Roderich is also the general chair of ANTS - The 15th International Conference on Swarm Intelligence in Darmstadt this year, June 8-10. Several from my team will be going - check it out!

About Roderich Gross

Roderich Gross is Professor and head of the Resilient Cyber-Physical Systems research lab at the University of Darmstadt. His lab conducts foundational research into information processing of distributed robotics systems. Before joining Darmstadt he was Senior Lecturer in the Department of Automatic Control and Systems Engineering at the University of Sheffield, and a Visiting Scientist at CSAIL, MIT. He received a Ph.D. degree in engineering science in 2007 from Université libre de Bruxelles, and has since been a JSPS Fellow (Tokyo Institute of Technology), a Research Associate (University of Bristol), and a Marie Curie Fellow (EPFL & Unilever).

Transcript (with a few edits)

Sabine Hauert: Welcome, Roderich. It’s great to have you on Swarm Stack.

Roderich Gross: Thanks, Sabine, for having me here.

Sabine Hauert: I’ve always been a big fan of your work. We’ve crossed paths many, many times over the years in the swarm community. And I was quite excited by some of your most recent work looking at how we get robots out into unknown environments and how they can navigate spaces like tunnels and mazes, and build bridges and whatnot. What is your motivation for getting these robots out into these unknown environments, and what are some of the challenges?

Roderich Gross: Oh, there’s lots of different angles to this. This work started when I was at the University of Sheffield. The Peak District is nearby, and there is a particular special place called Chatsworth. There is a maze there, which is absolutely awesome. And it’s quite hard to navigate to get to the centre. So this is one of the motivations for having a robot to help with this.

There’s more practical motivations too, of course. There are caves, underwater caves. There are pipe networks that are getting older and older. These are difficult to navigate - there’s not necessarily maps of these. And a lot of this research popped up when one of my PhD students looked into collective navigation, so how robots move on the go. Suddenly, his research took a turn and just went in this direction here.

Sabine Hauert: This work (“Design for One, Deploy for Many: Navigating Tree Mazes with Multiple Agents” by Jahir Argote-Gerald, Genki Miyauchi, Julian Rau, Paul Trodden and Roderich Groß, IEEE MRS 2025) was nominated for a Best Paper Award at the last MRS. Maybe you want to walk us a little bit through the algorithm. There’s some clever elements, like the robots move forward and then they replace each other and that allows them to explore a little bit more deeply.

Figure: Demonstration with Pi-Puck robots. Credit: “Design for One, Deploy for Many: Navigating Tree Mazes with Multiple Agents” by Jahir Argote-Gerald, Genki Miyauchi, Julian Rau, Paul Trodden and Roderich Groß, IEEE MRS

Roderich Gross: Absolutely, I’m happy to do that. So essentially, let’s start with the maze. We represent this as a graph - you have nodes and edges between them. And when we started, we looked just at tree mazes. These are graphs that don’t have loops, like simple graphs.

And all the robots start on one node of the graph, and they don’t know the graph, so they don’t have a map. They really have to explore it. As they move from one node to the next, more information about this environment gets revealed. And they need to find a goal. There could be one goal or even multiple, and they need to find it. They don’t know where it is. Whenever they discover it for the first time, then they are there, essentially. But one of the challenges in confined spaces is that if you have a lot of robots, there’s not a lot of space for them. So they can’t just all move at the same time, like a single agent would. When one agent moves first, another will follow, and they have to negotiate like who goes first, right? On one hand, there’s a challenge to find the goal, but on the other hand, there’s also this collective crowd agent, to avoid stepping on each other’s toes.

And so, how does the algorithm work? First of all, rather than solving it entirely, some other smart people have solved this problem for a single agent. There are a lot of single agent solvers. One of them would be, as many people know, following the right-hand rule or the left-hand rule, or following the wall. This works in simple environments like trees, and we assume that this exists.

