The rapid evolution of robotics is promoting new robotics related research fields to emerge. Taking insights from developmental psychology, developmental robotics is a new field which aims to endow robots with capabilities that enable them to life-long learning in an open-ended way. There are situations where engineers or designers cannot foresee all the possible problems a robot may encounter. As the number of tasks that a robot must do grows, this problem becomes more evident and traditional engineering solutions may not be entirely feasible. In that case, developmental robotics provides a series of principles and guidelines to construct robots which have the adequate cognitive tools in order to acquire the necessary knowledge. Self-exploration, incremental learning, social scaffolding or imitation. All are tools which contribute to build robots with a high degree of autonomy. By means of internally motivated self-exploration, a robot discovers what its body is able to do. Incremental learning techniques enable a robot to have ready-to-use knowledge by building new cognitive structures on top of old ones. Social scaffolding and imitation capabilities allows taking advantage of what humans --- or other robots --- already know. In this way, robots have goals to pursue and provide either an end use of learned skills or examples on how to accomplish a given task. This thesis presents a study of a series of techniques which exemplify how some of those principles, applied to real robots, work together, enabling the robot to autonomously learn to perform a series of tasks. We also show how the robot, by taking advantage of active and incremental learning, is able to decide the best way to explore its environment in order to acquire knowledge that best helps in accomplishing its goals. This, in addition to the autonomous discovery of its own body limitations, leverages the amount of domain specific knowledge that needs to be put in the design of the learning system. First and foremost, we present an incremental learning algorithm for Gaussian Mixture Models applied to the problem of sensorimotor learning. Implemented in a mobile robot, the objective is to acquire a model which is capable of making predictions about future sensory states. This predictive model is reused as a representation substrate which serves to categorize and anticipate situations such as the collision with an object. After an extended period of learning, and having encountered different situations, we observed that the acquired models become quite large. However, we realized that, at any given time, only small portions of it are used. Furthermore, these areas are consistently used over relatively long periods of time. We present an extension to the standard Gaussian Mixture Regression algorithm which takes advantage of this fact in order to reduce the computational cost of inference. The techniques herein presented were also applied in a different and more complex problem: the imitation of a sequence of musical notes provided by a human. Those are produced by a virtual musical object which is used by a humanoid robot. The robot not only learns to use this object, but also learns about its own body limitations. This enables it to better understand what it is able to do and how, highlighting the importance of embodiment in the interaction of a robot with its environment and the kind of cognitive structures that are formed as a consequence of this type of interaction.
|Date of Award||23 Mar 2015|
|Supervisor||Ramon López de Mántaras (Director), Jesus Cerquides Bueno (Director), Yiannis Demiris (Director) & Ricardo Juan Toledo Morales (Tutor)|
- Artificial intelligence
- Machine learning