e-MDBa cognitive architecture for lifelong open-ended learning autonomy in robotic systems

Abstract

The field of autonomous robotics is currently facing the challenge of designing robots with levels of autonomy that allow them to be useful in complex and changing service-related domains. This implies designing systems capable of operating in an open-ended manner, i.e., capable of carrying out useful tasks in domains that were not known at design time. It also means that robots must be able to reuse and integrate knowledge acquired in different and possibly distant learning processes, making subsequent learning challenges progressively more accessible. That is, they must be capable of lifelong learning. This doctoral thesis seeks to explore the basic principles that would be required to achieve this new level of autonomy: lifelong open-ended learning autonomy (LOLA). We address the issue of how a designer can go about designing robots that must operate in LOLA settings and, more importantly, that are useful to the humans that build them. To formalize the proposal, we propose the bases of an architectural approach to support LOLA involving the development of two main components: a motivational system and a longterm memory (LTM). With the objective of testing them, these components are implemented in a prototype architecture called epistemic Multilevel Darwinist Brain (e-MDB), which is tested in several experiments considering real robots under LOLA settings. The results obtained open up a wide range of research topics that are of interest to complement the work carried out and improve LOLA operation in robots. Among them, the most relevant are those related with the capability of autonomously obtaining appropriate knowledge representations that can facilitate on-line abstraction and generalization. Some preliminary work in this line is also presented to provide a starting point for further research in the framework of cognitive architectures for LOLA.