Spring M. Berman is an associate professor of Mechanical and Aerospace Engineering and graduate faculty in Computer Science and Exploration Systems Design at Arizona State University (ASU).
She is also a core faculty affiliate of the Global Security Initiative (GSI) at ASU and Associate Director of the Center for Human, Artificial Intelligence, and Robot Teaming (CHART) within GSI. Before joining ASU in 2012, she was a postdoctoral researcher in Computer Science at Harvard University (2010-2012).
She received the M.S.E. and Ph.D. degrees in Mechanical Engineering and Applied Mechanics from the University of Pennsylvania in 2008 and 2010, respectively. She received the B.S.E. degree in Mechanical and Aerospace Engineering and a Certificate in Robotics and Intelligent Systems from Princeton University in 2005.
Her research interests include the modeling, analysis, control, and optimization of robotic swarms; the analysis of collective behaviors in biology; and biologically-inspired control of distributed systems.
Per Sjöborg, host of the Robots In Depth podcast, interviews Spring Berman to learn more about her work as an associate professor in at Arizona State University. Below is a transcript of the interview.
Per Sjöborg: Welcome to the podcast version of Robots in Depth in this launch episode with Spring Berman in cooperation with Wevolver. Today I'm with one my friends Spring Berman here. She’s from Arizona State University and we're going to talk about like everything, everything in robotics. I'm so fortunate to have the opportunity to speak to you about what's going on in robotics and I find it so fascinating. Let’s start where we always start. How did you get into robotics? Why robotics? What was the first exposure to building mechanical things that could move around?
Spring: I always liked building with construction sets as a child like the Lincoln Logs and this little architecture kit so I liked design and building back then. Then I got very interested in math and physics on the theoretical side. I went to an undergraduate program in mechanical and aerospace engineering at Princeton University. I think that was where I first had my real exposure to robotics because there was a professor there Naomi Leonard who was working on underwater gliders and how to coordinate them to take measurements efficiently about the ocean like temperature and salinity biological bioluminescence. I worked in her lab for a few years and I became fascinated by this idea of using natural swarms or natural groups that exhibit these collective behaviors using that as inspiration for multi robot control strategies.
Per: Like emergent properties.
Spring: Emergent properties yes.
Per: Very hard to control I understand, emergent properties are something we usually try to keep out of our systems I understand.
Spring: Undesirable emergent properties...
Per: It's like herding cats.
Spring: It's such a challenge. That’s what I've been working on for since undergraduate but I think that's where I first got exposed to robotics was this underwater glider project.
Per: Was in deep underwater or was it surface? We have seen these gliders that have a surface part and a kind of suspended part 15 or 20 meters that use the waves to move ahead. Could you explain a little bit more about those gliders?
Spring: These were buoyancy driven vehicles so they move in the sawtooth pattern up and down. They can go out taking measurements for long periods of time.
Per: I understand that's the benefit of a glider. We talking weeks and months or even longer times rather than maybe a day or so on in a battery-powered vehicle. It is magnitudes more.
Spring: Yes so you can do things like persistent monitoring of the ocean. I think what fascinated me about that project was the idea that you're developing these mathematical and computational tools to control groups of vehicles not just a single vehicle but groups of vehicles to perform some kind of desired macroscopic behavior. The idea that you can see this in nature when you look at ant colonies or schools of fish or flocks of birds. There’s no leader in the swarm but everybody is, basically every member is following local information about their neighbors, about the environment and just through those interactions you get this very complex emergent behavior. I thought it was really interesting that we could possibly design those emergent behaviors by designing the controllers of the robot.
Per: Yes because if the birds had a leader that'd be just chaos. You couldn’t achieve what you want to achieve with a leader you have to design it without one.
Spring: I think there are some strategies that do use leaders like leader follower strategies but for the most part if you want a robust system where there's not a single point of failure then you need these decentralized strategies that are robust to disturbance and failures.
Per: can you talk a bit more about the development of these kind of leaderless systems?
Spring: I guess the challenge is guaranteeing that they'll do what you want them to do and that's a major challenge that's been approached by researchers in different fields. For instance some people use control theoretic concepts to guarantee stability. For instance stability of a formation that a flock will converge to a certain formation or there's methods from computer science from the study of formal methods to verify the correctness of behaviors using tools like linear temporal logic. These tools come from different fields and so one challenge right now is to kind of merge them and or to create you know a unified framework for understanding and controlling swarm behaviors. One challenge is that sometimes it's hard to develop an accurate model of the system because you can use, my group in particular uses models from chemical reaction networks so modelling the individual members as molecules that are colliding and interacting with each other. Then they can switch their tasks based on their interactions.