We assume we have any single agent-based solver, and one agent is special, it carries a token. In our case, who that agent is can be democratically elected. For example, the agent with the lowest ID on the current node. This agent has a token and executes a single agent maze solver, and then this single agent maze solver may say “stay in the current node”, or it may say “go in this direction to the next node”, and then essentially the agent looks to see, is there another robot in this other node? If not, if it’s empty, you just go and then you carry on. But if the adjacent node is occupied and there is already an agent on the node that the single agent-based solver wants you to go to, instead of you going, you just pass on the token.

That’s essentially the idea. The actual robot carrying the token may not be able to move there, but the token travels exactly the same route that a single agent, if it was alone, would travel executing the single agent maze solver. And thereby, it’s somehow clear that if the single agent maze solver solves this problem, also our multi-agent maze solver will do it.

The one core challenge we still have to address is that all the agents do not get lost, that their communication network remains connected at all times. That means that once the first agent hits the goal, then all the rest can follow as quickly as possible.

Sabine Hauert: I really like that algorithm, it’s very elegant. It also takes advantage of the swarm really being a collective and coordinating, rather than having every agent trying to figure out how to do every single thing to be able to negotiate these passages. So that’s really clever - and you tested it on real robots!

Roderich Gross: We tested it on real robots. Often, we start in simulation - this graph environment is of course a big simplification. We have tested them using Pi-pucks, which are robots that have a Raspberry Pi on board or on top. We put them into a simple environment. It’s not a real outdoor cave yet, so there’s a lot of challenges left. These robots can’t all start in one location because we would have to stack them, but we put them into a start zone. Then the computer creates a graph, and the information is locally revealed. All the robots in the start zone know the location of the adjacent node to the start node, and then one of them decides to travel there. Once the first robot arrives at this node, which we can check with an overhead camera, we reveal more information.

It is a little bit of a cheat because we are essentially using an overhead camera that sorts out certain problems. For example, that the robots essentially don’t have to perceive the environment and decide that this is a corridor, this isn’t. Rather, the information is revealed: okay, you can go there, you can go to this waypoint and so on. However, it is real robots in a real world, at least - it’s the first step. But there’s a long way to go to bring these robots into caves. And I’m not sure whether we have the full expertise in our groups to go to caves yet, but maybe people watching this video right now who say, hey, we could help you. If so, please write me an email.

Sabine Hauert: That’s really interesting. And for the proxies, you would have a robot with the LIDARs with the necessary sensors to be able to move down that tunnel or down that path, and then when they arrive at a location where they need to make their next decision, their sensors on board would allow them to do that. So the same algorithm would apply, ideally, given that more capable robot.

Roderich Gross: Absolutely. That should work.

Sabine Hauert: That makes a lot of sense as a first step. You’ve also been looking to use subtraction to create structures in 3D to navigate different environments. Do you want to walk us through that strategy? I really like that one as well.

Figure: Active subtraction illustrated using the HyMod platform. Credit: “Self-Reconfiguration in Two-Dimensions via Active Subtraction with Modular Robots”, Matthew D. Hall, Anil, Ozdemir, and Roderich Groß, RSS 2020

Roderich Groß: That’s great. The original papers of this go back to modular robotics, and it’s about how can robots form certain things via self-assembly. The traditional paradigm was always additives, like more robots will be added and you greatly form some interesting shape. It was the group of Daniela Rus that started looking at systems like Miche or Pebbles. These were systems that they were already assembled in a dense object, and you just shake off the robots that you don’t like. They disconnect and you’re shaking the structure. It’s almost like carving out what you like without the knife, because the robots just leave on their own.

A PhD student in my group, Matthew Hall, asked the question: rather than shaking the robots so that they just fall down, which could be quite expensive, what if they have to leave in an orderly fashion? What if they have to plan their paths to get down to the ground? And of course, no robots can be unsupported in flying in the air. They were not drones, so they were all walking on top of each other, etc. And they have to plan to make sure that at all times they are supported. Some robots may have to leave earlier than others and so on. That was quite an exciting study, a first step towards building structures that make physical sense, where essentially you need to be connected to the ground at all times.

