Nathan Gavenski is a PhD candidate in the Department of Informatics, King's College London, and is part of the UKRI Centre for Doctoral Training in Safe and Trusted AI and the Distributed AI Research Group. Supervised by Dr Odinaldo Rodrigues and Dr Matteo Leonetti, his research addresses a core limitation of imitation learning: agents trained to replicate expert behaviour often fail once deployed beyond their training conditions. Nathan develops methods that allow agents to generalise from observation alone, with contributions published at IJCAI, AAMAS and ECAI. A UKRI studentship fully funds his PhD.

Before his PhD, Nathan worked as a software developer in Brazil. He graduated Summa Cum Laude with an MSc in Computer Science from PUCRS, where he also taught as a Visiting Lecturer in Deep Learning and received a Google Research Award for Inclusion for work on Brazilian Sign Language translation.

Nathan also works as a Research Developer at the University of Aberdeen on a UKRI Frontier AI Discovery-funded project on formally verified goal recognition, alongside his PhD research. He is also a Teaching Assistant at King's and was nominated for the Outstanding Teaching Assistant Award in 2025.

Project Title: Generalisable Imitation Learning

Abstract: Imitation learning is a learning approach in which agents learn how to act in an environment by mimicking other proficient sources, such as other agents and humans. The field of imitation learning lies at the intersection of various other fields but borrows most of its techniques and processes from machine learning and reinforcement learning. In recent years, many researchers have focused on imitation learning, significantly increasing the number of researchers and publications. We argue that although imitation learning agents might achieve better results in each new publication, many of the original ideas of a socially inspired learning approach may have been neglected in favour of other aspects, such as training efficiency via machine learning techniques. Therefore, in this project, we examine the imitation learning literature and compare it with other fields. By doing so, we hope to find ways to achieve more human-like behaviour in imitation learning agents and avoid some of the common pitfalls of current approaches, such as susceptibility to non-optimal samples. Lastly, we explore the case for optimal teacher feasibility and present a toolkit and an environment.

Research Interests 

• Imitation Learning
• Generalisation
• Agentic Systems

Research Centres or Groups

• Distributed Artificial Intelligence 
• UKRI Centre for Doctoral Training Safe & Trusted AI 

PhD Supervision

• Primary Supervisor: Dr Odinaldo Rodrigues
• Secondary Supervisor: Dr Matteo Leonetti

Websites

• PURE Profile