Cathy Wu of MIT on The Robot Brains Season 2 Episode 6
Transcript edited for clarity.
Pieter Abbeel: We've had guests on this podcast that are building the brains of the cars and trucks of our future from Tesla, Aurora, Waymo, to Wayve. Today's guest is building the roadways of the future. She's using machine learning to predict the ideal infrastructure for our future mobility, the cost of building this infrastructure and most importantly, how do we do it? So noone has to waste time sitting in a traffic jam. Yes, that's right. Her research is ensuring our self-driving cars won't ever have to sit in bumper-to-bumper gridlock. I am, of course, talking about MIT Professor Cathy Wu.
Welcome to the show, Cathy. We're honored to have you with us. So great to have you.
Cathy Wu: It's great to be here.
Pieter: So Cathy, right now you are working on machine learning and optimization for the future of transportation. But as I understand it, it wasn't the immediately obvious choice as you were born into a Chinese family. How did growing up in a Chinese family land you doing what you're doing right now?
Cathy: It's a really good question. I mean, yes. Coming from a Chinese family, I was supposed to be a doctor. I joke that I am just bad at following instructions, and I am a doctor now, just a different kind of doctor. I mean, joking aside, I unfortunately found my biology classes very uninspiring in high school. So I knew at some point before going to college that that was not going to be a viable path for me. Sorry to disappoint my parents, but I kept that idea with me that I should be helping people with my work. And so starting from the first day of undergrad, I sort of looked around to try to understand like, “OK, if I'm going to be an engineer, which I knew I wanted to be, I wanted to be an engineer. How can I still help people?”
I really just looked into different kinds of societal domains, from agriculture to energy systems to education and so on and so forth. Yeah, transportation is not like a natural first pick, and I would say that it really was self-driving cars that struck a chord with me. I knew I wanted to study electrical engineering and computer science. All these other domains like these ideas, these skills are sort of not like they're good, but they're not necessarily critical, like mission critical, for some problem in health care, where maybe there's a lot of red tape that needs to be worked around. But with self-driving cars, you can see the pathway from that technology coming to fruition and impacting people's lives. But also it cannot happen without computer vision, without, you know, a path planning and a lot of ease. Yes, methodologies.
Pieter: So did you already realize this during your undergrad or when did this happen?
Cathy: Professor Seth Taylor, who has unfortunately since passed away during my junior year of undergrad, gave a lecture when I was competing in a robotics competition here. It was actually this very interesting moment, and I think those of us who get to experience this moment are very, very lucky where he gave that lecture. And then everything else I was thinking about just melted away, and I knew that this is what I wanted and I reached out to him to work in his group. I joined his group. He was no longer working on self-driving cars at the time. He was working more on assistive technologies all the time, so I worked with him for some time and after a while I told him I think I still want to work on self-driving cars. And so he actually helped me pave the way to working in other groups and talking with Google's self-driving car team and so on.
Pieter: Well, not sure if you remember, Cathy, but I remember talking with you during your visit to Berkeley. And I was so excited about reinforcement learning and the new opportunities there. And I knew your record. Of course, I'm like, you know, very hopeful you're going to be equally excited as I was about reinforcement learning. And you tell me, this is all very interesting and those are great tools. But I really want to start from the problem I'm going to solve. And the problem I want to solve is transportation. And I'd love to bring reinforcement learning into it as needed. But I'm not just going to do a Ph.D. in reinforcement learning to move the needle on that. I want to solve transportation. In my experience, I didn't have such a clear goal in mind. And you actually did it. I mean, you haven't solved transportation yet but we actually went down that path very successfully and you're on track to do it.
Cathy: Thank you for bringing that up. I think you're saying this more tactfully than me. In my recollection, I said, I don't want to do this in robotics, in like search, surgical, robotics or something. I want to work on transportation. And so actually, I forgot to mention the link between self-driving cars and transportation. I I actually made an attempt as a senior to work for Google's self-driving car team at the time. I think mostly it was the teams from CMU and Stanford. Everyone had huge deals and whatnot. I didn't have any degrees, so they weren’t that interested to talk with me. But this is actually one of those fortunate circumstances where I thought, OK, I can. I'm not prepared to work with Google Self-driving Car right now. I will build some skills that they will need in a couple of years. They're going to solve the one car driving around problem, not running into children, like identify stop signs, etc. I will start thinking about multiple cars and how they interact with one another. And whatever that means and all, I'll be able to join them in like two years. This was 10 years ago at this point, it’s just a lot harder than any of us ever thought.
