Michael Rubenstein describes how he has taken his robotics research from theory into practice by building cheap and small robots, 1024 of them to be exact. Michael is an assistant professor with a joint appointment in the Electrical Engineering and Computer Science and Mechanical Engineering departments at Northwestern University. His research interest is to advance the control and design of multi-robot systems, enabling their use instead of traditional single robots and to solve problems for which traditional robots are not suitable. Using these multi-robot systems can offer more parallelism, adaptability, and fault tolerance when compared to a traditional single robot. He is also interested in investigating how new technologies will allow for more capable multi-robot systems, and how these technologies impact the design of multi-robot algorithms, especially as these systems begin to number in the hundreds, thousands, or even millions of robots. Michael received his PhD from the University of Southern California in Computer Science. After his PhD, he was a postdoctoral researcher in the Self-Organizing Systems Research Group at Harvard University. He completed his undergraduate degree in Electrical Engineering at Purdue University. Michael first talks to our host Per Sjöborg about his PhD research in algorithms for modular self-reconfigurable robots, at the University of Southern California for Dr. Wei-Min Shen. He then shares his work on the Kilobot project and some of the challenges involved with building 1024 robots and how you can learn different things from actually building the robots than from a simulation. This project helped him fine-tune the algorithms from his earlier research. By working on Kilobot, Michael also learned how to make cheap robots, which fits the educational market well. He talks about this and the robots he has created that can be programmed by school children at robotics summer camps. The Kilobot work was done in the Self-Organizing Systems Research Group at Harvard University. Michael is now faculty at Northwestern University
This podcast is part of the Wevolver network. Wevolver is a platform & community providing engineers informative content to help them innovate.
Learn more at Wevolver.com
If you are interested in sponsoring the podcast, you can contact us at richard@wevolver.com
Per Sjöborg: Welcome to the podcast version of Robots in Depth Episode 5 with Michael Rubenstein in cooperation with Wevolver. Today I am honored to have a Michael Rubenstein here and we're going to talk about a lot of exciting stuff, thousand twenty four robots, science papers, papers in science magazine but we're going to start with how he got into robotics. How did you get, how did you notice that you can build this electromechanical devices that you could program.
Michael: I think originally in high school I was interested in electronics one of the natural extensions that is making things move with electronics. I think I one of the first things I did was I made a rocket that was able to steer itself with motors and then when I went into college I studied electronics as well. I was part of a solar car racing team so that's pretty close to a robotic system. You have you know actuators, sensors, power cells, things like this and it was really interesting to optimize the system to get it to compete against other people who do solar racing as well.
Per: Was it in a desert environment then I would presume?
Michael: I did this at my undergrad at Purdue so it's kind of Indiana. It’s not really desert environment but some of the races that we do, the race I did was from Chicago to LA along route 66. Racing 2,000 miles on solar power alone. They do some big races in Australia where it's The World Solar Challenge they go across Australia which is a big race as well.
Per: That brings in the whole complexity of doing a system with that kind of durability.
Michael: I guess yes. You’re trying to make a system that's robust to going across the country and rain and dust and cold weather and something like that. I like that but you also you're optimizing other things that kind of compete against that so the car needs to be low weight. Every pound of weight that you add to the car it takes a certain amount energy to push that down the road but a lightweight car is fragile and breaks easily. You’re kind of trying to weigh these two design choices.
Per: Then you went on to do other things I guess.
Michael: Yes so I did my PhD at University of Southern California and I worked with Wei-Min Shen and we worked on modular reconfigurable robots which are these cool robots that are kind of brick sized robots or smaller and they connect to each other kind of like Legos can connect each other. They can move around and change their shape that way. You have a robot that's made of tens or hundreds or thousands these little robots that are attached together just like cells in a multi cellular creature.
Per: This is my favorite area. It’s like you could do whatever you want. It’s Lego for grownup. It’s Lego for engineers.
Michael: I did that and one of the things I noticed when I had worked on my thesis is that the way these things, the way these robots interact with the environment is mainly by their shape. You think of any animal, any multi cellular animal and their shape tells you what they can do the environment. Birds can fly. Giraffes can reach tall things. Shape is one of the most important parts of a reconfigurable robot. I looked at how do you actually control shape in a reconfigurable robot. The idea is that you'd like to say I want this huge massive arbitrary large number of robots form an exact shape that I tell them to. Give them a shape that you don’t want to control every individual robot say you go here, you go here, you go here you just want to give them a shape and have them automatically form that shape for you.
Per: When we get there the world is going to look very different. That’s going to be like bringing the transistor into the world or bringing electric power into the world. It’s just going to be a game changer we won't believe but we're quite far from that sadly aren't we?
Michael: Yes so I guess that leads to the kilobot project which is so my thesis I wrote software and simulations where you can draw a shape in the computer and give it to all our simulated robots and they form that shape. I wanted to do that with real robots and since if you're making a shape out of individual like pixels or units the more you have the more fidelity of the shape you'll have so the cleaner the edges will be, the more precise looking. I mean you can always approximate anything with a circle right but if you could approximate it with a thousand circles it's a better approximation.
