Dissertation defence (Information and Communication Technology): MSc (Tech) Sajad Shahsavari

MSc (Tech) Sajad Shahsavari defends the dissertation in Information and Communication Technology titled “Towards Machine Learning-Based Digital Twinning for Autonomous Mobile Robot Control” at the University of Turku on 13 November 2026 at 12.00 (University of Turku, Natura, IX, Turku).

Opponent: Professor Sadok Ben Yahia (University of Southern Denmark, Denmark)

Custos: Professor Juha Plosila (University of Turku)

Summary of the Doctoral Dissertation:

Mobile robots will soon change the way we live both at home and at work, because they are becoming more intelligent and easier to collaborate with. In contrast to previous generation robots that were merely mechanical operators in industrial environments, now robots can “understand” their surrounding environment and perform long-horizon actions on it. Thanks to the modern AI models, robots now extract abstract knowledge on their environments, such as various types of objects, humans, tasks, etc., by processing vast amounts of sensory data. This intelligence, however, comes with the cost of heavy computing power on a robot working with a relatively small battery. 

The focus of this thesis is to present modern robotic systems as a compound cyber-physical system that is composed of multiple different parts such as mechanical motors, computing boards, sensors, communication equipment, and battery unit. Our claim is that these parts need to be managed jointly together so that the robot can exhibit its best performance. In simple words, the idea is that the required computing resources depend on the motion of the robot and vice versa. This means that the robot must constantly adjust its motion and computing resources in tandem to each other so that it performs optimally. For example, when a robot drives faster, its computer must process camera images faster to avoid obstacles. When it slows down, the computer can take it easier. 

In one experiment in this thesis, we experimented this claim on an autonomous ground rover and demonstrated that (1) computing power is comparable to the mechanical power for motion, and (2) adjusting the computing resources at runtime with respect to robot’s motion can save up to 39% in its energy consumption while keeping the task accomplished. 

To achieve these, we developed a virtual model of the robotic system, which is called a Digital Twin, and utilized its predictions to assess various future scenarios. This virtual model is then used to learn the actions that the robot should take.

This thesis makes robots more affordable and efficient. In practice, this means robots can work longer on a single charge, and their batteries can last more years before they need replacing. On the other hand, using the outcomes of this research, one can make robots smarter by spending the saved energy on smarter AI models.

This research has opened new novel directions that are being actively pursued in our research group. One promising example is taking the battery subsystem into account with its short- and long-term dynamics to extend robot usage lifespan. Another research direction is to apply this compound cyber-physical overview to a swarm of collaborative robots that can enable them to outsource computation to their peers or use others collected knowledge.