We have later extended this to look at the forces that occur in these structures. If you assemble a large cantilever, then this might eventually bend so there’s a lot of forces and connections could break. And then you can add force angles into the models that they can make smart decisions.

Sabine Hauert: How do you imagine it being used? I’m imagining that there’s a gap in an environment and you bring a big block of cubes or something like that, and then they self-assemble into a bridge. How do you imagine this then translating to an environment?

Roderich Gross: There are examples of ants and particular species that are able to form structures, to pull certain things or even to traverse. They do it quite smartly. What’s the best analogy - like potholes on the roads, right? You have these potholes. We engineers, we go and we fix them, and the ants, they just go inside the pothole and others just walk on top of them. And if no other ant on top of them is walking over, they potentially get out and realise they’re no longer needed, that there’s no traffic that is passing through.

These are dynamic structures that form in nature and that could be done in disaster zones where particular locations are inaccessible. We were also thinking about rivers. There was a disaster in Germany where there was a lot of rain and the Ahr river went from 1 metre to 10 metres in the night. Quite horrible - this environment was really, really damaged. I visited the place, there were I think 7 or so bridges - they were all damaged. You couldn’t even cross the river. To have a robot that can make a temporary bridge out of some modules - that could have helped. Then you have a bridge. It doesn’t have to be an extremely good bridge that withstands the challenges of times, but maybe just that is available for a couple of weeks until something better is built - so infrastructure that can be rapidly deployed. That will be cool, having water structures going over the river - maybe floating platforms even, that can move along. That will really be of interest to us.

Sabine Hauert: That makes sense, as emergency structures that just pop up and rapidly organise into the right shape, and into those challenging environments as well. I think the adaptability plays to that.

Roderich Gross: That’s right.

Sabine Hauert: On the adaptability side - another thing that I’ve really enjoyed is some of your latest work, looking at how robots can learn models of their own capabilities. So that if you throw a bunch of robots in an environment, they can learn their own capabilities and adapt, so that they assign themselves to the right task depending on the capabilities that they have. Do you want to say a little bit more about that work? I think it’s really interesting to see how you used drones and had them consider their battery levels, etc.

Figure: On-demand delivery scenario with a fleet of UAVs committing to deliver orders arriving at a fulfilment centre. The fleet is heterogeneous, as the UAVs differ in battery health, and hence in their true energy storage capacities. Credit: “Ready, Bid, Go! On-Demand Delivery Using Fleets of Drones with Unknown, Heterogeneous Energy Storage Constraints”, Mohamed S. Talamali, Genki Miyauchi, Thomas Watteyne, Micael S. Couceiro, Roderich Gross, ArXiv

Roderich Gross: Absolutely. That was driven by Salah Talamali from Sheffield University, who is a brilliant guy. If you look at the study, it was about drones that learn to deploy themselves - they decide, if there is a parcel, can they deliver it or not? I think fundamentally it’s quite broad. It’s about capability learning, as Salah always said it to me.

It’s like how if you have a large collective, often the capabilities in the collective are not uniform. You might have a drone where the battery is more healthy than in another drone. You might have drones where, due to wear and tear, the propellers are a bit worn, and even though you bought identical models, there might be just slightly subtle differences that impact performance.

And what if we don’t know this? We don’t assume that we know everything and from first principle can come up with a perfect model, because there’s always something, some small differences that we may miss. So how can we use machine learning and auction-based methods in combination, where drones bid for jobs and through a trial-and-error process, find out precisely what the capability is. We don’t really care about why the differences are there. There could be a difference in the propellers, in the battery health, and so on. It’s just that we want to get the most out of it, and we are interested in fair usage policies. Can we use hardware in a more uniform way perhaps, avoid robots decaying more than others.

In this particular study, there were these policies that we learn where, given a particular job, a drone can say, given my current charge level – e.g. the battery says 80% - and the parcel has to be delivered a certain distance and has a certain weight, I’m happy to take this job on. The unknown in this specific study - but it can be anything - was the battery health. The drone didn’t know its own battery health. It could be 50%, it could be 100%, and it was not modified during the experiment.