Pieter: A lot of people, when they think about the future of transportation, they think about how can a self-driving car avoid accidents? How can we maybe build flying cars so it can be up in the air and follow a straight-line path instead of, you know, roadways? But you're actually thinking about something quite different than those two specific things. Can you say a bit more about what you're thinking about?
Cathy: Yeah, I say I really think about self-driving cars in this system's context. So we have a lot of systems level challenges like congestion, like emissions safety as well. It can be a systems challenge as well like air pollution and things like this. And they are not the result of just a single vehicle moving around in space, but they're a phenomenon that's created through the interaction of many agents in the system, many cars interacting. So there is a really nice study from about a decade ago that demonstrates the formation of traffic jams just purely due to human driving. So it's a scenario with no stop signs, no traffic lights, no lane, no extra lanes or anything, and just humans driving together for a stop and go traffic jams. And it's this interaction between the drivers that creates interesting phenomenon. And so the extent to which we have this new tool, in some sense of self-driving cars, how can we improve the system? Its is a really exciting opportunity.
Pieter: So it's interesting because he's saying human driving, even when there is no traffic lights, no stop signs, no lane changes required still causes traffic jams. And then in my mind, the question becomes, is it fundamental to safe driving or is it something just from the style, the way humans drive and maybe an autonomous car can avoid this?
Cathy: So actually, to counter to this this point, there are studies about how stop and go traffic causes jams. Variations and speeds in roadways actually correlates with more accidents. And so the extent to which we can mitigate congestion, we can potentially also improve safety at the system level, which is another really exciting aspect to this. I would not say that human driving is like the gold standard by any means. And I think there's this is another aspect that I have come to really appreciate about working in this domain that oftentimes when we think about robotics and I and I do have some robotics background, oftentimes we sort of know what we want the robot to do. Like we know we as people are pretty good at picking up objects. We're pretty good at walking. But when it comes to driving, like, yes, we can drive and we cannot crash into things, but we are not actually able to drive very well in a way that improves the system. And so that's a nice opportunity for self-driving cars.
Pieter: Say a bit more about that. How can the self-driving car actually improve our traffic situation?
Cathy: Oh, yes. OK, so this is now where our worlds intersect. Ane we start digging into reinforcement learning and and studying these systems with these pretty advanced A.I. tools. And basically, what we find is so, so more broadly, I'm very excited about how reinforcement learning can actually sort of help accelerate the discovery of insights into these complex dynamical systems. And here's just here's one example where if we use reinforcement learning actually to model self-driving cars, we don't actually know what self-driving cars will do in the future because they don't. They don't really. They're not really pervasive yet. So let's use reinforcement learning to model them. And these these reinforcement learning agents can optimize whatever we as a systems designer desired. So let's say we're we wish for the agent, the self-driving cars to optimize for the speed of the system. And what we found numerically in numerical experiments is that just a very small fraction of vehicles controlled. So five to 10 percent can have really outsized impacts on the system. So five percent of autonomous vehicles can basically eliminate traffic congestion in these model highway settings in a way that we didn't we didn't realize was possible. And it can also do so at some theoretical also, which we can reference through control theoretic techniques. Even more so like the on congestion, we there are a variety of self-driving car traffic phenomenon that these vehicles are able to to help with. And we're starting to uncover these. So, for instance, in highway bottleneck scenarios, there are other forms of congestion that might form. And we're finding that again, like 10 percent of autonomous vehicles can can mitigate this phenomenon called a capacity drop. And we're also finding that self-driving cars can potentially aid emergency medical vehicles in getting to their destination, and that has some safety implications as well. And so there's a there are a lot of cases where we can start to think about, oh, we can model self-driving cars. Is this like helpful agent for the benefit of the system?
Pier: And as I understand it, from your work. It doesn't need to be that every car is a self-driving car can be just a few or even just one that is doing the right thing and helps alleviate the traffic jams.
