Väitös (tieto- ja viestintätekniikka): MSc Sahar Salimpourkasebi

MSc Sahar Salimpourkasebi esittää väitöskirjansa ”Deep learning for Robust Multi-Robot Perception and Control” julkisesti tarkastettavaksi Turun yliopistossa tiistaina 27.10.2026 klo 12.00 (Turun yliopisto, TIERS Lab, Tieronkatu 8, Turku).

Vastaväittäjänä toimii tohtori Matti Kutila (Teknologian tutkimuskeskus VTT Oy) ja kustoksena apulaisprofessori Widhi Atman (Turun yliopisto). Tilaisuus on englanninkielinen. Väitöksen alana on tieto- ja viestintätekniikka.

Tiivistelmä väitöstutkimuksesta:

Robots are increasingly expected to work alongside people and other robots in everyday environments. To do this reliably, they need to understand their surroundings, recognise misleading information, and move appropriately when conditions change. My dissertation investigates how artificial intelligence and sensor measurements can help robots meet these challenges. It brings together methods for determining where robots are, assessing the information they share, and learning how to move and respond to human guidance.

The first part of the research focuses on detecting unreliable information and building trust between robots. By comparing distance measurements, robots can identify sources of faulty information and improve their estimates of where they are relative to one another. Unreliable information can also come from other sensors, including cameras. The research shows how camera images can help robots detect new objects or changes in their surroundings without retraining the artificial intelligence models for every location. Comparing images collected by different robots also helps identify a robot that may be sharing misleading observations. A shared set of rules allows the team to distribute responsibility for deciding which information to trust.

The second part examines how robots can learn to move under different forms of guidance. In a “follow-me” setting, robots learn from examples to follow a leader accurately and smoothly using distance and direction measurements. The research also demonstrates how people can give robots commands through body postures recognised by small wearable sensors. For movement without direct human guidance, robots learn to navigate and avoid obstacles through practice in a computer simulation. Experiments show that this learning can be transferred to physical robots without retraining their movement models in the real world, although moving obstacles remain more challenging.

Together, these findings provide building blocks for more reliable robots that cooperate with one another and respond to people. Potential applications include inspecting buildings and industrial sites, transporting goods, and supporting people with mobility. By helping robots recognise unreliable information and learn useful movement skills, the research supports their wider use in everyday environments.