doing this work, there were people in full Tyvek suits, you know, covered head to toe in these white

suits with respirators on and they're picking through the rubble and they're sort of carrying out little bits of metal. And those were contractors that had been hired by the Environmental Protection Agency to do the first phase of the cleanup. So their job was to go into these burned areas and find hazardous materials. So that might be like cans filled with

paint, things like that, and to take those out. The next stage is that the Army Corps of Engineers is going to be overseeing this sort of bulk cleanup of these areas. So they're going to be coming in and just removing all the debris and hauling it

And so that is probably going to sort of make all this ash come up into the air. There's questions about what effect that will have on the sort of local air quality there. And then even after that, there's questions of like, well, OK, so what's the condition of the soil? Is there any pollution that's seeped in water, drinking water in the area or anything like that? And so these scientists are going to be there for the long haul. Yeah, sounds like it. Thanks, Warren. Thanks, Warren.

Thanks, Sarah. Great to talk to you about it. Warren Cornwall is a contributing correspondent for Science. You can find a link to the paper we discussed at science.org slash podcast. Stay tuned for a discussion of using different kinds of physics for collecting and processing data on board small mobile robots. This week in Science Advances, Mensa Chen and colleagues wrote about using a different kind of physics to do computing on robots and drones.

Think about it this way. Sensors are getting better. We're asking more of robots, more of drones. Navigate this forest. Walk across this uneven terrain. There's just a huge amount of sequential data constantly coming in. It creates a very high demand for computation, for energy on these devices. We want to keep them small. We want to keep them light. We want to keep them cheap.

So in this paper, they're kind of taking the hard part of capturing the data and making it readable by machines in a different way. Normally, when I'm reading an abstract for a paper, I have to look up, you know, a word every couple sentences. In this case, it's

It was like every third word when I was getting started. So this is not easy to understand. Here's Menza describing the problem. So traditional robots, they are usually based on like software-based and digital hardware, like our computers. And they need like higher consumption and...

complexity to deal with dynamic system control, such as like a rover and a drone control. The reason for that is that they need to process those time sequential signals, like input versus the time. And those time sequential signals, the calculation for those are power consuming. And so a more power efficient control structure is desired.

for the future robotic vehicle and the miniature robots. So that's the original motivation for our research. So basically robots, and we're going to talk mostly about robots here, they have sensors on board, they have real-time data coming in, coming in over time, and it's just an overwhelming amount, especially as you increase your sensitivity, it's in a dynamic setting. And

The way that this is taken on board now, it requires a lot of computing power, a lot of actual power. And you looked here into a way of dividing up the tasks so that capturing the information and making it digitally readable is the low power part. Yes. Okay. So we basically use the intrinsic property of the memory register to do that power efficiency calculation. Right.

Right. Okay. That's a good lead-in to what is a memristor. So we break it into two parts of the word. It's memory and resistor. So it has a little bit of a memory for what's happened recently in terms of how much current is going through and it changes its behavior. So by definition, the memristor means that its conductivity or resistivity could be modulated by the charge

previously flow through it. So for example, if we have a high current flow through the memristor, its conductance will be increased. Or for some of type, it will be decreased based on difference. So that's a basic property of the memristor. Okay, so in order to use memristors to encode this data that's coming in, the sequential data,

We got to talk about reservoir computing, which is a type of artificial intelligence. Can you talk about what's different about it? So reservoir computing is a type of artificial intelligence algorithm, and it's

famous for its lower power consumption and computation simplicity. How do you do it? So the first, reservoir computing, mathematically it has lots of nodes and a big network, but it's maybe too abstract. So one thing is that the reservoir computing, the reservoir can non-linearly map the input temporal information to a high dimensional space.

So in other words, the reservoir could differentiate the historical information from the temporal signal, like the input temporal signal. And this single input signal can really be mapped to around maybe five to 10 output signals. And those outputs carrying the historical information of the input signal.

One example I read about that helped me understand this was think about it as disturbing the surface of water. So if that's your reservoir, you can have these inputs that create ripples. The ripples interfere with each other and there's a slight memory over time of what has come before. And then you take a picture of that. That's the state of the reservoir. And you use an output layer to analyze that data. But you don't have to train something to categorize all of this input.

It just affects the physical state of the reservoir, and that can be interpreted by a much simpler system downstream. So the data that's going into the reservoir, and it comes out the other end, then what happens to it? Yeah, so first the input will go through a reservoir, so it differentiates with respect to its time. And so its temporal information could be extracted to lots of outputs.

And then it will go through a weight matrix or the readout layer. It's what they call in the AI structure. And then we can guess the actual output we need. We haven't really talked about specific output you might want from a system like this. But the main point is the reservoir does all of the hard work of sequential data crunching, kind of turning it into this physical format. And then there's an output layer that will classify that. And this is how you can save energy. You can say it's low power.

All the stuff that you normally associate with artificial intelligence, classifying things, training on data, using a weighted matrix, it's after the reservoir has turned your data into something readable. You're analyzing the state of the reservoir instead of the raw data coming in from sensors over time. Okay, so let's go back to what the reservoir is actually made of. That way we may come back to why we need to use memory register to achieve this.

