MINDS alumnus | Postdoctoral Research Associate, University of Liverpool
Using human feedback to build safer autonomous agents
Research focus: Leveraging human feedback to build safer autonomous agents
Ilias Kazantzidis is a MINDS alumnus whose doctoral research investigated how human feedback can help autonomous AI systems learn and operate more safely.
Career at a glance
| MINDS status | Alumnus |
| Graduated | June 2026 |
| Current role | Postdoctoral Research Associate, University of Liverpool |
| Doctoral project | Towards Safe Autonomous Agents via Human Input |
| Research focus | Leveraging human feedback to build safer autonomous agents |
Doctoral research
Ilias’s research aimed to make autonomous AI systems, including robots and self-driving cars, safer, more reliable and better aligned with human intentions and expectations.
These systems may learn to complete a task without fully understanding what people consider appropriate or safe behaviour. Ilias’s work built on learning from human preferences by enabling users to explain not only which behaviour they preferred, but also why. This helps AI systems understand and prioritise important considerations relating to safe learning and deployment.
Ilias’s PhD research explored how human feedback can help autonomous AI systems learn to make safer decisions. A major challenge is that these systems often learn through trial and error, known as reinforcement learning, which can be dangerous in real-world applications such as robotics and self-driving cars.
He developed two approaches, JPAL-HA and DROPJ, that allow people to guide an AI agent by comparing possible actions or behaviours and explaining why they prefer one over another, particularly where safety is involved. This additional information helps the agent learn to avoid unsafe behaviour, either through direct interaction with a person or within a virtual environment learned from previous experience.
His research combined learning from human preferences, safe reinforcement learning, world models and model predictive control. Experiments, including a study involving 40 human participants interacting with a real robot, showed that these methods could significantly improve safety without placing excessive demands on the people providing feedback.
Current role
Ilias is now a Postdoctoral Research Associate at the University of Liverpool, working on robotics and artificial intelligence for chemistry laboratory automation. His research focuses on learning-based robotic manipulation methods, including vision-language-action models and reinforcement learning, with an emphasis on safety and reliability.
He is also involved in an EPSRC-funded Centre-to-Centre project between AIchemy and the Acceleration Consortium, which aims to develop transferable robotic workflows for chemistry laboratories in different countries.
Outreach and public engagement
Together with fellow MINDS CDT researchers, Ilias organised and delivered an outreach activity at the Southampton Science and Engineering Festival in 2022. The six-hour activity attracted 315 visitors, with 59 people taking an active part.
The activity used the JPAL-HA model developed during Ilias’s PhD to train a Thymio mobile robot to navigate a small arena, find the shortest route to a goal and avoid dangerous areas. Visitors could train the robot using a laptop interface or natural-language instructions, demonstrating how human feedback can help autonomous systems learn more safely.