Cathy: Yeah, that's exactly right. So that's one of the aspects that was really important to me to study because at the time when we were starting this work, it it felt like there was a lot of there are a lot of people studying, like one self-driving car, navigating the local environment, not crashing into things. And then there are people who study. There's a lot of work that studies like systems where all vehicles are autonomous. And to me, that's like decades away. So I was really interested in what can we do when we start to have vehicles rolled out, like one percent, five percent. So it was a really important part of our study. And this is actually one aspect where reinforcement learning really enter the picture as important because what we find for systems at the sort of periphery of what I call the mixed autonomy spectrum between zero and 100 percent adoption of autonomous vehicles. The periphery actually has a fairly well-established tools from control theory and from stochastic systems. But in the middle, where we have this interesting space where the problem is actually more complex because we don't control everything and the scope of the problem is in sometimes larger.
Pieter: Now one thing I'm really curious about, I can definitely see how the mix of autonomous and human drivers makes a problem a lot harder than just autonomous, because just autonomous, you control everything. But I'm curious…you mentioned how a single autonomous car can alleviate traffic jams, right? And I'm curious, do you see it as generalize to multi-lane roads, let's say highways and is there any intuition into how this autonomous car is doing this? And let's say if I could drive, if I followed the right recipe, can I leave the traffic jams just like the autonomous car would be doing in your system?
Pieter: Yeah, yeah. These are great questions. So we do have some studies that demonstrate generalizing the this single way and makes a lot of results for congestion mitigation to multiple generalization. The result is really interesting. So there is one interesting phenomenon that highway patrol officers actually engage in, which I didn't know about, but sort of emerged from some of our experiments. And this is called a traffic brake. So oftentimes on a highway, when there's an accident, there's a lot of congestion build up and something the highway patrol officers will do is they just sort of swerve in a pattern across five lanes of traffic to hold back traffic so that vehicles don't sort of just accelerate to maximum speed and then like cause congestion downstream. And so that actually fell out of some of our simulation analysis through the use of reinforcement learning. That was really interesting.
More recently, we've done a follow up on this study and we've actually uncovered that the reinforcement learning agent will actually learn to have the self-driving car use its blinkers to also avoid some other vehicles from maneuvering too aggressively as well. It’s interesting that if you're signaling to change lanes and other people will sort of let the car do so. And that's another way of issuing control, exerting control in system. So I think the the extent to which, like multi lanes are going to be fully addressed is still out there for at least like a single autonomous vehicle. But we can always just, you know, have as many autonomous vehicles as we have drivers. I forgot your other question.
Pieter: Well, my other question was, can I adapt some intuitive driving pattern that you discovered in your simulation so I can personally alleviate the traffic jams?
Cathy: OK, that's an interesting question. And it touches something I'm really passionate about right now, which is our sustainability and climate challenges. One potential for congestion mitigation is that we can reduce CO2 emissions, potentially in the near term if self-driving cars were ready, but they're not ready yet. We still have safety and robustness issues. And so recently, my group has been really interested in whether lower levels of autonomy could also mitigate congestion through more low-grade types of instructions. We’ve been studying what we're calling ‘human compatible driving’ in which we use reinforcement learning to understand the engineering requirements, and the technical requirements for mitigating congestion. We model am autonomous vehicle as a vehicle that can follow a instruction very well every like 0.1 seconds. You have a new instruction, but a human driver or a or another level of autonomous vehicle like a level two or below vehicle cannot necessarily follow instructions at such high granularity in the sensing, maybe lower fidelity as well. So we just went to the far extreme where we're considering like, OK, for the human driver, you can follow instructions every I don't know, five seconds, 10 seconds, 30 seconds can we still mitigate congestion? And so we have some preliminary work out that indicates that up to like fifteen to twenty five seconds per instruction, human drivers can still mitigate congestion. It’s indicating some early results that human drivers can also mitigate congestion where you can imagine a smartphone app giving some guidance on the right speed or acceleration to drive.
Pieter: So instead of just running a regular chips that tells us which turn to take, what kind of instructions will they give us?
Cathy: Yeah. So what we're hoping for, this is still early. We still need to engage in the human factors aspects of the study, but it'll be something as simple as possible like the acceleration of the speed.
Pieter: So it might say speed up a little bit or slow down a little bit.
Cathy: That's right. That's right. And hold this for 30 seconds or so.
Pieter: Wow, that's interesting. That brings me to another question, which is simulation and reality. Like how well are they matched up and how do you make sure that the close match is there?
Cathy: Yeah. So this, I actually don't know. So here are a couple of things I can say. We are starting to form some collaborations with doctors and researchers, so that we can put some of this into practice. My students and I have very roughly prototyped an app. And then I have been personally testing it. But it's not scientific at all right now. But just to see like, is this totally off base or not? We really want to know. We’re really serious about whether this kind of work can see the light of day.
