Buildings Are Not Robots: Why Critical Thinking Matters More Than Ever in the Age of AI

Interview with Zoltan Nagy by Lada Hensen Centnerová

Dr. Zoltan Nagy is Full Professor and Chair of Building Services at TU Eindhoven, where he leads research that transforms how buildings interact with energy systems and occupants through AI-driven sustainable solutions. A roboticist turned building engineer, he previously directed the Intelligent Environments Laboratory at UT Austin (2016-2024), pioneering the application of machine learning and artificial intelligence to building controls and urban energy systems.

Dr. Nagy's work on reinforcement learning for energy management culminated in CityLearn, now recognized as a UN Digital Public Good for advancing sustainable development goals SDG7, SDG11, and SDG13. He actively bridges building science and energy systems communities, serving on committees of IBPSA and ACM SIGEnergy, demonstrating the transdisciplinary nature of his research spanning building performance simulation and computational energy systems. His innovations have earned recognition including, Fellow of IBPSA, Outstanding Researcher Award from IBPSA-USA (2022), multiple Best Paper awards from CISBAT and Building & Environment, and Applied Energy's Highest Cited Paper Award.

LHC: Your PhD is in robotics. How did you move to the built environment related research?

ZN: After finishing my studies, I wanted to work on important problems. I worked on medical micro-robots—that's important, helping people with technology. But then I attended a talk that opened my eyes to buildings. They have a huge impact on climate change but also huge potential to improve.

The challenge fascinated me: buildings aren't like robots. You don't build 100 identical units. Each building is unique, even when built by the same people, at the same time, with the same materials. There's an opportunity for technology to help bridge this challenge—to be unique for every building while having impact at scale.

So, I went to the architecture school and did a five-year postdoc. Moving between the technical and artistic fields was challenging, but that's where the interesting problems are.

LHC: How would you define Artificial Intelligence (AI) in general and in relation to built environment?

ZN: This is a complex question. AI was coined in the 1960s, focusing on automated reasoning—can machines make logical decisions and deductions? The ideas were there, but computing power wasn't.

Then came neural networks and what we called Machine Learning (ML) for a long time. It's essentially pattern matching: you have input data, you want to match them to output data. There's no real decision-making or reasoning involved—it's pattern recognition.

Machine learning gained momentum in the mid-90s as computing improved. By the 2000s, it started entering building science—not mainstream yet, but the idea emerged that buildings generate lots of data, and we could use computing to recognize patterns when we lack time or human resources.

Now with ChatGPT and other Large Language Models, people call these "AI" because they seem human-like. But they're sophisticated auto-complete machines—they guess the next word really well, but there's no actual decision-making. We're still doing very strong computing, not true intelligence. Whether we should create actual decision-making AI is an ethical question we need to address.

LHC: In your paper/essay ‘The future of artificial intelligence in buildings’ [1] you write that AI is a technology that aims to emulate human intelligence, including understanding natural language, recognizing images and making decisions through learning from past experiences. But you also said that access to data is very limited. What can be done about that?

ZN: This is our biggest challenge. We generate data but don't store it in usable ways. Think about cars—manufacturers collect data continuously, download it during service, and share it to improve all vehicles. But buildings? Once you build a house, nobody goes back to collect systematic data.

Even when we collect building data, people resist sharing it, often citing privacy concerns—though these can be addressed technically. Without data, you can't do machine learning. There's nothing to learn from.

It's not just a technical problem; it's legal and cultural. People share everything with Alexa/Siri/Google Home but won't share building performance data. It's about the value proposition—we haven't demonstrated enough value for giving up that privacy.

LHC: Well, I can give you data from our family house but you cannot use it for a house in Texas.

ZN: You're right—I couldn't use it in Texas. But I could use it for buildings in Amsterdam or even Germany. Climate matters, but there are regional pockets where data transfers well.

The real challenge is occupants. Put two identical buildings next to each other with different families, and you get different results. But that's where machine learning can help—teasing out the contributions of people versus systems. We can find patterns even in this complexity.

LHC: Do you have an example?

ZN: As we electrify homes, they become more interactive with the grid through solar panels and electric vehicles. When should you charge your EV for the lowest cost? Companies might say "charge between 8 PM and 6 AM," but that might not work for your schedule.

An intelligent home energy management system can learn your individual patterns and optimize accordingly. Your solution will differ from your neighbor's because your needs differ. You can't pre-program this—it needs to adapt.

The bigger picture is grid congestion. The challenge is controlling the built environment—this huge battery—with many degrees of freedom to help everyone collectively without impacting individuals negatively.

LHC: In your paper ‘MERLIN: Multi-agent offline and transfer learning for occupant-centric operation of grid-interactive communities’ [2] you write: Decarbonization needs grid-interactive efficient buildings (See Figure below). Could you explain that?