These chemical models that abstract to differential equation models are very useful in describing the expected or the mean population dynamics. There are fluctuations around that expected behavior based on things like the number of agents, their sensing ranges, their capabilities. You also want to be able to model that variation and to hopefully produce bounds on the performance so confidence bounds that you can be certain that the swarm will execute this behavior with this degree of confidence. That’s an ongoing, that's an open problem that we're still working on is how do you characterize that variance. What is it a function of in terms of the robot programs and their physical capabilities? Then of course you have issues like they're exploring an uncertain environment you don't know what obstacles there are and there are what features. Some of them might fail. They might get stuck somewhere. These things are very difficult to model. We try to address that by also introducing mapping strategies where we can deploy this and then get information back from them that we can use to reconstruct the environment to modify the dynamics in our model so that it's more accurate. That’s another problem that we're addressing because we do we do use these both differential equation and stochastic models to describe the robot behaviors.
Per: Say you're doing this in the air or you're doing it on the water how do you convince regulatory authorities that they're not going to do anything bad?
Spring: That's an excellent point and that is something, there's ongoing research that's being funded for instance by the Department of Defense because they really want to know things like if your adversary has a swarm and it starts interacting with your swarm can you guarantee that your swarm is not going to go awry or that it's not going to kill your swarm so things like that where you actually have interactions between different swarms as well as between a swarm in an environment. There’s two different problems like can you guarantee that it will go into this desired set of behaviors or can you guarantee that it will avoid this undesirable set. Those are two questions and they are excellent questions. They’re ones that we're trying to work on using our methods.
Per: Where would you say the field is in maturity? I mean I guess you've been working on this for quite a while and others who might worked before you. Where are the field now? Is it kind of in the beginning or in the middle?
Spring: I've kind of been a part of this field for a while at least 10 or 12 years. I've seen it developing and I think one thing that's been enabling the swarm technologies is the development of these really low-cost, small robot platforms that you can manufacture quickly and cheaply and then deploy in large numbers. There’s been a whole variety of platforms so aerial, ground, surface, underwater that have been coming out so that's kind of exploded recently. That’s on the just the technological feasibility side that it's actually becoming feasible to use these but I think one of the main challenges is the control problem. How do you control them and have them, I think the next frontier now is developing resilient, autonomous collective systems. That is that you can deploy them into an unknown environment possibly without GPS or reliable communication and that they can adapt to novel circumstances so unexpected situations, that they can even learn how to deal with these different scenarios so you have adaptation and learning and that they can persist for long periods of time like a single glider but now you have a swarm of a thousand that is robust to failures, that can adapt and learn. That’s really the next challenge and I think it's still hard to do that because we are not there yet in terms of the broad applicability of control methods and verification methods to all kinds of circumstances. That’s still an open problem. People are working on that frontier of research. That’s why a lot of it is still confined to research labs for smaller groups and simulations basically, multi-agent simulations.
Per: The area is kind of maybe it's towards the end of the beginning but it's not that far along. It’s simply that complicated, that hard.
Spring: There's been so many interesting ideas you know from coming from different communities from applied math, kinetic modelling, study of self-organization and biological systems, control theory, computer science, formal methods. There’s been a lot of ideas. There’s been a lot of collaborative work and discussions and workshops. That’s been great because it's kind of been this knowledge exchange between different fields. I think the next phase of that is to combine them in these rigorous ways, in these analytically rigorous ways and where you can't do that develop computational techniques that can answer the question if someone says I'm deploying this swarm and I want it to not do this undesirable behavior that you can assure them that yes, this is ready for deployment and we can guarantee this and that.
Per: Because otherwise you simply won't be allowed to deploy them. These will be deployed then somewhere in the future in in large groups I would presume not dozens but hundreds of thousands. Where do you see the application? I mean say we've solved these problems. We have hardware. It’s available. What would be your dream application?
Spring: I think that there's a lot of potential in areas like environmental monitoring humanitarian operations like search and rescue, disaster response where you have to cover a large area, possibly manipulate things in the environment. Other applications might be for exploration in environments that are inaccessible to humans like deep underwater or in outer space and other planets, doing construction and manufacturing and assembly in these environments where humans can't necessarily go.
Per: Or don't want to go.
Spring: exactly, yes. There’s other applications that are also new but at a very different scale so at the micro nano scale. There are people working on micro nano systems such as nanoparticles or other types of controllable systems at the micro nano scale.
Per: That is difficult in a whole other level. If you build a mechanical device as your swarm member I'm thinking you have an easy day because if you go to the nano scale everything becomes just weird.
Spring: You have effects that you don't have in larger scales.
Per: We have a hard time relating to them because we never experienced them. We are not nano scale entities on our own.
Spring: Then people are developing these kind of hybrid platforms even with biological components like cells or bacteria plus synthetic components so that you can kind of exploit the propulsion or other behaviors of the biological component and don't necessarily have to engineer that which is really still a challenge at this scale.
Per: It's better to use what nature already provided.