Per: Anything digital so this is actually a digital matter.
Michael: The idea is extended to the Kilobots where we draw a shape in the computer and we gave it to all the robots at once and they form that shape in the environment. It’s kind of realizing the idea that we had from the simulation and things that are interesting in modular robots but obviously kilobots aren't as capable even like a simple modular robot is because they can attach each other. They’ll form the shape but you can't reach out and pick up the shape. It will fall apart. You just pick up individual robots.
Per: Yes but it's a very important step down this road towards the things that you both you and I dream of.
Michael: A lot of the things that we investigate like the control and the algorithms and building the robots can be applied to more complex robots that can form shapes that are more useful is what we hope.
Per: You call it the Kilobots project and I know what the Kilobots project it is actually extremely cool of you saying I'm going to build a 1024 hence Kilobots robots. That’s just an enormous amount of robots and there had to be an enormous challenge for you to really remove everything that was generally possible to do without.
Michael: It's like going back to the solar car issues. You have two designs that are competing against each other. I want to build a robot that does interesting things. You'd want to have like a laser scanner on it and accelerometers and GPS and other things but also want to build a thousand them so they have to be low cost. The balance is I want to build capable of robots but I also want to build cheap robots because I want to build a lot of them so figure out what can I get rid of on the sensing and the capabilities of the robot and still have enough capabilities that I can do some of the algorithms I'm interested in.
Per: They also have to be very reliable because if you have two robots and one of them breaks you work and you fix it but if you have a 1024 of them and 10% of those break you have a huge issue so they have to be durable and robust in one sense.
Michael: They have to be durable robust in hardware but you never have a perfect hardware so even if you build a robot that is really good one in a thousand will still break. Your algorithm you have to deal with those hardware failures in algorithms as well.
Per: Then you have to simply throw away that robot that's dead.
Michael: Or you let it carry on and it just doesn't cause any large-scale damage.
Per: The shape will still form so to speak. That’s really cool. You then with this Kilobots showed that the theoretical framework you developed earlier actually is verifiable so to speak.
Michael: The simulation that we wrote originally for my thesis didn't really model some of the errors that we detected in the real hardware.
Per: You went back and update that.
Michael: We went back and modified the algorithm for shape formation to adapt to this new hardware. We couldn't take the algorithm from my original thesis and put in the hardware. It wouldn’t have worked so we had to modify it to take into account more errors and different types of errors that the robots have that we didn't expect which is one of the important things of building a robot. People always say why don't you just simulate a thousand robots.
Per: Because you learn a lot of actually building them.
Michael: Simulation doesn't test any of the assumptions you make about the simulation so my original simulation I assume certain things about something sensing that in real life didn't actually end up happening on the real robots.
Per: They say that simulation is doomed to success. It will always work in simulation and then when you bring it into the real world you learn, you take the next step because simulation is an important step.
Michael: Yes and we could feedback so now we know a lot more about what the robots are capable and we could add that to the simulation later on.
Per: Then the Kilobot is now a commercial project. You can go out and buy them if anybody wants to play with their own set of 1024 robots they can actually do that and they can do research with them and without having to build the platform which is a huge undertaking.
Michael: I know Roderich Gross from Sheffield, I believe he bought 900 robots or so. He has a huge group of robots and he didn't have to worry about design for the robots as much as you would if you built them yourself. Hopefully that's useful for him in the research.
Per: He's going to be doing valuable research so much earlier than he would have if he would have to start from scratch.
Michael: That’s the hope.
Per: You seem to be both a theoretical guy and a hardware guy but most of us are either or and if you're a software guy building 1024 robots will just be daunting I believe. Now we have this amazing platform and that's going to lead to so much great stuff and you then wrote this up. This is became a paper in science.
Michael: We spent a long time getting the hardware working, making sure all the bugs are worked out of the hardware and then we started working on the algorithm for shape formation. I originally tried my thesis work and a lot of things didn't work so we've modified the algorithm until we were able to get the shape formation. We thought it was pretty cool so we made a big paper about it. We also spent a lot of time analyzing the control for the robots so we're able to prove that it is guaranteed to form the shapes that we draw in the computer.
Per: That's very important. There’s no doubt.
Michael: The proof, the theoretical proof we made a set of assumptions about what the shape looks like, how the robots behave and we guaranteed with those assumptions that it will always form a shape but if you give me a shape that's outside those assumptions, it's really thin, has a hole in it, things like that it would not be able to form.
Per: The assumptions are probably pretty basic like it has to be enough robots there and stuff like that.
Michael: At least in the proof we assume that there is enough robots to always form a shape.
Per: Otherwise it will be pointless.
Michael: The algorithm that there's too few robots it'll partially form the shape. If there's too many robots it'll form the shape and then one robot runs around the shape says that the shape is formed and can stop afterwards. They know if the shame is complete.
