Interview Transcript
Q1: Could you tell us your name, where you’re from, and a little bit about your background story?
Bengt Lennartson: My name is Bengt Lennartson. I’ve been a professor at Chalmers University in Sweden for quite many years. I’m 70 years old, so I studied at this university as well. After some time, I got interested in feedback control and later also in robotics and automation. One thing unrelated to robotics that I work quite much on is energy consumption.
I used to say that due to Obama’s initiative—saying to General Motors, “Okay, if you would like to get some money from the state, then you should go for green technology.” Since I had some collaboration with General Motors 15 years ago, they contacted me at that time, and by that, we started to think about what we can do related to robotics and sustainability. After some time, we realized: let’s go for soft movements and in that way try to reduce accelerations and reduce energy consumption. We got some very good results on that.
Q2: Have you always been working in Sweden?
Bengt Lennartson: Yes. Actually, I moved as a student to Gothenburg and I stayed there my whole life, with one exception. I have been quite many times in Newcastle in Australia for research. First time for half a year, and then I was there totally six times or something. So that’s an interesting experience.
Q3: How did you get interested in robotics in the first place?
Bengt Lennartson: Well, you could say that it was not robotics in itself—it was more automation. Sweden is extremely good at automatic control, with a lot of famous professors. I realized maybe I should try to move a little bit into a related area, and thought manufacturing automation, programmable logic controllers, and more logical control aspects could be interesting for me.
So I started to look more into that field and developed some interesting theory for hybrid systems where you mix continuous and discrete behaviors. Later, I was also establishing an interesting bachelor program at our university, which now has about 120 students starting from the first year to study automation and mechatronics. From the beginning, that program was actually called robotics. But 30 years ago, robotics was not so popular because people were afraid that robots should take our jobs. Things have changed a little bit since then.
Q4: What are some of the main research questions you have when studying in this area?
Bengt Lennartson: Since my background is in feedback control, it was a question of how we can make these types of systems correct on a higher level. We looked quite much into formal methods to verify and be sure that when we generate a control function, it’s working correctly.
For instance, a very typical question related to computer science is deadlock. If a deadlock occurs, nothing more will happen. We looked at how to avoid that kind of logical bad behavior to guarantee safe and live behavior. Later, looking at a robot as a component in a production system became mostly my focus. Regarding the energy reduction I talked about, we were very early on looking at a couple of robots working together in the same area, avoiding collisions between robots and other equipment while simultaneously saving energy. That was something new at that time.
Q5: How did your research interests evolve over time?
Bengt Lennartson: It’s interesting that you say “research questions,” because when I was young, we didn’t talk so much about research questions—it was more what kind of goals or ideas you have that you could apply. Nowadays I realize it is a lot about research questions.
One thing you could say is that sustainability is something we have added to our portfolio to see how we can make all types of systems more sustainable. From a political side, I find that climate change and all those kinds of things are so important, so anything we can do to improve sustainability aspects is important.
Q6: Have you faced any major challenges in your career—institutional, technical, or social?
Bengt Lennartson: The challenge is always to get money for your research. At the same time, the area I worked in is considered to be very important.
One thing that has changed more recently is that now everyone talks about AI and machine learning. I used to say, “Well, I worked on that 30 years ago, but then I called it system identification.” You learn a model for a system, and based on that model, you design controllers that should be adaptable. Now we call that reinforcement learning, which I didn’t know about at all at that time, but I gave courses on adaptive control, system identification, and discrete behaviors related to expert systems and old-fashioned AI.
Now I can put all these things together. My vision is to write a new textbook on logic, learning, and decision, where I give the whole picture of modern AI.
Q7: When is your new textbook coming out?
Bengt Lennartson: I have talked about this for a long time, and I am still doing some research to develop algorithms even further. I hope to really start writing on this this summer. I am already giving one course in machine learning and one course in logic, learning, and decision. My dream is to put these two courses together into one textbook. I’m really excited about it.
Q8: Given that your work touches heavily on sustainability, do you work mostly with industry companies or government funding?
