Creating Curious Robots

Applying machine learning to a robotics problem typically requires substantial human oversight to design the learning system, tune the parameters, define the task, determine the input and output representations, and create the training data set. In contrast, biological organisms are able to learn autonomously from unlabeled data in an open-ended fashion. Developmental robotics is an emerging field that strives to build better robots by applying insights from biological developmental processes. In this talk I will review several recent approaches from developmental robotics that use prediction to generate teaching signals. This results in a task-independent kind of learning in which the robot focuses on novel stimuli.