Per: That's cool so it doesn't have to be the exact number of robots.
Michael: It's really hard to do that to take the exact number of robots. In my thesis work out it was a little bit better of an algorithm where it forms the shape at a scale proportional to number of robots. It forms of shape and if it's too big for the number of robots it'll shrink the size of shape until it fits the robots perfectly. You add more robots it'll make it bigger. It’s a little bit nicer. You don't to worry about does the scale, the shape match the number robots perfectly.
Per: You just get the size the number of robots corresponding to. That’s pretty cool.
Michael: When I was doing the work for science we had a fixed shape scale. I had to kind of guess what is the density of robots per unit area and how big is the shape going to form in the table. I would scale a shape to fit all 1024 robots in it but there's ways of doing that automatically which are a little bit more advanced but we kind of did the most basic shape formation we could initially.
Per: Then coming from (0:11:58.7) lab and coming from the connected three dimensional self-configuring because in one way you could say that Kilobot is a self-reconfiguring robot, it's a self-reconfiguring shape. Do you see us coming to the 3D world and the connected world anytime soon?
Michael: I think it's easier to go to the connected world than the 3D world so you could do a 2D connected robot. That’s relatively simple to do depending how strong the connection needs to be it modifies the robot. You adding a connected to robot will double the complexity of the robot, double everything about the robot basically. Modular reconfigurable robots are basically built around their connectors and their motors. It’s hard to add a connector to a robot after you built it. It’s better to add a connector and then build a robot around it.
Per: Yes certainly but that's eventually where we wall want to go.
Michael: I mean this is a test for some futuristic system like this.
Per: That's just amazing that we are on the road towards that because that's just going to be a huge thing in the world. You then also developed an educational robot now because you kind of got into this doing them cheaply because that makes them available to people.
Michael: One thing about building swarm robots is I've become very good at making really cheap robots. Some of these skills actually work towards making a good education robot. I mean the robot that I built recently costs $10 so for like a couple cups of coffee you can have every student in the class have a robot.
Per: That really gives each their own robot and I understand they even participate in assembling the robots.
Michael: It helps them feel ownership to the robot. It improves the learning is what we found.
Per: Ties them emotionally to it.
Michael: They build a robot. They put a sticker on it. They can draw on the robot, put googly eyes on it, things like this.
Per: They can also program it. They can make it do a few things because it's also what we could call a Bristlebot. It also navigates on a flat surface with these vibration motors from cell phones that are producing it.
Michael: The main reason we did that is that cell phone vibrators are much cheaper than a motor with a wheel attached to it.
Per: It works the same way.
Michael: it works. It’s a little bit slower but when you're learning from nothing to being able to do wall following, line falling the speed of the robot isn't as important as the learning processes.
Per: It's the learning process and making them instead of costing $50 it's costing $10. That difference is huge in the reach we have out to children and allowing them to play with the robot. Sometimes children, there's accidents and the robot breaks it's much better to break a $10 robot than a $50 or $150 robot.
Michael: The robot is relatively simple. It’s very robust. It’s hard to break them. You can drop them and they're fine.
Per: You're bringing this project out into to small scale tests now with 50 or students at one time.
Michael: Me and some people in Radica Doug Paul's lab have been working on developing courses that work with these robots. You have a week-long course for students like say 5th to 8th graders and you give them all a robot and they learn. They start from nothing in a week later they'll learn how to use the robot, do interesting things. They have the basic ideas how robots work, how to program robots, how to program in general. We started this summer with there's a company that runs summer camps and we taught 40 students how to use these robots and a week-long course for each student.
Per: Amazing that that age how old did you say they were?
Michael: They were 5th to 8th graders so probably 10 to 13 is my guess.
Per: They incorporated robotics in the teaching of those. I mean we didn't even have computers at that age when I went to school. It’s amazing what we are developing.
Michael: The hope is that this encourages students and gives them experience in software, how to program, how to use robots and they're interested in learning science and math and engineering.
Per: Breaks down the barrier. They’re going to be really interested in building a better robots.
Michael: That's the hope.
Per: That's really cool and this project is also going to be spread further I hope. We’re going to see it commercially you think?
Michael: We're looking at commercially having someone produce these robots for us because I don't want to build education robots. I built a hundred this summer and it's kind of painful building. Hopefully someone will build them for us and we're working on further developing this course so hopefully next summer more students will use it.
Per: Yes really cool and I think that having it commercially available it's not going to be $10 anymore because that's I guess the material and basic costs.
Michael: It's probably double or triple that cost if it's commercially available.
Per: Yes but it's still a very cheap robot and very durable. Now you give them to the students but if you have them in a lab environment where the students come in and use them for a subset of the time they're going to survive in that environment. I think that's just that's just great. Thank you very much for being part of the show and I hope to hear from you soon again. Just a couple of science papers and we'll be right here again talking to you.
Michael: All right thank you.
Per: Thank you. I hope you liked 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 is how 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. Until the next episode thank you for listening.
END OF TRANSCRIPT