Development of Hierarchical Concepts in Humanoid Robots
One of the important problems of a developing robot or an organism is the development of concepts and linking it with language. Concepts are very important for intelligence and cognition for, due to them, an organism can generalize and abstract over his experiences; he can transfer his experiences to other situations or events; he can think, reason, remember, and talk about situations, events or entities; and, maybe the most important of all, concepts are essential in language comprehension.
Although there is abundant theoretical literature on concepts, there is little computational work on how a developing robot or an organism can acquire concepts and related these acquired concepts to language and these problems, as can be understood from the lately funded EU and TUBITAK projects, are beginning to attract the attention of many research groups. The existing studies focus just on first-order concepts and mostly focus on affordance-based or haptic-based concept formation. However, first-order concepts (such as APPLE, ORANGE, DOG, CAT2 etc.) are not sufficient for a robot which we expect to interact with us, humans; an intelligent and cognitive organism should be able to abstract over abstractions (and acquire concepts like FRUIT, ANIMAL etc.), i.e., acquire concepts of the concepts. For example, when shown an apple and told “Look this is an apple”, the robot should be able to link the word “apple” to the APPLE concept; however, when later told a different sentence for an apple like “Look this is a fruit”, the robot should be able to handle the conflict in using different words for the same concept. We claim that resolving such conflicts and communication/interaction with humans at this level require hierarchical representations of concepts.
As known from developmental psychology, main modalities affecting concept formation are: language, appearance and function; i.e., babies use the symbols from adults (i.e., words), the shape and the texture, and the affordances of the objects for grouping and conceptualizing them. This way, a single object (for example, a little red rotten apple) can be a part of many different concepts (e.g., APPLE, ROLLABLE, FRUIT, THROWABLE, GROWS-ON-TREE) based on the context and the goal.
In this project, we will study how a (cognitively) developing and embodied humanoid robot can acquire a hierarchical representation of concepts from its experiences. For this goal, by going beyond the current literature, we will use language, appearance as well as affordances of objects and investigate how these three modalities can affect the formation of the hierarchy. The proposed methods and mechanisms will be demonstrated on a concise scenario involving a humanoid robot iCub interact with a human on a clearly defined task.
Invited Talks, Seminars
Videos (coming soon)