Welcome, kind visitor!
My name is Stevan Tomic.
Here you can find out about my life trajectory relevant to my research, a breaf overview of my research, and get a sense of what makes me tick.
I received my bachelor’s degree in Computer Science from Technical University, Sofia, Bulgaria. For a couple of years, I have been working as an AI and gameplay developer in Ubisoft, Sofia. While the work was very interesting and stimulating, at that time, it could not fully satisfy my deep curiosity about the nature of intelligence and general principles behind human cognition. This drow me to enroll in master studies in Cognitive Science, which significantly affected my views on AI. I received my master's degree in Cognitive Science from New Bulgarian University in Sofia. After graduating I got an opportunity to do research and also work with robots as a Ph.D. student at AASS lab, Orebro, Sweden, supervised by Alessandro Saffiotti and Federico Pecora.
If robots are to be accepted into human society, it is essential that they are capable of reasoning about and learning from human social norms. Thus, the main research question in my study is how to make robots 'socially aware'. This includes the following underlaying questions:
- How to represent social rules understandable for humans, operable by for robots?
- How to reason about and learn behavior from (multimodal) norms specifications?
- How to re-use norms in novel domains (both in planning and learning)?
The early stage: Norms and Socially Aware Planning. My research started by extending the known principles behind context recognition, to support the recognition of social context in a situation. This allows me to model social norms depending on the context of the underlying social situation and adding this model in a planning domain language. Coupling planning with social context recognition allows proactive behaviors of autonomous robots directed towards achieving their normative requirements in human social environments. This novel approach to planning - socially aware planning (SAP), breaks the social space into social contexts with their sub-contexts, and generate normative (socially acceptable) plans in human-robot environments. During the Monarch project, the approach was demonstrated in many real-world problems, in which humans and robots were cooperating together. The approach is described here. The midle stage: Abstraction; Formal Model of Institutions for Robots. The main shortcoming of the discussed approach is that norms were ad-hoc, and consequently, changing objects, humans, or robots relevant for a given situation, had to be manually reflected in the planning domain language. It became evident that norms are abstract constructs independent of a particular physical domain. Addressing these problems led to the development of the formal normative framework based on the notion of institutions (often investigated in social sciences and multi-agent systems (MAS)). An important characteristic of the framework is that it clearly separates social constructs (norms) from the purely objective world (domain). While a domain describes agents, their behaviors, objects, and their dynamic (trajectory); social (institution) level describes norms through higher-level categories that have particular meaning for social interactions. Norms can describe deontic concepts like an obligation, but also temporal and spatial relations (and other relations of interests). Norms can be easily understood by humans, however, robotic agents must rely on the norm semantics, which provides detailed specifications of whether a physical execution (trajectory) adheres to a norm or not. Grounding an institution (a set of abstract norms) into a domain, bridge the gap between the abstract norms and physical execution, and also provides social meaning to that execution. The framework provides a way to systematically classify reasoning/learning problems concerning (social) behavior, such as the problem of interpretation or recognition of an institution given the trajectory (execution). My attention was directed on the following two problems:
- Verification: Whether physical execution adheres to norms or not;
- Planning: Generate adherent plan by automatic translation of abstract norms into concrete operators and constraints in planning domain language;
Another characteristic of the framework is that affordances in a domain are explicitly modeled, which makes several computational problems much easier to solve. The framework is described here.
Final Stage, Reinforcement Learning: Guidance, Interpretation and Abstract Learning (To be updated soon...) The list of my publication can be found here.
Inspiration and free time
In my free time, I like building things (like computer games or small robots), reading books, listening to podcasts about the future and philosophy behind AI, the human mind, and technology in general. Somehow, I am also very inspired by the advances in theoretical physics. Other than scientific/technology related stuff, I enjoy spending my time with my family and friends, travel, playing chess, etc. Biking, jogging, or hiking through the Swedish landscapes is a special kind of experience. My Artistic side expresses itself in (low-skill) guitar playing and even sometimes in 3D modeling (blender).