RL | Meta-Learning | Multi-Agent AI
Pablo Magariños, AI researcher
Aerospace engineer researching how embodied AI can learn from its environment
Aerospace engineering was my first encounter with the passion I have had since childhood: technological progress. Human advancement.
My fascination with artificial intelligence goes back to my adolescence, and my ambition to contribute led me to learn more and pursue a Master's in Industrial Mathematics.
Today I work in research and, above all, strive to contribute to this field. My work sits in reinforcement learning and meta-learning under partial observability, applied to multi-agent systems and embodied AI. Specifically, I study the role that world models and the belief state play in an agent's ability to adapt to novel situations and, eventually, learn from them.
New environments. New teammates. New knowledge.
Current Work
I am starting a PhD in Aerospace Engineering at the Universidade de Vigo, centred on embodied artificial intelligence.
Alongside the doctorate, I do research with ATRG, the Aerospace Technology Research Group at the same university, and I have just started teaching at the School of Aerospace Engineering, where I trained as an engineer.
There is more detail on each of these in Experience and Research.
Education
PhD in Aerospace Engineering
Universidade de Vigo 2026 – Present
I am just starting the doctorate, centred on embodied artificial intelligence: how an agent can learn to act competently in the physical world from its own experience of it.
MSc in Industrial Mathematics
Universidad Carlos III de Madrid 2024 – 2026
I took the Master's to go deeper into the field, motivated by the ambition to better understand the language that allows one to manipulate reality through logic and rigour.
Degree in Aerospace Engineering
Universidade de Vigo 2020 – 2024
The degree gave me a strong engineering foundation, but also a solid grounding in mathematics and physics, and a deeper understanding of reality. It led as well to my first published paper, which you can find in Research.
Research Focus
My research is centred on the development of intelligence that genuinely understands its environment, rather than merely identifying statistical correlations in data.
This is my research tree. ClickTap on each element to find out more.
-
How I tackle it
Embodied AI
Right now I work on AI for drones and aerial robotics, but embodied AI is a cross-cutting technology: progress in this field spreads to every other. The drones are simply an anchor.
Long-Horizon Tasks
Long-horizon tasks are diffuse: often we do not even know how they could be solved. Above all, it is hard to tell whether we are getting any closer to the final goal, because the signal that says so arrives late, if at all. Reasoning, world modelling, and turning events into discrete symbols we can reason with may be what makes distant goals reachable.
Meta-Learning
The goal is to push our models not to simply memorise what they see, but to abstract from it, so that what they know helps them learn even more. That lets them adapt quickly to new tasks and unseen situations.
Continual Learning
A model that is trained once and then frozen stops learning the moment it leaves the lab. An embodied agent faces a world that keeps changing, so it has to keep learning on the go without forgetting what it already knew, and fold each new experience into its own knowledge.
World Models
Being able to predict what will happen in the world is closely tied to understanding it. It is not only useful for knowing what comes next: it is useful for building rules about how reality works.
Belief State
We never get access to reality as it is. We cannot know for certain what a rival will do, nor the physics behind everything we see. That is why building a belief state, one that “fills in” the reality we do not know, is essential.
Reinforcement Learning
Learning from existing data is efficient, but it makes it hard to diverge and create new knowledge. The progress of the last few years shows that letting an agent explore and discover new policies leads to fascinating results.
Multi-Agent AI
Theory of mind and the ability to cooperate are still, to this day, some of the great challenges in artificial intelligence. Learning in a social setting, rather than a purely selfish one, opens the door to swarm AI. Adapting to others in real time is a key step.
More About Me
I love photography. What draws me in is experimenting with my camera, understanding optics at a physical level, and trying new things to see how light and images actually work.
I'm also passionate about music production, which has been part of my life for more than eight years. I played guitar for several years and later picked up the piano, but music production is what has truly stayed with me, especially for its creative side and the joy of experimenting with sound and software.
I also enjoy letting my thoughts wander, having ideas, and thinking for the sake of thinking. I am a curious person, and I write some of those reflections in Thoughts.
A few moments from the last few years, each with the story behind it, are in the Gallery.
