It is the stated goal of the Helmholtz Visiting Researcher Grant to promote knowledge exchange, enable new research collaborations, and explore new or emerging research topics in the field of information and data sciences. With this grant, HIDA supports PhD students and postdocs in undertaking short-term research stays of one up to three months at one of the 18 Helmholtz Centers. This opportunity was seized by Volker Strobel, a postdoc at the Université Libre de Bruxelles. Together with CISPA researcher Mario Fritz, he delved deeper into the security aspects of swarm robotics and came up with some completely new research ideas.
Dear Volker, what drew your attention to the Helmholtz Visiting Researcher Grant and what brought you to CISPA?
HIDA issued a call to apply for the Helmholtz Visiting Researcher Grant, and it came at just the right time. I had recently secured major research funding for work on the security aspects of swarm robotics. I was collaborating in Brussels with Professor Dr. Marco Dorigo, an absolute expert in swarm robotics, who has been researching various topics related to robot swarms for over twenty years. But we didn’t have much expertise in security aspects yet, which is why I saw this opportunity for exchange as a fantastic opportunity to create synergies. I then looked at the focus areas of the Helmholtz Centers and quickly came across Mario’s work. He also researches the security of robots and AI systems. I was able to start in July 2024, after my proposal was accepted by HIDA.
Before we talk about your research, tell me what the HIDA support looks like exactly.
The Helmholtz Visiting Research Grant offered by HIDA provides extensive support, including a research fellowship from the host Helmholtz Center, which may fully cover the costs that accrue during the research stay. In my case, the university in Brussels continued paying my salary for practical reasons, but the grant covered a housing allowance, mobility assistance, a contribution toward health insurance here in Germany, and even a subsidy for childcare costs. That’s really fantastic.
What exactly is swarm robotics, and what are robot swarms used for?
Swarm robotics is based on the fundamental idea that, instead of relying on a single, highly complex robot, you might have many less complex robots work on specific problems. This field of research is strongly inspired by observations in nature, for example by the behavior of bee and ant colonies. Collectively, the swarm arrives at solutions for problems that would be unthinkable for its individual members. The focus is on accomplishing complex tasks without central control. The robots gather information and share their findings only with the nearest local member of their swarm.
Robot swarms are not yet widely used, but there are many potential application scenarios. For example, they could play a major role in disaster relief, such as using drone swarms to locate avalanche victims. They could also be beneficial in agriculture. Additionally, swarms are valuable in research, they can for example be used for deep-sea exploration.
What is the advantage of having many autonomous units instead of one central control unit?
At best, robot swarms can handle deviations much faster and with more flexibility. They are also more fault tolerant. Even if parts of the swarm fail, the remaining members can continue to work. Robot swarms are also very scalable. But it is still challenging to ensure that they make good and reliable decisions, even if, for example, one of the robots is hacked. This automatically raises the question of where humans need to be involved in the process and exercise control. The more autonomously robots can decide, the more effective they can be. But you have to be able to trust their reliability. We always have to find a compromise between autonomy and security.
Security, that’s my cue: What exactly did you and Mario work on?
Initially, I thought that Mario's background in so-called federated learning would form the basis for our collaboration. After all, federated machine learning is also about decentralizing the learning process and training many small models instead of one large one. But currently, Mario is particularly active in researching the security aspects of large language models, and we quickly found a completely different approach that is super exciting.
What does this approach look like?
In one experiment, Mario had several software agents, each backed by a large language model, negotiate with each other and make decisions together. For example, a scenario such as: The mayors of various cities have to come to a joint decision even though they represent very different interests. Large language models are actually extremely good at this. They possess a great deal of world knowledge and also some ethical understanding. If we can incorporate these abilities into robot swarms, it will greatly increase the security of the swarms regarding the quality of their decisions.
Large language models, or LLMs for short, are trained on the basis of neural networks and can understand and even generate complex texts. They underlie chatbots such as ChatGPT, for example. How did you show that they actually enable robot swarms to make better decisions?
For example, we equipped a swarm of robots with LLMs and cameras and gave them the task of searching a field and reporting back to us where grain was growing and where weeds. Then we created two problems without specifically preparing the robots for them or telling them how to react. First, we pretended that there was an injured person in the field, at least according to the virtual data. The robots examining this area of the field then autonomously reported back to us that there was an injured person there. In real life, a human could then have gone to check and get help. Second, we programmed one of the robots to ignore the actual data and only to report weeds. Again, the other robots alerted us to the fact that this robot's statements were probably not true, since all the others were detecting both crops and weeds. With the help of LLMs, the robots can communicate with each other better and we can immediately understand what is happening.
That sounds super interesting, but is it really that easy to equip these robots with LLMs? I thought the focus was on keeping the robots as simple as possible.
Yes, that's true, and the hardware limitations of current robots are indeed an issue. Running a large language model on a robot is certainly more complex than using a simpler controller. But it’s not unrealistic, because it’s already possible to run LLMs on very small devices. We’re also assuming that hardware will become even smaller in the coming years, and software will become more efficient. So, we don’t see this as a major obstacle.
Could this mean that in the future, robots will be smart enough to make decisions entirely without human intervention?
It is at least conceivable that humans will only have to intervene in exceptional cases. LLMs have at least some kind of common sense and they know what human decisions look like.
The interdisciplinary exchange that HIDA made possible for you seems to have been very valuable at this point.
Absolutely. It was really exciting to get Mario's perspective on my topics. And even though we have completely different research foci, he immediately understood everything and had some great ideas. This has had a huge impact on my research and I think we will continue to work together.
That sounds fantastic, Volker. Many thanks for the interview.
