Team

I am looking for motivated PhD students to join my team. For more information about available projects check here!

Current members

PhD Students

Aleksandra Stawinska (2025-2029)
Anthropomorphic Cues and Trust in Human-Robot Interaction
Coadvised with Prof Lasana T. Harris and Prof Nadia Berthouze
PhD in Experimental Psychology
Department of Experimental Psychology, University College London
Abstract
Abstract:
As robots are becoming increasingly integrated into social and applied contexts, understanding the factors that shape trust in human-robot interaction is becoming ever more important. Anthropomorphic cues are often proposed to influence trust and engagement, yet it remains unclear how these processes unfold and how different aspects of anthropomorphism shape people’s perceptions of robots. In particular, human-like appearance and behavioural features may make distinct contributions to expectations of a robot’s warmth and competence.
This research aims to examine how anthropomorphic cues shape trust in robots. It explores how trust is formed, repaired and maintained in interactions, and how people perceive and evaluate robots in social and assistive settings. In doing so, this work contributes to a broader understanding of trust in human-robot interaction and helps inform the development of robots that support appropriate calibration of trust.
Adithya Venkatadri Hulagadri (2025-2029)
The Evolution of Deception Through Emergent Communication
Coadvised with Dr Ștefan Sarkadi
PhD in Computer Science, Safe & Trusted AI CDT
Department of Informatics, King’s College London
Abstract
Abstract:
Deception is the act of one agent manipulating another agent to believe a statement that the first agent himself believes to be untrue. This is very common in competitive environments such as zero-sum games as well as real world applications such as espionage. A crucial prerequisite for successful deception is to have a good model of the agent’s reasoning, so as to take advantage of its blindspots. This type of reasoning can quickly become expensive to compute, as the other agent may also be aware that its internal processes are being modelled, and may resort to similar modelling of the attacker to defend its interests. This is often modelled using interactive partially observable Markov decision processes (I-POMDPs), which can contain recursive models of other agents. This perspective-taking ability, also called Theory of Mind, is often used by humans to communicate, collaborate with and deceive other humans. This ability develops in children around the age of 3-4 years old, and is measured by their ability to recognise that other humans may have false beliefs (the Sally-Anne tests). Therefore, this project seeks to leverage the insights from cognitive science on how humans can solve puzzles that require perspective taking to inspire efficient ToM artificial architectures. Humans are also hypothesised to simultaneously develop language and perspective taking abilities, and this thus opens the door to studying how language, theory of mind capabilities and consequently deception can emerge naturally with the right evolutionary conditions. The applications of this work are potentially wide-ranging, with applications for more empathetic AI applications whose actions can be explained using theory of mind approaches. Empathy is a double-edged sword: used for good, it can magnify the impact of AI applications for humans, and conversely, can be used to take advantage of our vulnerabilities. To defend oneself from deception, one must first learn to how he might be deceived.
Ma’ayan Armony (2024-2028)
Common Sense Planning for Robotics
Coadvised with Dr Albert Meroño-Peñuela
PhD in Computer Science, Safe & Trusted AI CDT
Department of Informatics, King’s College London
Abstract
Abstract:
Robot task planning techniques are being developed constantly, aiming to maximise the feasibility and eﬀiciency of the generated plans, to better achieve the desired goals. Language models are being incorporated into planning strategies more frequently, and the use of LLMs to improve planners is now common. Due to LLMs training on textual corpora, in which common sense knowledge is rarely included, they lack certain knowledge which is intuitive to humans.
This PhD aims to leverage common sense knowledge graphs to increase the safety of planning approaches and the user’s trust in them. This is to be achieved by integrating knowledge about human behaviour into the planning process, whether for plan creation, evaluation, or repair.
Lennart Wachowiak (2021-2026)
Knowing When and What to Explain: Explanation Strategies in Human–Robot Interaction
Coadvised with Dr Oya Celiktutan and Dr Andrew Coles
PhD in Computer Science, Safe & Trusted AI CDT
Department of Informatics, King’s College London
Abstract
Abstract:
Background. To generate robot behavior, researchers employ a multitude of complex technical solutions, such as large neural networks for dialog and vision, or search-based planning algorithms for determining high-level actions. For a user interacting with such a robot, it consequently becomes increasingly difficult to understand why a robot acts the way it does. The field of explainable robotics tries to tackle this problem by enabling the robot to communicate insights into its internal decision-making.
Research Gap and Aim. However, much of the current explainability research considers explanations in a vacuum without taking the full interaction context and the user into account. Human–robot interactions are characterized by complex dynamics, with users’ mental models changing over time and users exhibiting varying needs regarding explanation timing and content. Therefore, we investigated two novel aspects of explanations in human–robot interactions. Firstly, we examined the timing of explanation by asking:
- When do users want an explanation, and can robots detect such situations to offer explanations proactively?
Second, we examined the issue of selecting the right information, asking:
- What questions do users ask, and how can a robot select the information most relevant to answer users’ questions?
Findings I: User Needs and Literature Analysis. As a first step, we grounded our research in the findings of the preceding literature and the needs and preferences of users. Based on a scoping review of the field, we present a taxonomy of explanation types and provide an overview of the causes driving the need for explanations. We then verified these results empirically through a large-scale online survey. In this study, we asked around 200 participants about their communication preferences in various interaction scenarios. The results demonstrate the value of explanations across a wide range of scenarios, such as when a robot makes errors or is uncertain about how to solve a problem.
Findings II: Proactive Explanations. In a second step, we used these findings to show how a robot can identify critical moments that require the robot and the user to communicate. To achieve this, we demonstrate how social cues such as eye gaze, facial expressions, speech features, and posture can be used to train multimodal time series classification models for detecting user confusion and robot errors. Additionally, we evaluate the potential role of large vision–language models in detecting users’ communication needs from video and text data.
Findings III: Explanation Selection. To advance user-centered explanation selection, we first provide a dataset of questions commonly asked by users to household robots. We then sketch out how to design a neurosymbolic system that uses large language models to collate the information most relevant to a user’s question from a range of symbolic and sub-symbolic explainability components, such as causal traces from classical planners or attention-based explanations from neural networks for object detection.
Impact. With this thesis, we provide HRI researchers with insights into when users want explanations and what the explanations should be about. These insights can be used to design more explainable and user-friendly interactions. Additionally, we provide technical solutions for integrating the findings into robotic systems. Applied to real-world robot design, these findings can improve users’ experience of interacting with robots, allow for appropriate trust calibration, and increase safety in human–robot collaborations.

