Open-ended interactive robot learning in human-centered environments

Robots that operate in human-centered environments, such as homes and workplaces, must be able to understand everyday situations, interact naturally with non-expert users, and adapt to new tasks over time. In his thesis, Giorgos Tziafas investigates how such robots can become more interpretable, generalizable, and data-efficient by combining two key ideas: natural language grounding and memory consolidation.

First, Tziafas studies natural language as a shared representation between humans and robots. Language is used not only as a way for users to give instructions, but also as an intermediate layer that connects perception, reasoning, and action. Tziafas develops methods for language-guided grasping in cluttered household scenes, aiming to equip robot with the right tools to identify and resolve ambiguities, adapt to novelty and be robust under clutter.

Next, Tziafas focuses on lifelong robot learning with humans as teachers. He introduces memory-based approaches that allow robots to learn new object concepts, manipulation skills, and long-horizon behaviors from limited human guidance. By storing new knowledge in semantic, procedural, and episodic memories, robots can adapt during deployment while avoiding catastrophic forgetting.

Together, the thesis shows that language and memory are central mechanisms for building robots that can be instructed, corrected, and taught by everyday users in dynamic human environments.