And then Salah did simulations for 7, 8 weeks of simulated time, so for millions of parcels being delivered. He showed that when the drones bid in an auction, each of them has a certain confidence, right? Either they don’t bid, or they bid. And they bid, with high confidence or medium confidence. It is a real value, not just three levels. We found out that the drone that should win the bidding will be the least confident. So that was counterintuitive at first.

Sabine Hauert: It is.

Roderich Gross: Yes, it will be the least confident that win the bid. As long as they want to bid - we’re not giving it to anyone that doesn’t want to bid. The reason is, that if it’s the least confident, it is the one that is really close to its capability model threshold.

Sabine Hauert: Mm.

Roderich Gross: And it is these that most benefit from the feedback - can they do it or not - to perfectly evolve the model to their true capability. You can also think about it this way: why should a drone that is extremely powerful take on a small job, even if they are super confident? You can save this asset for something more demanding in the future, right? So this is another possible interpretation.

Sabine Hauert: That’s really clever. So just a more optimal use of the resources that are available within the collective. On Swarm Stack, we’ve been talking about this vision of lots of robots with different capabilities entirely. There they all delivered parcels, but you could imagine a world where each one of these robots has different capabilities, and they need to perhaps learn their own capabilities and how to communicate that with other robots, so that this heterogeneous collective can do things that are useful. I wonder if this type of model, of learning your capabilities and being able to communicate that in a way that the collective can auction for tasks, could be translated to those more heterogeneous systems as well.

Roderich Gross: I think it could. I mean, in our case, we had only a single currency or a single capability - can you do these parcels? But it was multi-dimensional. Like you have the characteristics of the jobs, and we have even included things like weather patterns now to make it a bit more broad.

There could be different heterogeneous capabilities where certain robots have certain classes of capabilities and all of them can share it. Something we haven’t looked into is teamwork, like perhaps there’s tasks that require multiple capabilities at the same time. It’s a cool idea, perhaps we can work on this together.

Sabine Hauert: All right, a future project.

Roderich Gross: Jointly, you need the correct capabilities, and you can only be confident about what you can offer. Like in a real team, you have some estimation of what others can do, and together you need to have group confidence.

Sabine Hauert: This brings me to the last question. A lot of the things we’ve discussed today have to do with putting robots in unknown environments with unknown conditions where you need to learn the model of the individuals and adapt to the environment you’re in. How do you see this taking shape for the future? What are you excited about in terms of making some of these things a reality, and what are some of the challenges that remain?

Roderich Gross: So this is quite difficult. We have the capability model that was the internal state of a particular robot, that we can use machine learning to find out about. But one of the environments could not just be known or unknown, but really rapidly changing, almost to the extent that you need entirely new strategies to solve tasks. And how resilient can such systems be made? This is a huge open problem, right?

I think we humans solve this fairly well on our own when, for example, the bus doesn’t arrive, we walk - we just adapt different strategies, right? But how can we make robots in this way? To me, it’s an open problem. We need to bring novel expertise in from machine learning, from people studying resilience, and to capture their knowledge to make it work. Of course, we can always use optimization approaches, but the thing is that we don’t know what function we optimise for necessarily. We can only do so much in simulation because we don’t capture the world as it is, the complexity of the real world.

This is why, in practice, I would hope that a lot of systems in the future involve humans, where you have humans and technology collaborating. And you have this creativity of the humans and the common sense of the humans. If you pair this with the incredible speed of computation and potentially the strength, such as lifting heavy loads, and the scalability, perhaps, of using robots, then you can really do a lot of stuff, I hope.

Sabine Hauert: I really like that analogy. I think that would look quite organic to have these robots and humans all being able to interact in a natural way, in the way we operate on a day-to-day basis. I think that’s a really good way to put it. Thanks, Roderich, for being here with us on Swarm Stack.

