UNIVERSITY OF HERTFORDSHIRE COMPUTER SCIENCE RESEARCH COLLOQUIUM presents "Object Empowerment: Learning Tool Use through Intrinsic Motivation" Dr. Faizan Rasheed (University of Hertfordshire) 20 May 2026 13:00 -14:00 Room B160 Everyone is Welcome to Attend Refreshments will be provided Abstract: Artificial agents are often trained to achieve predefined goals, yet many real-world scenarios require autonomous exploration and the discovery of useful behaviours without dense external rewards. Intrinsic motivation offers a promising route to such capabilities. Among existing approaches, empowerment provides a principled notion of control by measuring how much influence an agent can have over its environment. However, standard formulations of empowerment do not explicitly account for objects or tools, which are central to intelligent behaviour. In this talk, I introduce object empowerment, a formulation that quantifies an agent’s influence over specific environmental objects. This provides a natural framework for modelling tool use, where tools act as intermediaries that extend the agent’s influence over task-relevant objects. I then show how object empowerment can be integrated into reinforcement learning as an intrinsic reward, enabling agents to discover functional relationships between tools and objects even in sparse-reward settings. Results from grid-world and MiniHack environments demonstrate improved exploration, faster learning, and interpretable tool-selection behaviour. Finally, I discuss extensions to multiple tools and objects, and introduce empowerment-based measures for characterising tools in terms of persistence, latency, and reliability. Together, these results position object empowerment as a unified framework for understanding and generating tool-use behaviour in artificial agents. --------------------------------------------------- Hertfordshire Computer Science Research Colloquium http://cs-colloq.cs.herts.ac.uk