Pieter: You've been saying self-driving cars could alleviate congestion, but some people also say, and you must have thought about this, too, that self-driving cars will make people more eager to jump into the car because, you know they can do something else. They can maybe sleep. And so how about that side of it?
Cathy: It's a really good question. It's one of my greatest fears that we as engineers do all this work and we improve the efficiency of our systems and then all the improvements build what's called induced demand, which is what you've described. And so my sense is that we have to pair engineering improvements, engineering innovations with policy and other strategies that manage the induced demand of whether it's OK to, like, be wasteful, for instance. And that can take the form of like pricing. It can take the form even beyond policy. You can take the form of public messaging around like social responsibility. I don't think that this is a purely technical problem.
Pieter: Yeah, excessive pricing. I'm thinking, OK, I now want a reinforcement learning agent on my behalf to learn the cheapest way for me to get to the places I want to get to.
Cathy: That’s right. That's right. And that might mean traveling when there's less traffic or something like that.
Pieter: Now, just the cars, as you said, policy might be needed in many ways. When I think of your work, I think of it as a way of somehow using existing cars to have the same effect as traditionally is achieved by building infrastructure. Normally, infrastructure will maybe have a stop sign or a traffic light, and you're essentially saying, Oh. If I can control a few of the cars on the road, they can induce a lot of those patterns. But are there still things that you think in our road infrastructure would need to change beyond just what, you know, a change to the software and self-driving cars?
Cathy: Yeah. So I mean, definitely given my background, I'm a high proponent of software solutions to these problems. But yes, I mean, infrastructure comes hand in hand with these solutions. One of my favorite examples in the self-driving car space is these highways that have magnetic sensors embedded in the pavements. And if we were willing to put in the investment, which I believe is not so expensive, this essentially solves the localization problem, the problem of determining where is the car relative to the lanes. It addresses the problems of the lane markers or if there's snow or heavy rain. The self-driving car can know exactly where if there's some little piece of infrastructure that is sort of helping. So I think that infrastructure is a really big part. The intersection between these infrastructure questions and my work is that I view reinforcement learning as a really powerful modeling paradigm for understanding where we need sensing. What kind of sensing do we need? And is it going to be fully on board of the vehicle? Or is it going to be sensing that can only be provided by the infrastructure? So for instance, one hypothesis that we've long had is that in order to mitigate congestion, you need to have a notion of density of traffic. And the best way to get a notion of density of traffic is through infrastructure sensing rather than through onboard sensing. And we are able to use reinforcement learning to understand whether or not that is the case. Our research has indicated, at least at a preliminary level, that we only need local sensing for mitigating certain kinds of congestion, and that actually has implications on the cost of the system. It has implications on whether we need to install infrastructure. On the other hand, I will say to answer your question and more concretely, I will say I am a big fan of dynamics speed limit signs. I think if we have those in a lot of places, we would be able to mitigate a lot of congestion without anyone retrofitting any cars.
Pieter: And that's not too expensive to do, probably. But it sounds like you still would need sensing right to guide the speed limit. You said dynamics is always changing based on the current situation.
Cathy: Yes, yes we would. That would be an infrastructure solution rather than a pure software solution.
Pieter: So Cathy, you've been mentioning how reinforcement learning is a really powerful tool at the core of what you're doing here. Can you say a bit more about what is reinforcement learning and how does it play a role in this work?
Cathy: Yeah, great question. So reinforcement learning is essentially this paradigm at the intersection of machine learning and also control, and it is essentially about how agents learn from experience and in particular through trial and error. Agents will take some actions in the world, interact with the world and get some feedback signals, some sense of how well it's doing, which we call a reward. And over time, the idea is that the agents are trying to optimize some objective. So in our case, this would be, say, the trying to optimize the speed of traffic in a system or trying to lower emissions or trying to prevent accidents. And this is a nice framework from which we can start to ask questions about, you know what, if the agents saw the whole system, then could they achieve congestion mitigation? Could they achieve congestion mitigation if all vehicles are autonomous? If five percent of vehicles are autonomous and we can sort of ask these questions in a scientific manner and conduct experiments like we are now, we are scientists and we're using this. There's this framework to answer our scientific questions around whether these autonomous vehicles can achieve some capability. And the capability could be, for instance, like I mentioned, the degree of autonomy like zero two on a person. But it could also be along the lines of how much of the world does the does the vehicle observe? Does observe all the vehicles as the only observe local information? Does it observe some some other privileged information, like the density of the system and the how much the the vehicles observe have direct implications on the cost of deployment, on timelines of deployment, on whether we need to dig up for the roads and install sensors or not. And so through trial and error on our side, on the researchers side, we can we can sort of ask we can start to restrict the observations that the agents can, can sort of see in the system and we can try to get at what is the smallest amount that the system the other agents need to see and still be able to achieve some sort of capability. And so what we have we find in some of our model highway settings is that just local sensing, just observing the vehicle's own speed, the distance of the vehicle in front of it, the speed of the vehicle in front of it appears to be sufficient to eliminate congestion theoretical optimal levels, which we find remarkable, and it still requires further investigation to understand why this is possible. This is not something that we expect to be possible.
