Julia Ferraioli, Leslie McTavish and George Dahl
Mentor: Professor Doug Blank
Developmental robotics involves designing learning systems for robots such that they are capable of displaying sophisticated behaviors that have not been directly programmed into them. Traditionally, robots can only do what we program them to do. They are not capable of acquiring any new information and therefore are not able to adapt to new situations. Because of the limitations of this traditional approach, new methods are being explored.
The developmental approach to learning attempts to create a models that mimic human learning development. The robot should first develop a sense of its self and its environment by detecting and establishing a relationship between its sensor and motor values. We will try to do this by experimenting with combinations of various established learning algorithms.
Learning algorithms are methods by which the system autonomously makes generalizations about input which is provided to it. These algorithms allow the system to learn a task or task without explicit programming or memorization. To achieve our goal we will design developmental algorithms which incorporate several learning processes such as neural networks, self-organizing maps (SOMs), resource allocating vector quantizers (RAVQs), growing neural gas (GNG) as well as others.
In the next stage of development, the robots will be assisted by social interactions with other robots or human guidance. We believe social interaction is an important aspect of human learning and may also prove to be useful in developing intelligent robots. Our experiments will test the effect of social interaction in the developmental process. We hope to design a developmental learning system that can be applied to both simulated and real robots and is independent of environment and robot type.