Roderich Gross: Thank you very much, Sabine. Thanks a lot.

Special thanks to Ella Scallan for editing this post!

About Merihan Alhafnawi

Dr. Merihan Alhafnawi is a Postdoctoral Researcher at Princeton University, working with Professor Radhika Nagpal at the SSR lab to implement embedded swarms in self-adaptable structures. Her research focuses on the hardware design and build, as well as algorithm development, of social robotic swarms. Ranging from creating robotic elements on façades to make them responsive, to creating smart robotic sticky notes for brainstorming, her robots have been used by more than 500 people. Merihan obtained her PhD in 2022, from the Hauert Lab at Bristol Robotics Laboratory, supervised by Professor Sabine Hauert and Dr Paul O’Dowd.

Transcript (with a few edits)

Sabine Hauert: Hi, Merihan. Welcome to Swarm Stack.

Merihan Alhafnawi: Hi, Sabine. Thank you so much for having me.

Sabine Hauert: It’s so good to have you on the show. We’ve obviously worked together in the past, and you’ve done beautiful work over the years in swarm creativity and using swarm robots as a new way to interact with people, which I think is really exciting. Congratulations on your new Science Robotics paper.

Merihan Alhafnawi: Thank you.

Sabine Hauert: Do you want to tell us a little bit about Swarm Garden?

Merihan Alhafnawi: Swarm Garden is this collaboration between swarm roboticists, civil engineers and architects. We were trying to figure out how to combine these fields in order to make adaptive and responsive architectural façades. Currently, most of our buildings are static. Can we add robotic elements to make them responsive not only to us, the people inside the building, but also outside, to the environmental changes and climate changes?

Swarm Garden came out of this collaboration, and it’s this architectural swarm that has elements that you can, for example, place on windows, and each element opens and closes. Then these elements that we call the SG bots, Swarm Garden bots, each open and close, and they coordinate with one another to figure out the best lighting for the room based on the user’s preferences, the lighting conditions, and the sunlight intensity. And because these systems are going to exist in our spaces, we also looked at how they enhance well-being of occupants - how they can bring a sense of joy.

We also explored Swarm Garden as a form of creative expression. We exhibited the bots at Princeton, and we had people interact with them. We created this wearable people could use to influence the swarm, based on their hand gestures. We gave that wearable to a dancer who performed live in the exhibition, and twice later. When we interviewed her, and she said that the Swarm Garden enhanced her creativity and made her think differently about what dance is and how technology can also enhance dance and creativity.

Sabine Hauert: It sounds magical. It’s a lot of work to make the actual robots because you design everything from scratch. Do you want to describe the robot modules themselves? I’ve seen them in action and they’re really beautiful - they look like little flowers. Could you also describe the modularity of how you can add them and assemble them in different ways – help us imagine them?

Merihan Alhafnawi: Sure. Each robot is made of this very thin plastic sheet that when we confine it into a circular opening, or pull it through a circular opening, the sheets kind of crumble. And then they fold. When the sheet is flat, it blocks the light, and then when we pull it, it creates this flower shape and then allows light to enter the room.

The postdoc at the time who was looking at this with me, who’s now a professor at Northwestern - Professor Lucia Stein-Montalvo - had previously studied the properties of this thin sheet when it folds. And we thought that we could turn it into a robot and make many modules. It becomes modular because each module is independent, so you can create different shapes with it and make them into different reconfigurations. It’s highly reconfigurable. And each module is simple - it has one motor. It has a couple of sensors. They can be put together really quickly and scaled really fast.

Sabine Hauert: When you came to do this project in Radhika Nagpal’s lab, having worked on swarms and creativity and interactions with humans, was that the starting point of Swarm Garden? And then you realised, wow, this is actually a wonderful interface for architecture. Or did you start with architecture and then it happened that it has all this creativity to it as well? How do you go about designing these tools that can be used in so many ways?

