Research
This interview is part of a broader editorial project by Imminent, featuring conversations with expert professionals collaborating on Experience AI — which begins when machines learn and interact with the real world in real time — within the DVPS project.
DVPS is among the most ambitious projects funded by the European Union in the field of artificial intelligence, backed by an initial investment of €29 million. It brings together 20 organizations from 9 countries to shape the next frontier of AI — one rooted in the interaction of machines with the real world. Building on the success of large language models, DVPS explores the future of AI through multimodal foundation models. Unlike current systems, which learn from representations of the world through text, images, and video, these next-generation models are designed to acquire real-time empirical knowledge via direct interaction with the physical world. By integrating linguistic, visual, and sensor data, they develop a deeper contextual awareness, enhancing human capabilities in situations where trust, precision, and adaptability are critical. The overall initiative is led by Translated, which coordinates the project’s vision and implementation. The team brings together 70 of Europe’s leading AI scientists. The potential applications span across several domains, including language, healthcare and environment.
Prof. Aldo Faisal is Professor of Neuroscience at Imperial College London and holds a Chair in Digital Health at the University of Bayreuth, Germany.
Aldo Faisal
Professor of AI & Neuroscience, Imperial College London
Aldo Faisal is Professor of AI & Neuroscience at Imperial College London, where he leads the Brain & Behaviour Lab, and Director of the Behaviour Analytics Lab at the Data Science Institute. He is also affiliated with UCL and the MRC London Institute of Medical Sciences. His research focuses on AI, neuroscience, and behaviour, with applications in healthcare and human–machine interaction. He serves as Associate Editor for Nature Scientific Data and PLOS Computational Biology and has received international recognition, including a Toyota Mobility Foundation award and election to the World Economic Forum’s Global Futures Council.
Can you describe your work and how it connects to the DVPS project?
My work straddles artificial and human intelligence—not just drawing on neuroscience to build better systems, but also using those systems to help restore function in the brain and elsewhere in the body.
The human brain already integrates multiple modalities to understand the world, and I think we still have far more to learn from that than we currently draw on.
Take multimodal or generative systems aiming for more general capabilities: A system needs short-term memory as much as long-term memory, and the ability to integrate information correctly across both. It also needs to be aware not just of what the world looks like, but of what it can do to change it, because knowing your possible actions is what shapes your decisions.
We focus on healthcare, but we think it’s essential to treat the human as part of the system. Human–machine interaction matters to us just as much as building something highly capable.
In your eye-tracking work, gaze reflects intention, attention, and decision-making. How far does understanding how we act in the world help us understand intelligence itself?
This is what we call embodied intelligence, or embodied AI: the idea that intelligence cannot arise in the universe without a body. The only way your brain interacts with the world is through movement: to speak and act. A disembodied machine—a computer that just sits somewhere—doesn’t really make sense.
Concretely, embodied AI is a system that operates within a real-world behavioral loop. It isn’t something where you ask one question and get one answer; it interacts with you, with others, and with the world. It knows what it’s dealing with, and how it can act to change things.
EMBODIED AI
Embodied AI is a paradigm of artificial intelligence that emphasizes the role of an agent’s interaction with its environment in producing intelligent behavior. Rather than treating intelligence as abstract computation that maps inputs to outputs, embodied systems operate through a continuous perception–action loop: they sense their surroundings, select and execute actions, observe the consequences, and adapt their behavior accordingly. The “body” in question need not be humanoid, or even physical in the conventional robotic sense. It refers more broadly to the sensors, actuators, effectors, and other interfaces through which an agent is situated in an environment, perceives the environment and its own state, and can act upon it. What distinguishes embodiment is therefore not simply having inputs and outputs, but the coupling between perception, action, and the environment. Technically, embodied AI often draws on reinforcement learning, in which an agent learns policies that maximize expected reward through interaction with its environment, as well as imitation learning, planning, world models, and other forms of learning from experience. Such systems may also integrate multiple sensory modalities (vision, sound, etc.) into representations of both the external environment and the agent’s own state, often across different spatial and temporal scales. This combination of multimodal perception, continuous interaction, and physical or environmental constraints makes embodied AI particularly challenging to train and deploy. Real-world interaction data is costly to collect, errors can have physical consequences, and environments are noisy and only partially observable. Simulation reduces the cost and risk of training, but the gap between simulated and real-world environments makes reliable transfer difficult. Embodied AI applications span robotics, autonomous vehicles, prosthetics and assistive technologies, surgical and rehabilitation systems, and interactive healthcare. In these settings, systems must not only understand the world but also determine available actions, predict their consequences, and adapt as conditions change. This coupling of perception, reasoning, and action lies at the heart of embodied AI.