Former members

Postdocs

Benjamin Krarup (2023-2026)
Projects: Advancing Impact in AI Planning and Robotics, COHERENT
Coadvised with Dr Andrew Coles
Department of Informatics, King’s College London

PhD Students

Peter Tisnikar (2021-2025)
Capability Estimation for Adaptive Human–Robot Collaboration
Coadvised with Dr Matteo Leonetti
PhD in Computer Science, Safe & Trusted AI CDT
Department of Informatics, King’s College London
Currently a Postdoctoral Researcher at University of Oxford
Abstract
Abstract:
This project will investigate service robot adaptation across two time scales: short term scale, where the goal is to adapt to a single user, and a long-term adaptation framework, where individual user adaptation is reduced and optimised through discovering patterns and commonalities between different users. The particular focus of this project will be assistive robotics, where tasks like feeding and shoe dressing will be investigated.
The proposed output of this project is a learning framework and a behaviour adaptation algorithm that builds an interaction model with the user by inferring beliefs about the user’s abilities and physical properties from ad-hoc collaboration (table-top assistance, putting a shoe on). It then uses this learned model to update its global beliefs about any future interaction based on key insights from the single interaction. This should allow the framework to operate at very long timescales without need for recalibration or wiping of memory or experience. This work will also be underpinned by physical experiments, where a service robot will perform in a variety of scenarios that prove the framework design.

Visiting PhD Students

Alberto Olivares-Alarcos (2023)
Contrastive Explanations of Robot Plan Adaptation
Coadvised with Dr Guillem Alenyà and Dr Sergi Foix
PhD in Automatic Control, Robotics and Computer Vision
Institut de Robòtica i Informatica Industrial (IRI, CSIC-UPC)
Department of Informatics, King’s College London
Amar Halilovic (2024-2025)
Planning of Robot Explanations
Coadvised with Dr Senka Krivić
PhD in Computer Science
Universität Ulm
Department of Informatics, King’s College London
Silvia Izquierdo-Badiola (2022-2023)
Use of LLMs for Goal Generation and Task Adaptation in Robot Planning Problems
Coadvised with Dr Guillem Alenyà and Dr Carlos Rizzo
PhD in Automatic Control, Robotics and Computer Vision
Institut de Robòtica i Informatica Industrial (IRI, CSIC-UPC) and Eurecat
Department of Informatics, King’s College London

Undergraduate Students

Nicole Lehchevska (2022-2023)
Multi-Floor Robot Navigation in Indoor Buildings (BSc Thesis) (2023)
BSc in Computer Science
Department of Informatics, King’s College London
King’s Undergraduate Research Fellowship (KURF) (2022)
Coadvised with Dr Sachini Weerawardhana and Dr Peta Masters
BSc in Computer Science
Department of Informatics, King’s College London
Elisabeth Halvorsen (2022)
King’s Undergraduate Research Fellowship (KURF)
Coadvised with Dr Sachini Weerawardhana and Dr Peta Masters
BSc in Computer Science
Department of Informatics, King’s College London
Maria Vila (2018-2019)
Autonomous User Feeding by a Physically Assistive Robot (2019)
Coadvised with Dr Guillem Alenyà and Dr Cecilio Angulo
BSc in Computer Science
Facultat d’Informàtica de Barcelona (FIB-UPC) / IRI
FIBAlumni award to the best 2019 BSc thesis in Computer Science at FIB
Abstract

Abstract:
In this BSc thesis, we propose a robot capable of feeding a person autonomously. Its behavior takes into account the state of the world and the user requirements in order to decide which are the best set of actions that need to be performed in each situation.
Safe Feeding Strategies for a Physically Assistive Robot (2018)
Coadvised with Dr Guillem Alenyà and Dr Cecilio Angulo
BSc in Industrial Technology Engineering
Escola Tècnica Superior d’Enginyeria Industrial de Barcelona (ETSEIB-UPC) / IRI
Abstract

Abstract:
The goal of this project is to program a robot to be able to feed a person with reduced mobility. To do so, a planning of the actions that the robot will follow will be computed, given some sensor-based inputs. These inputs will consist firstly in the detection of the dish that contains the food and the amount of food left, and secondly in the detection of the user’s face and whether the mouth is open or not. Finally the motions of the robot will be designed to be natural and safe. All the movements of the robot will comply with the established regulations.