Merihan Alhafnawi: That’s a great question. I think it actually came first from architecture. Radhika Nagpal, who I work with, she had this collaboration with Professor Sigrid Adriaenssen from Civil Engineering, and they were thinking about how to merge the research in order to create adaptive façades. And then Radhika said that she needs human-scale robotics - not micro robots or small robots, but actually robots that scale in big sizes. And then when I came into the project, I thought, these robots are going to exist in our spaces, so they need to have a human interaction aspect to them. This is where we started thinking of not just purely the hardware design, of how we can make robots that adapt to sunlight, but also how can we make them adaptive to people, and how can we make the interaction interesting? That’s also when we started thinking about the creative expression applications. If these are going to exist in our spaces, we need to make them beautiful, and we need to make sure that people would be comfortable and happy while interacting with them.

Sabine Hauert: Were they? Did they enjoy it?

Merihan Alhafnawi: They were. In the exhibition, we surveyed people and asked them to describe the robot or their experience in general, just with one word. And we got about 96% positive sentiments from the responses. People were really happy. They used a lot of words like cool, amazing, floral. And I think generally everybody was very animated and happy in the exhibition, especially when the dancer danced with the robots. That was also an impressive piece of the exhibition.

Sabine Hauert: I was just thinking about the floral element. You always have flowers on, and you do today! And then, you know, you often say that of your pets as well. I’m sure Mango’s somewhere around.

Merihan Alhafnawi: Even my dog has a lot of floral things.

Sabine Hauert: That’s amazing. I also wondered where next with this, because now you’ve done the proof of concept that this can integrate into human environments. Is this something that you see could translate to architecture, or is there just so much more research that needs to be done in these smart spaces?

Merihan Alhafnawi: It’s true. As part of the Science Robotics paper, we did some experiments where we put them on a real office window. We stacked around 4 rows of robots, so 16 in total. We saw how they interacted to sunlight and, using an opinion dynamics model, how they took into consideration not just my preference as a user and as an occupant of the room, but also how much sunlight there is out there. And they reacted successfully to sunlight. But then the next step is actually to work with architects to see how this can scale to big buildings. How can it be integrated into buildings? And also, algorithmic-wise, the next step is to see how they adapt over time to different user preferences. So how can we have customisable experiences for every user that is in the room, or different rooms, and how it learns from people’s behaviour?

Sabine Hauert: There are beautiful concept drawings from the architectural side in the paper as well. So you can really start to imagine what this might look like at larger scales. It’s very cool. This is not your first robot that interacts with humans or swarm that interacts with humans. In the past, you’ve worked with other systems like the Kilobots, using them as canvases that people could paint on in an interactive way. And you’ve designed 100 robots in lockdown called the Tiles that can assemble into MOSAIX. What have you learned through designing all these systems? And what do you hope for the future in terms of where this field is going?

Merihan Alhafnawi: So the robotic canvas work that I did with the Kilobots and the tiles, which are part of the system called MOSAIX – this was all work I did when I was your PhD student.

Sabine Hauert: We used them this weekend in my neighbourhood actually, to ask the neighborhood what they thought about robots for the future. And all your little robots were post-it notes for my neighbours.

Merihan Alhafnawi: Oh my God, you need to send me pictures.

Sabine Hauert: I’ll talk about that another time. Let’s carry on. So, what have you learned and where is it going?

Merihan Alhafnawi: When I first worked with the Kilobots, it was super early stage research that I did as part of my master’s with you. And I learned a lot about how the robots themselves can interact with one another, and how I can exploit these interactions between the robots and what they can sense from in their environment, so that I as a user can interact with them. As we know, swarm robots, they’re decentralised. So, when we want to add a person, we don’t want to break this decentralisation. Robotic Canvas taught me a lot about how to exploit the environmental and the local interactions to add a human aspect to the system. So that was great.

We did these paintings where people could collaborate with each other to create paintings with the Kilobots. They could also move and be used as a sculpting material to create shapes. While the robots were actually not made for interactive purposes, throughout this project, we transformed them into interactive robots. They were very limited in what they can do, and they needed specific conditions, for example, ambient light.

