Machine intelligence has entered a new phase. Artificial systems now interpret images, language and sensor data, generate text, code and designs, support consequential decisions, and increasingly act in the physical world through robots, autonomous vehicles and networked infrastructure. What began as a set of narrow tools is becoming a general technology that extends into science, healthcare, industry, public administration and everyday digital services. This expanding reach exposes a central scientific challenge. Every such system learns from a finite record of the world as it was, yet it must operate in a world that is uncertain, interconnected and continually changing. The conditions it meets are rarely those it was built for, its outputs feed other systems and carry its errors with them, and its own actions alter what it will observe next. The question is therefore not only how to make machines more capable, but how to enable them to keep learning, reasoning and adapting in operation while remaining reliable, efficient and open to human scrutiny.
Our research develops the principles and methods required for this next generation of machine intelligence. We combine machine learning, computer vision, signal processing, and mathematical modelling, connecting foundational research with demanding problems in healthcare, manufacturing, robotics, energy, autonomous systems and digital decision support. Our activities range from multimodal foundation models and generative artificial intelligence to trustworthy adaptation, distributed learning and intelligent interaction with the physical world.
We focus specifically on three research fronts:
● Multimodal foundation models and machine reasoning investigate how machines can build useful representations of the world from images, language, audio, sensor measurements and other forms of data. Although modern foundation models can recognise complex patterns, genuine machine intelligence requires more than statistical prediction. Systems must be able to connect observations across modalities, identify relevant concepts, reason about relationships and consequences, and use their knowledge in situations that differ from their training data. We study self-supervised and weakly supervised learning, multimodal foundation models, generative models, world models and agentic systems that can interpret information, simulate alternatives and interact with people or other computational tools.
● Adaptive, data-efficient and trustworthy learning addresses the fact that the conditions encountered during deployment rarely remain identical to those represented in the training data. Sensors change and unexpected operating conditions arise. We develop methods for transfer learning, continual learning, domain generalisation and test-time adaptation, enabling models to learn from new experience without requiring complete retraining. A fundamental question is how a learning system may change while preserving previously acquired capabilities and avoiding silent deterioration.
● Embodied, distributed and collaborative intelligence studies systems whose predictions influence the world from which their future data arise. Examples include robots, autonomous vehicles, sensor networks, digital twins, intelligent manufacturing systems and decision-support platforms. In these settings, intelligence forms a continuous loop of perception, reasoning, action and feedback. We investigate learning-based perception and control, planning under uncertainty, human–AI and human–robot collaboration, multi-agent coordination, and efficient inference across devices, edge infrastructure and cloud resources. Particular attention is given to systems that must make decisions under constraints on time, communication, computation, energy and safety.
Across these research fronts, our ambition is to understand and develop the mechanisms that allow artificial systems to remain useful when knowledge is incomplete, conditions change and decisions have real consequences. We provide a meeting point for foundational AI research, interdisciplinary collaboration and responsible translation into the complex digital and physical systems on which society increasingly depends.