Bengt Lennartson: It has been a mix. For the robot energy consumption project, the interesting thing was that General Motors was in the end not so enthusiastic about our results. They said robots don’t actually consume so much energy compared to other aspects of car production, like the paint shop and ventilation. But Daimler and Kuka here in Europe were very interested, and we got a big European project on that.
Otherwise, it has been more local Swedish projects. In Sweden, we have a demand that industries should be involved, but we don’t ask them to give us money—the money comes from the government side. To guarantee that our research is important for industry, companies need to commit that they are involved in the project.
Q9: Has there been any project you would have liked to pursue, but had to give up due to career challenges?
Bengt Lennartson: For the last seven years or so, I haven’t applied for any research money because I like education and giving courses, which allowed me to finance my position at the university. For some time I even gave five courses to finance myself. I’ve had nearly 30 PhD students produced during my career, but now in the end I spend most of my time on education, which leaves time for my own research.
Now I really have the opportunity to be what I call a “true professor,” meaning I can sit and work on my own theorems and proofs. The problem is I have to do all things myself, which is good and bad, but I think I am having the best time of my life right now.
Q10: Do you have any advice for students who aspire to be in the same position as you?
Bengt Lennartson: Some of my colleagues say they should do like me—skip some parts and focus on being more independent. I am very independent now; I can decide exactly what I want to focus on. That means I have the opportunity to learn on a much broader umbrella.
I worked on formal methods, optimization, feedback control, and so on. A word I use very much is “unification”—trying to understand how computer science and control engineering explain things. Everyone is talking about nearly the same thing, but how are they related to each other? That’s something I like very much to explain to the community.
Q11: How do you think robotics has changed through the years?
Bengt Lennartson: There is a dramatic change now when we talk about humanoids. I worked mostly on traditional industrial robots, and now we talk about all these humanoids and their opportunities. That is the obvious change that has happened.
Q12: How do you see the field of robotics and automation developing over the next 10 to 20 years?
Bengt Lennartson: Our society is called Robotics and Automation, and I am from the smaller part—automation. I would rather say: what will happen with robotics and automation when you join them together?
Robotics is about one robot and how it can be more intelligent, while automation is when you have ten robots working together. A major trend right now is reinforcement learning, such as what Chinese companies like Unitree are doing with imitation learning and reinforcement learning.
In the past, we learned based on physical models. Now there is a tendency for industries to ask what can be done purely with data-driven learning. But we already have a lot of physical knowledge. How can we combine old-fashioned physical knowledge together with data-based methods? Model-free reinforcement learning relies only on data (a black box), while model-based introduces physical knowledge. My vision and focus for the coming years is to combine model-free and model-based learning to get better, verifiable results through formal verification and physical models alongside data-based strategies.
Q13: Do you think researchers use “robotics” and “automation” interchangeably, or do they position themselves distinctly?
Bengt Lennartson: To be honest, at big conferences like ICRA, it’s more or less only the word robotics. Most leading persons talk about robotics and forget about automation. I understand that because robotics is so huge—ICRA has around 9,000 participants now, whereas the automation conference (CASE) has around 1,000 participants. Robotics is much bigger; automation is the small sister. But leadership in the society recognizes that we need to look at the combination between robotics and automation while accepting that robotics is the main issue.
Q14: What are you most excited about in your recent research projects?
Bengt Lennartson: There are three things I am excited about:
First is combining model-based and model-free reinforcement learning—combining physical knowledge with data. We are currently applying this approach to eco-driving for trucks. In September, I will teach basic feedback control students how simple tools from feedback control connect to reinforcement learning, giving them a new perspective.
Second is unifying ideas from computer science and control engineering. We use verification methods based on bisimulation abstractions. For instance, a car’s locking alarm system has 1 billion states. How can you be sure there are no deadlocks? We developed an abstraction method that takes 1 billion states and, in a modular fashion, reduces it down to a state model with only three states while preserving safety information.
Third is optimization. Google developed a tool called CP-SAT (combining constraint programming and satisfiability solvers). I’ve been using it for scheduling problems and find it very promising. I am currently submitting a paper on how to use CP-SAT for scheduling applications.