So that’s how the MOSAIXs came into life. How can we design robots that are highly interactive, that we can deploy in very different environments and that people could interact with really intuitively? That’s how we had the idea to do it with a touch screen, because everybody knows how to use a touch screen. It’s not going to take them a long time to learn, but also we could walk people through the interaction, through the touch screen. We could show different kinds of videos or pictures or buttons.

Then that’s when we took it out to shopping centers, to a museum, to schools. And we did learn that having a highly interactive system makes deployment really easy and makes interaction really easy. And at some point, even museum staff were able to deploy the robots on their own. I would take breaks and I would go, for example, have lunch and they were able to still keep using the robots and explaining to people how it works without a lot of training from me. One of the insights is that highly interactive robots, highly intuitive robots, are very important, not just for the users, but for the people deploying them. And, we learned a lot about how multiple people can interact with multiple robots. In previous research, there were a lot of human swarm interactions, which is one person interacting with the swarm.

Sabine Hauert: Like an operator.

Merihan Alhafnawi: Exactly. With MOSAIX, the beautiful thing about it is that it’s opened up the opportunity for us to explore multiple people interacting with multiple robots in public settings, with people who might not be technical. And we gained a lot of great insights about what people thought worked in the design, what they didn’t think worked, and how we can enhance our systems and how we could add more interactive aspects to our system.

Sabine Hauert: Because you keep designing new hardware, do you think we need to rethink hardware for this type of interaction? Is there something that we should be thinking differently about in terms of what swarms look like when they’re in these interactive spaces?

Merihan Alhafnawi: Something that I really appreciate in your research as well, and I think is very important, is how we have focus studies that tell us what people really want. Because even though I worked with three different swarm systems, two of which I designed, they look completely different. So, the Kilobots are these small robots that don’t have wheels and they buzz and vibrate to go to places. The Tiles are like those touch screen robots that have wheels and can integrate into our environment. And then the Swarm Garden is these robots that you can put on your windows or in your space. They don’t have wheels. They’re like robots in a grid. Completely different designs because they’re for completely different purposes.

It’s one of the challenges of, interactive swarms - thinking about the form that the robot takes. It’s not your usual humanoid, it’s not your usual drone. Now they have to be very creative and very suitable for the task in hand. I think keeping an open mind about what a robot is, what makes a robot, and that we could have really out-of-the-box designs that really work with people. And I think people in this context can really help us - tell us what works for them and what doesn’t. And that’s why - going back to the first point - focus studies are really important. Having early on deployments where people can give us their feedback is also very crucial.

Sabine Hauert: That makes me dream, and hopefully makes people dream more broadly about what swarms could be like within their everyday environment. Thank you so much, Merihan, for joining us on Swarm Stack.

Merihan Alhafnawi: No, thank you. It’s my pleasure to be here.

Special thanks to Ella Scallan for editing this post!

About David Garzón Ramos

Dr. David Garzón Ramos is Assistant Professor and Ad Astra Fellow at the School of Mechanical & Materials Engineering, University College Dublin (UCD), Ireland. He received a degree in Electronics Engineering and an MEng in Industrial Automation from the Universidad Nacional de Colombia, Colombia. He also received an MSc in Automation and Robotics from the Universidad Politécnica de Madrid, Spain. From 2017 to 2024, he conducted a PhD on swarm robotics at IRIDIA, the Artificial Intelligence research laboratory of the Université libre de Bruxelles, Belgium. Before starting his position at UCD, he was a Senior Research Associate for swarm robotics at the Bristol Robotics Laboratory, University of Bristol, United Kingdom.  Dr. Garzón Ramos’ research is at the intersection of artificial intelligence, collective intelligence, and robotics. He focuses on investigating how to use optimization and machine learning concepts to design large groups of intelligent and self-organizing autonomous robots—the robot swarms. He is active in the popularization of swarm robotics research through science communication activities.