Pieter: Wow, that's so interesting. I can also imagine these agents that run reinforcement learning, they become better over time, that maybe if you expose them to more information, that they learn different strategies than if you expose them to less and somehow with less information, they figure out something more clever that still achieves a very good behavior.
Cathy: Yeah, that's true. And I think that's we touched on this earlier about the maybe the use of blinkers that the agents are have been taking advantage of, which we didn't like. We didn't and we didn't realize that the simulator models like human response to blinkers, which is pretty interesting.
Pieter: Now you've been working on this for a while, and I'm curious as you've been working on transportation and machine learning for transportation, how has your vision evolved of how you see this play out? And where do you see the future of transportation as a whole, but also especially the role of machine learning and AI in transportation?
Cathy: So I would say early on, I was really focused on very specifically understanding the role of self-driving cars in impacting the transportation system and in particular with congestion as a starting point. And I think along the way, I realized how powerful this reinforcement learning tool is. And so I think that has opened up. I’m very excited more generally about how reinforcement learning can help us accelerate the discovery of insights because I think I was trying answer one question about self-driving cars and interactions with this complex transportation system. And now I'm sort of envisioning that there may be other complex dynamical systems that it will take us as researchers a really, long time to reason through ourselves. But reinforcement learning can help us discover these capabilities just like what we have done with congestion mitigation, as we are doing now with emergency vehicles, with other kinds of scenarios. We can potentially use this tool to help us to help sort of point the way to what is even possible with the future in this case, mobility systems. What should we be designing for? One of my motivations is the world is changing quickly, which is a really interesting juxtaposition on infrastructure decisions. Infrastructure decisions are long lasting and basically permanent, like if you build a road, we can move it. But it might take 50 years because we must get everyone to agree to move the road so effectively. A lot of our decisions are permanent where we put sensors and we put them here where we don't have the money to put them somewhere else. So we have to be really careful about how we make infrastructure decisions. At the same time, the world is changing, so we can't just use the requirements of mobility systems today to make those decisions. And so this is one aspect where I'm really excited about the role of reinforcement learning as a modeling paradigm because we can model what we anticipate we will need for future systems. What level of congestion do we want? Are we OK with in the future? What level of safety performance? What level of emissions? Of course, we can't model everything. The future is very uncertain, but there are sources of uncertainty that we at least know we should plan for. So, for instance, we should plan for self-driving cars and other emerging technologies. We should plan for electric vehicles. We should plan for climate change and other kinds of disruptions. We should plan for changes in consumer preferences. We have an aging population in the US and in many parts of the world, and a lot of requirements will change. These are sort of known unknowns. So we know we need to plan for these and potentially leverage reinforcement learning as a modeling tool to help us understand how we can better plan for this. The other aspect that I'm really interested in is how big of a role that learning can play in these really large scale systems.
A lot of our work so far has been restricted by scale reinforcement. It's very, very data hungry. Mobility systems, cities, urban systems are huge. We're currently able to study like one intersection, one road, two intersections. That's like if you look at the research, that's also where we are right now. Like a city has like typically like twenty thousand intersections and we are nowhere close. There’s a really exciting opportunity and we're also doing some work in this space in terms of how do we address scalability?
Pieter: Now I'm curious if you look at U.S. cities or even beyond cities in the world, you're saying this infrastructure is often permanent or it stays the same for a long time. Are there any cities that you think are a great model and that will be better off with their current infrastructure?