Spring: Yes, in some case. The applications there at that level there are a lot of biomedical applications. Those include things like targeted drug delivery for cancer therapy. Getting the chemotherapy directly into tumors.
Per: But into the rest of the body forcing intolerable side effects. I mean of course everybody supports that. That’s really interesting. I would presume with that when we have so many systems we're not going to go really big. Really small is one thing and medium-sized is one thing but we're not going to see them sizes of buses. We don’t want 6000 buses flying around in the air.
Spring: That starts getting into autonomous driverless vehicles and managing traffic.
Per: Of course there’s applications in logistics and transportation.
Spring: I think that may be a different class of problem, well different constraints obviously because you definitely don't want collisions. You don't want collisions in swarms anyway but it just depends on the platform and how resilient they are.
Per: But I think that when you use them for monitoring especially if there's a small chip unit you could really monitor the environment in a way that is magnitudes more exact and spread out than what we have today. Usually we have a few sensors telling us at this point this is what's going on but we don't know what's going on in the bigger picture.
Spring: That's true. You can get a much richer data from a swarm. You can get data as a function of both space and time, dynamically changing data. Then you can also have your sensors adapt to where the interesting features are. Maybe they can detect a gradient or you know some maximum in a field and they can kind of home in on that feature and get more detailed data about that feature.
Per: Fish for instance is a huge resource for mankind. That has to be managed and also has to be used wisely. Do you see swarms interacting with swarms of biological animals? Could we have swarms in, could we have every other fish be an artificial fish and simply lead them to the net and we could determine that these fish are ready for harvesting and these fish should stay in the ocean and grow further? Could we interact with animals to better manage them?
Spring: There's definitely been work going on that kind of combines these artificial animals with real animals and investigates their interactions. It’s also been done with cockroaches to influence the real cockroach to aggregate in some area near a robot that may not look like a cockroach but it has the pheromone scent of a cockroach.
Per: Do you think that we're going to see swarms interacting with natural swarms anytime soon? Is that an emerging field?
Spring: It's hard for me to say how soon. I know there's definitely been work on it for different animals including birds and fish and insects. There’s definitely interest from the community that studies these biological aggregations. They’re very interested in robotics from the point of view of well, can they test out different plausible biological behaviors to see if that's what the animals are doing.
Per: When we can design a robot that does what the natural animal does we know how they potentially could do it. We don't know that's actually how they do it but that's at least one way.
Spring: It's a test bed that you can very lots of conditions whereas a colony of ants might be more finicky in laboratory conditions. Another tool that robots provide for the biologists is a way to perturb the biological system and see how it reacts. If you introduce a predator or like a nest mate that is actually artificial and you program it with certain behaviors how does the real colony, the natural colony react. You can kind of develop models of the animals’ behaviors and interactions based on that. I think as we develop these robotic tools for biologists and develop more sophisticated models of the biological swarms then we can think more about controlling them so that they avoid aircraft or leading them like a Pied Piper away from insect infestations away from somebody's house.
Per: Instead of using chemicals and other things that's really bad in many ways. We could also use these swarms to take care of many things in agriculture I would believe. We can detect moisture levels in soils. Should we water? Should we not water? Do we need to use pest control in some way? Is this plant ready for harvesting?
Spring: I know that robotics in agriculture is becoming really big now. I mean you can use drones and other vehicles for monitoring the plants. I was on a postdoc position on a project called Robo Bees where they were developing robotic bees for things like crop pollination and that was really as a supplement to natural bees which were dying off in large numbers.
Per: We have seen that. It’s a really disturbing trend.
Spring: That was really to help supplement for doing this really important pollination activity that produces so much of our food.
Per: We're talking dozens of percent of what we every day. I mean it's a very significant chunk. We don't want to lose it. Very interesting, very challenging problem I mean if you look at robotics in general which could be said to be challenging enough. Doing it with hundreds of thousands or millions of units and this just adds and the poor nano guys they do it even harder. They really found a nut to crack there. That’s just crazy stuff but they can also achieve things that are just amazing fighting cancer, monitoring the environment. Fascinating, it's been so interesting to talk to you and I'm sure we're going to do it again soon. Thank you very much for taking the time to do an interview.
Spring: Thank you. Thank you for interviewing me.
Per: I hoped you like this episode of the podcast version of Robots in Depth. This episode is produced together with Wevolver. Wevolver is a platform and community providing engineers informative content to help them innovate. It helps engineers stay cutting edge. Aptomica is the founding sponsor for Robots in Depth. Aptomica runs anything in modular robotics. Dream, rent, build. Visit Aptomica.com to connect. I am your host Per Sjöbor.
End of transcript
The Wevolver Robots in Depth podcast is published multiple times per week. If you, or your company is interested in supporting this podcast and reach an audience of professional engineers and engineering students, please contact us at richard [at] wevolver.com