DFF Research Project 2025-2029

Research Questions and Project Goals

To enable effective coordination in multi-agent environments, such as hospitals, artificial agents like robots should take into account the knowledge, beliefs and goals of other agents, a concept known as Theory of Mind. Without it, robots run the risk of becoming obstacles rather than aids. Epistemic planning, which develops algorithms to support reasoning about other agents, lays groundwork for socially intelligent robots. The logical framework underlying epistemic planning is highly expressive, supporting arbitrary levels of higher-order reasoning (e.g. “doctor A knows that doctor B doesn't know that patient C believes to suffer from X”) and reasoning over arbitrarily large domains. This makes it computationally very demanding. While restricting either the reasoning depth or the extent of the domain seems like the obvious way to go, it is not trivial how to do this without limiting the abilities of the robot significantly.

The human brain, however, seems to have achieved this. It has developed a strong ability to focus only on the aspects of the world relevant to the task at hand, for instance ignoring the aforementioned piece of higher-order knowledge when the goal is to simply get a cup of coffee for the patient. This is achieved by means of our attention system. Attention is the mechanism by which we handle the otherwise overwhelming complexity of the external stimuli we get exposed to, and the complexity of our internal world model. AI systems such as robots can also suffer from the sheer volume of sensor inputs to be processed as well as complexity of their internal state representations. In this project, we take inspiration from studies on human attention in cognitive science and philosophy to integrate attention mechanisms into logical frameworks and AI systems, leading to our first research question: “How to incorporate attention into epistemic logic to create a framework for epistemic planning that enables efficient, robust problem-solving in multi-agent settings?”

Exploring this question is expected to provide new foundational insights into attention mechanisms, as well as to impact the epistemic and modal logic communities with new logical frameworks, techniques, and algorithms. No rich logical models of attention exists at the beginning of this project. Providing these will lead to formally precise models of the core mechanisms of attention, having independent epistemological interest in many areas (philosophy, psychology, economics, social sciences, computer science, AI), and forming the foundation for building robust, theoretically well-founded attention-based AI systems. Further, the proposed attention-based epistemic planning framework will enable innovation in our existing framework for epistemic planning on humanoid robots. Through proof-of-concept human-robot interaction experiments, we will validate ecological validity, assess interaction quality, and test how accurately our models represent the limited attention of humans. This leads to our second research question: “How to use attention-based epistemic planning to create more realistic models of human reasoning, and hence improve the quality of human-robot interaction?”

The AiEP project has the following overall goals:

Goal 1 Develop formal logical models for attention including goal-driven attention and attentional limitations.

Goal 2 Develop the theoretical framework, algorithms, and implementation of attention-based epistemic planning. Investigate the (theoretical and practical) efficiency of the developed algorithms.

Goal 3 Implement epistemic planning with attention in multi-agent simulations and humanoid robots, and evaluate the dynamics and quality of agent interactions (including human- robot interactions).

Project Members

Project members of the AiEP project:

• Gaia Belardinelli (postdoc in AiEP from October 2026)
• Thomas Bolander (PI)
• Hans van Ditmarsch, IRIT Toulouse.
• Ludovico Deponte (PhD student in AiEP, 2025-2028)
• Hermine Grosinger, University of Örebro.
• Jens Ulrik Hansen, University of Roskilde.
• Nir Lipovetzky, University of Melbourne.
• Rineke Verbrugge, University of Groningen.
• Sebastian Watzl, University of Oslo.

Events

Events organized and (co-)funded by the AiEP project.