Cathy: That's a good question. I'm certainly not an expert on international infrastructure, but I am aware of there are some places that are sort of just building cities from scratch, and that's admirable. So I believe this is happening in China and a couple of places. I think it's a nice opportunity to at least capture what we know now. I mean, I think they'll have the same problem as as the world changes
Pieter: When people think about the future of transportation. One thing that comes to mind for a lot of people is flying cars. I mean, that's at the heart of a lot of science fiction and so forth. So what's your take on that?
Cathy: I think it's exciting. I mean, I'd love to receive my Amazon packages by drone, and it's really compelling. You know, if we can lift our 2D traffic into 3D, it's possible that it'll take a long time before we hit congestion again. Oh, so that would be one way to to avoid congestion. I do have a concern around the energy cost of lifting all the vehicles up in the air. But I think it's very exciting.
Pieter: And then the other direction is the tunnels, right? I mean, it’s the Boring company’s next thing it is supposed to make.
Cathy: I love that idea. I mean, it is also an infrastructure solution, right? You can't just have software. You have to go dig that tunnel. So I mean, I I love the idea. I love the name.
Pieter: Yeah. There isn't too many of them yet. But is this something you've been thinking about yourself? And how would the traffic landscape change? Where would the optimal channels be? Or is this a little too far out?
Cathy: It's a really good question. I have not thought too much about flying cars or underground cars too much myself. I would say it’s an area of research that I'm really interested in, which has to do with learning for combinatorial optimization. And this is the idea of trying to devise some general-purpose algorithms for combinatorial problems. So for instance, when we do have flying cars and underground tunnels, these are going to be different network optimization problems than we have today. And these are all going to be large scale. They're going to be difficult to solve. We're going to need a lot of heuristic methods because these are generally computationally intractable problems. So I'm really excited about trying to anticipate the pace of change and design algorithms such that when the change starts to occur, we have algorithms already ready. Typically for combinatorial optimization methods, we may need years or decades to perfect a heuristic that is specialized to the problem, especially for large scale problems. So we've started some work and this is actually accepted to this year's NeurIPS as a spotlight talk that is on
“How do we design algorithms for large scale combinatorial problems, in particular vehicle routing problem?” So we basically designed an algorithm that’s a supervisory approach that helps to accelerate existing heuristics by up to 100 times. It’s a decomposition oriented method. This is a first step for us to produce some algorithms that aren't so reliant on hand designed heuristics that may take decades to form when these new problem variants arise.
Pieter: Now, let’s talk about COVID 19…during the early days, much of the world abruptly stopped driving because we were all confined to our homes. Did this unique period of driver behavior create any sort of data that was helpful to your research and have any type of impact?
Cathy: Not to me personally but I know a lot of colleagues that got involved in in these new traffic dynamics for other solutions related to congestion. I am also a fan of what's called a mobility solution related to work from home like telecommuting VR. If we can make the meeting room perfect and then we don't need to commute. That's a wonderful solution to to congestion. We're unfortunately not there yet, although we did a remarkable job of transitioning to virtual in general.
Pieter: So Cathy, you pretty recently finished your Ph.D. and became a professor at MIT. It's a really exciting career path that I imagine many people are aspiring to. I'm curious, you have any advice for PhD students or any advice for people who are starting as professors somewhere?
Cathy: For people starting out in grad school. I highly recommend going to as many seminars during your first year and meeting as many people in your cohort as possible. These are the people who are going to be with you for the next five or however many years, and you will learn way more from them more than the half hour or so you spend with your advisor every week. So that's what I would say for folks starting out in grad school, for folks starting out as faculty. So I'm in my starting my third year as a faculty member now, I'm actually going to echo some advice that Peter actually gave me before I started, which is to really not worry about anything just and just to focus on doing great research. And I didn't actually understand what he meant at the time, but now I very much do. There are so many, so many kinds of things that compete for your attention when starting out as a faculty member. The biggest one is teaching and preparing new courses. There is also fundraising. There’s getting to familiarize yourself with a new environment. There's recruiting students. There’s all sorts of tasks that one is not necessarily trained for or prepared for. That sort of takes away time from research. But what I found in even the last couple of years is that if one focuses on doing great research, the other things sort of work themselves out. And so its not worth worrying about those things. So thank you, Peter.
Pieter: Glad that advice had worked out, Cathy. That's been propagated so now everyone has it. Well, that said, Cathy, thank you so much for coming on. This was really fun.
Cathy: Thank you so much, Pieter. This was really fun.