So traditionally, we usually have a huge array of memory registers. And each single memory register, you can consider it as a computation node. So for example, if I input data signal as a current,

going through this memory star array and we know this input value. And the only thing we need to do is measure the voltage at each memory star. By doing so, we can do a simple multiplication because we also know the resistance of each memory star. So by the simple Ohm's law, they can do the matrix multiplication at a very low power consumption and very fast. All you need to do is just measure the resistance.

the voltage. So that's why a lot of people have used MemoRaster to do a low power consumption matrix calculation. And for our case, it is a little bit different, but the fundamental principle is the same. So in our research, we have a MemoRaster network. So network has a lot of channels, right? So one channel we put as input and we measure the voltage at different

So also as inputs going in, we measure the voltage value at different outputs and we can finish the computation. And the power is really low. So

So yes, it's not as simple as disturbing the surface of water, but these memristors, the nodes can, you know, capture a lot of data and it's measurable as voltage. You don't have to do operations on the actual data coming in. That's like kind of the key of the reservoir. You're not doing the same normal computer stuff. You're using a different kind of physics, right, to do this work.

heavy lifting. And then after that, you have this simpler job for your next layer from what you get from the output. Yes, that's correct. Okay. Okay. So now we have kind of all of the nouns. We have Memrister, we have reservoir computing, we have robots. So why don't we talk a little bit about what your experiment was designed to show? What were you trying to demonstrate with your setup and

Yeah, so what we are going to do as we talk about our initial motivation, we want to use this low power system to achieve robot control. So the first task is that we use this system to control a rover to tracking red target swinging in front of it. So this is a little, I looked at the video, it's a little guy, it's a little robot and it's chasing basically a spot. Yes. So for that demo, the whole thing is kind of simple because that's our first demo, right?

So we have a camera that captures the red targets and we use software-based AI algorithm to extract the position information. For example, if it's in the middle, the position will be like one, and if it's in the left, it will be zero, and if it's in the right, it will be like two. And those signal positions information will be sent into our hardware reservoir.

And then, as we talked before, we measure the voltage value at the output channel of our memory-based network.

and those values will be exported to the trained-weight matrix. And finally, the output will be the power for the wheels. So basically, the output is still a single output, and it's the difference between the left wheel power and the right wheel power. So if the value is 1, it's going straight forward. And if it's 0, it means that it's turning, maybe turning left. And if it's 2, it's turning right. It's an example.

We're not talking about tons of sequential data pouring in through audio-visual sensors. We're talking about, for this proof of principle, zero, one, or two is being fed into the reservoir. The memristors are converting that to a voltage that's read by the readout layer and interpreted as directions for the wheels of the rover. You also looked at how this might function in a drone. Can you describe that? Yeah, sure. Sure.

So a future application of our system is like we want to integrate them into the drones, which need them to control the propellers. So our second demo is that we use this control system to control the power of a propeller to try to lift it up to the horizontal and keep it stable. So basically in that case,

The input will be the angle of the lever. It will be like a 30 degree. And then we're targeting to lift it up to 90 degree, which is horizontal. So basically the input is angle and the output is propeller power.

Okay. And so this is kind of a proof of principle that you can use reservoir computing with memristors on board small robotic devices. And then scaling that up means what? What kind of inputs would you consider this being ideal for?

what kind of sensor systems or robotic tasks? That is actually our future direction of the research. We're trying to implement into more complex system. And the first thing we can consider is still drones because drone, when it's flying the skies, there are a lot of variables that you can control: the power, the angle, and the speed. And also, for example, the drone have at least four to six propellers for them.

and each of them needs control. So when we want to scale up, I think those kind of tools is a potential application. So what did you show in terms of the improvement of power usage with this system compared with some other approaches people have taken with these either target finding tasks or balancing the drone tasks? So we did some calculation on the power consumption. We only comparing the computation units, not other things.

We have like a hundred times smaller power consumption than a traditional software-based computational algorithm. Because we only need to measure the voltage value, we don't need to do like the analog to digital signal conversion and signal transmission. We are much closer or much simpler than the traditional computer. That's why we can save that much of energy. All right. What did I miss? What else should we talk about?

There are several researchers already that use memory register to achieve relative computings. For example, they have used relative computing to detect if your heartbeats indicate you have certain illness, or they use it to do some voice recognition for certain numbers. So however, those tasks, they

They can only do like yes or no tasks, or basically the result is digitized. They just have certain several results. However, for our system, our output is continuous and it's analog output. So that's a big difference from the previous work. And this is thanks to our unique device structure. So previously, researchers usually have a big array of memory stores. They are doing those calculations.

But our memory-sensitive network has a shorter memory effect. Our memory-sensitive network, we basically constantly measure it. So we have analog input, but also the export result or output is also analog. So we can use this analog signal to control a robot rover. This is something that I haven't done before by other people.

Mingzhe, thank you so much for talking with me. This has been complex, but very interesting. Thank you. Okay, Mingzhe Chen is a graduate student in the mechanical engineering department at the University of Michigan. You can find a link to the science advances paper we discussed at science.org slash podcast. And that concludes this edition of the Science Podcast. If you have any comments or suggestions, write to us at sciencepodcast at aaaas.org.

To find us on podcasting apps, search for Science Magazine or listen on our website, science.org slash podcast. This show was edited by me, Sarah Crespi, and Kevin McLean. We had production help from Podigy. Our music is by Jeffrey Cook and Wenkui Wen. On behalf of Science and its publisher, AAAS, thanks for joining us.

From April 25th to 30th, the American Association for Cancer Research will host the AACR Annual Meeting 2025 in Chicago. This meeting is the critical driver of progress against cancer, the place where scientists, clinicians, and other healthcare professionals, survivors, patients, and advocates gather to share and discuss the latest breakthroughs in cancer research. Both in-person and virtual registration options include access to live sessions,

Q&A, networking, CME and MOC credits, on-demand access to sessions after the meeting, access to the e-poster platform, and more. Learn more about the AACR Annual Meeting 2025 at aacr.org slash aacr2025 and register for the world's most important cancer meeting.