ZN: Let's look at history. Post-war buildings from the 1950s-60s had thick walls, no insulation, single-glazed windows—you heated them when needed. The 1970s energy crisis changed everything. We added insulation, better windows, and now we have passive houses.

But here's the problem: efficiency lowered the energy curve, but buildings got bigger. The curve actually went up! Per person, we use more energy because we have higher comfort standards and more space.

Until recently, we focused on annual energy balance—a Zero Energy House balanced over a year. But that's an accounting trick. The real problem now is grid congestion. When everyone uses electricity simultaneously—like during peak evening hours in the Netherlands—the grid struggles.

We need to look beyond individual buildings to entire neighborhoods. It's not just about using less energy annually; it's about when you use it daily. With renewable energy, timing matters for emissions. Use electricity when it's from solar or wind, not coal.

LHC: The next question is about your paper Ten questions concerning Reinforcement Learning (RL) for building energy management [3]. What’s the main message?

ZN: RL is interesting because it can potentially make decisions beyond pattern recognition. But researchers got lost in fancy algorithms. What we need now is to run these on real buildings and prove they work.

We don't need better algorithms—we need to demonstrate their value. They work great in simulations, but buildings aren't games you can play millions of times. We need real-world results.

I believe we'll see two main applications: First, adapting buildings to individual occupants day-to-day. Second, managing communities of buildings to address grid congestion—a much harder challenge. We should stop focusing on single buildings and tackle these harder problems.

LHC: Generation Z grew-up with computer games and I read that the best results in RL are made in game industry. What can we learn or use from it in built environment?

ZN: That's exactly the problem! In gaming, you can play millions of times and learn. Deep Blue beat Kasparov at chess through brute computing force. AlphaGo mastered Go through endless practice.

But buildings don't have clear rules or end goals like games. You can't "play" a building 100 million times. There's no clear victory condition.

Autonomous cars are similar—unpredictable, but cooler and better funded. Tesla shares data between all vehicles, improving continuously. Buildings need similar data sharing, but nobody wants to watch an autonomous building like they watch self-driving cars. We lack that "wow factor" demonstration equivalent to make people care.

LHC: Speaking of challenges, you organize the CityLearn challenge [4]. How does this accelerate progress?

ZN: We developed CityLearn as building simulation software, but individual PhD projects often disappear after graduation—such wasted potential! So we made it open-source with competitions.

The challenge requires winners to release their solutions publicly before receiving prize money. It's about open science and democratizing development. I am envisioning it to become a community-driven tool, like Linux for building control—anyone can and should use and modify it.

LHC: As an educator with a robotics background, is AI more useful for building design or operation?

ZN: It's not either/or. We won't outsource building design to AI—that's not smart. But AI can handle boring, repetitive tasks and speed up processes.

When CAD emerged in the 1980s, people said it would end designers. It didn't—it made them faster. Same with parametric design and scripting tools like Grasshopper. These are tools that enhance, not replace.

For operations, there's more potential for autonomous agents making decisions, because you can't manually control everything continuously. But we still need human oversight and understanding.

LHC: The last question is to you as a teacher. What is the most important for HVAC students to learn in this AI age?

ZN: Focus on basics and critical thinking. Now you can generate so much content so quickly—but also so much garbage! Without understanding HVAC fundamentals, you might design something that looks good but is actually terrible.

My kids use AI tools constantly. This isn't going away. But that makes critical thinking even more important. We need to teach students to use these tools critically, understanding when output is valuable versus when it's nonsense.

RL and AI are just tools. You must be able to evaluate them critically. If you don't understand the basics, you can't judge whether AI-generated solutions are brilliant or garbage.

The future isn't about AI replacing human intelligence in buildings—it's about using these tools wisely while understanding their limitations.

References

[1]     R. Labib and Z. Nagy, “The Future of Artificial Intelligence In Buildings,” ASHRAE J., vol. 65, no. 3, pp. 26–32, 2023.

[2]     K. Nweye, S. Sankaranarayanan, and Z. Nagy, “MERLIN: Multi-agent offline and transfer learning for occupant-centric operation of grid-interactive communities,” Appl. Energy, vol. 346, 2023, doi: https://doi.org/10.1016/j.apenergy.2023.121323.

[3]     Z. Nagy et al., “Ten questions concerning reinforcement learning for building energy management,” Build. Environ., vol. 241, p. 110435, Aug. 2023, https://doi.org/10.1016/J.BUILDENV.2023.110435.

[4]     Z. Nagy and K. Nweye, “The CityLearn Challenge: Four Years of Advancing Common Task Frameworks for Energy Management in Smart Buildings,” in E-Energy ’25: Proceedings of the 16th ACM International Conference on Future and Sustainable Energy Systems, 2025, pp. 944–948, [Online]. Available: https://dl.acm.org/doi/10.1145/3679240.3734667.