Balint Gucsi

MINDS CDT doctoral researcher | University of Southampton | MINDS CDT, 2021

Helping social robots adapt to individual users

Balint Gucsi is developing ways for social robots to adapt their decisions and behaviour to the person they are working with. His research aims to make human-robot interactions more useful and less frustrating by helping robots respond to individual preferences, characteristics and tolerance levels.

Profiles: Research website | LinkedIn | Google Scholar | ORCID | GitHub

Research at a glance

Research focus	User-aware, adaptive human-robot interaction
MINDS theme	Agent-Based Adaptive Systems
Methods	Multi-armed bandit learning, constrained decision-making, multimodal signal processing
Robots used	TIAGo and Pepper humanoid robots
Open data	Two multimodal human-robot interaction datasets on Zenodo
Current status	Doctoral researcher

Robots that respond to the person in front of them

Social robots are increasingly being designed to help people with everyday tasks, from passing objects during a shared activity to finding a colleague in an office. Most still follow the same behaviour with everyone. People differ in their preferences, in how much interruption or delay they will tolerate and in how they communicate, so a robot that treats every user in the same way can easily become a source of frustration.

Balint’s research investigates how robots can behave in a user-aware way. He develops decision-making and learning frameworks that allow a robot to adjust its behaviour during an interaction, using what it learns about the individual user.

That learning can draw on explicit feedback, such as what a person says to the robot, and on implicit signals people show without meaning to, such as facial expressions and body posture. Implicit signals are useful but imperfect: they can be noisy, inconsistent and vary between individuals and cultures. Part of Balint’s work is designing approaches that use these signals while accounting for those limitations.

Why user-aware robots matter

Research approach

Much of Balint’s work builds on multi-armed bandit learning, a family of methods for making repeated choices under uncertainty. A robot using this approach tries different actions, learns which ones work well for a particular person and balances exploring new options against using what it already knows.

He combines this with constrained decision-making methods, including knapsack algorithms, which help a robot make the best choices within limits such as how often it should interrupt or ask a user for input. To estimate how a person is experiencing an interaction, he develops models of user frustration informed by psychological research.

These approaches are evaluated through simulation using realistic human-behaviour data and through user studies with humanoid robots. Studies have included participants collaborating with a TIAGo robot on an assembly task and interacting with a Pepper robot in a cafeteria-style scenario, where the robot made recommendations and took orders.

Research keywords: Human-robot interaction, social robotics, adaptive AI, multi-armed bandits, multimodal signal processing, implicit feedback, user frustration

Research and outputs

Balint’s doctoral research has been published at leading international robotics and human-robot interaction conferences. The selection below shows how the work has developed.

HRI-SENSE: A Multimodal Dataset on Social and Emotional Responses to Robot Behaviour

Presented at the 20th ACM/IEEE International Conference on Human-Robot Interaction (HRI 2025), Melbourne. This paper introduces a dataset recording how people respond, socially, physically and emotionally, to different robot behaviours while working with a TIAGo robot on a shared assembly task. It contains over six hours of interaction from 146 sessions with 18 participants.

View the HRI-SENSE paper on IEEE Xplore

User-Aware Collaborative Learning in Human-Robot Interactions

Presented at the 2025 IEEE International Conference on Robotics and Automation (ICRA 2025), Atlanta. This paper develops frustration-aware learning and decision-making models for human-robot collaboration, evaluated in simulation and in a user study with a TIAGo robot.

View the User-Aware Collaborative Learning paper on IEEE Xplore

RoboButler: Frustration-Aware Assistive User Localisation for Social Robots in Office Environments

Presented at the 34th IEEE International Conference on Robot and Human Interactive Communication (RO-MAN 2025), Eindhoven. This paper looks at how a robot could find and assist colleagues with errands of varying importance in an office, without continuous contact or live feedback from users, while aiming to keep frustration low.

View the RoboButler paper on IEEE Xplore

Investigating the Role of Implicit Signals in Adaptive User-Aware Human-Robot Interactions

Presented at the 2026 IEEE International Conference on Robotics and Automation (ICRA 2026), Vienna. This paper explores how signals such as facial expressions can be built into a robot’s decision-making, tested in a user study with a Pepper robot.

View the implicit signals paper on ePrints Soton

Open datasets

Balint has made two multimodal datasets from his user studies openly available, so that other researchers can study how people respond to robot behaviour.

• HRI-SENSE: A Multimodal Dataset on Social and Emotional Responses to Robot Behaviour (Zenodo, 2024)
• Investigating the Role of Implicit Signals in Adaptive User-Aware Human-Robot Interactions: A Multimodal Dataset (Zenodo, 2025)

Research profiles

• Balint Gucsi on Google Scholar
• Balint Gucsi on ORCID
• Balint Gucsi on IEEE Xplore
• Balint Gucsi on GitHub
• University of Southampton research project: User-aware adaptive human-robot interaction