Take medical applications: It’s not enough to know a patient’s state. The system needs to know what a doctor or a healthcare system can actually do to help that patient. If action and thought are disconnected, the answers you get won’t be meaningful.
That kind of interaction is essential if technology is going to serve people rather than replace them, whether in education, healthcare, or elsewhere. We know this from language itself: If you speak and nobody responds, or responds inappropriately, communication breaks down.
For our bodies, error is a source of learning—while AI, for now, mostly tries to avoid it. Is that a fundamental gap? And how does design change when errors become signals rather than something to eliminate?
We often learn to do new things precisely because something we tried didn’t work. In standard supervised learning, a model has a “teacher” who simply says whether something is right or wrong—that’s the only feedback it gets. We work with reinforcement learning instead, where the system pursues a broader goal—say, treating a patient—and receives positive or negative feedback along the way. It then has to figure out for itself which changes actually helped. That reward-based approach turns out to be much better-suited to solving bigger, messier problems, whether the learner is artificial or human.
But even if a system can work things out for itself, what it learns ultimately depends on the data it has access to. And if you want to understand human behavior—movement and interaction—you need a description that is as objective as possible, not filtered through our own assumptions about what behavior should look like. Ideally, you learn this end-to-end from natural, everyday behavior rather than staged experiments.
That’s why we’ve set up a range of “living labs” in our laboratory, where we simply observe people going about their daily lives and collect that data. It lets us capture human behavior in a different way. Once you have that, you can bring it together with other foundation models for health, language, and beyond.
The goal isn’t a single multimodal model, but an ensemble of specialists working—and talking—together.
What are the main technical challenges of building multimodal models in this context, compared with more traditional domains? Is a universal embodiment possible, or does each application domain need its own body and interactions?
The very first issue is about training data: Biomedical data is very sparse. A lab test might be ordered a week after a patient falls ill, analyzed days later, and written up in a report a week after that—there’s a strong asynchronicity. We’re working across wildly different timescales—an ECG plays out in milliseconds, while a disease (or a recovery) can take years—and the data itself comes in scattered, inconsistent formats.
On top of that, there’s simply very little of it: Language models are trained on nearly all the text humanity has ever produced, while in medicine, even pooling records from all 450 million European citizens gives you only a fraction of that scale.
None of that, though, makes me think the answer is a separate system for every application domain. Walk around Rome, and you’ll see people who speak several languages, who translate, who might be surgeons, and also tend a garden—all at once.
We don’t build a different brain for every one of those things. One way of thinking can solve multiple problems. So why would we need a different embodiment for every domain? That’s a narrow, purely engineering view of intelligence.
The value of DVPS, and of this European funding, is precisely that it gives us the opportunity to find one way of thinking that can solve multiple problems at once. That’s something that is deeply human.
Is Europe a supportive environment for this kind of interdisciplinary research?
I’m a great proponent of what I call sovereign AI: We need to have our own AI technologies, maybe not in every field, but in the core fields that are important for Europe.
Europe is not a “United States of Europe” yet. It is made up of many individual European countries, and that brings with it a great deal of diversity. We know from AI that diversity is one of the things that helps us build smarter systems.
There is therefore a real opportunity to innovate in Europe. And when it comes to problems of a societal scale—whether climate, the environment, or healthcare—the European way of doing things can be a real strength.
We have universal healthcare, and we have data from patients spanning their entire lives, from cradle to grave. That gives us the opportunity to think about and build systems that can genuinely help citizens feel better.
Outside Europe, by contrast, data spaces are often highly fragmented, with private players competing with one another and reluctant to share data. So we have a unique opportunity to build something stronger precisely because Europe is the way it is.
Not despite it.
By the end of DVPS in 2028, what legacy do you hope to have created—scientifically, but also in terms of approach, infrastructure, or this European way of doing AI?
What I treasure about DVPS is that it brings together some of the best people in Europe around a common goal—good competition, not bad competition. We bring different strengths and perspectives together, and think collectively about how to find better solutions. That is always valuable for science and innovation.
But what’s genuinely exciting is the sheer talent and opportunity we’re bringing together in Europe to solve problems the European way. That, to me, is what sets us apart from the corporate efforts being made elsewhere in the world.