Abstract
The area in which this thesis is based on, is referred to as Adaptive Robotics. The main objective of this approach is to synthesize agents that evolve or develop their skills autonomously through interaction with a natural or artificial environment. The research aims at identifying the possibility of self-organizing systems to build an internal representation of the input space, able to handle the case of an unknown environment, and moreover, examine the causes, consequences and solutions to the conflicting problem of catastrophic forgetting that these systems are prone to.The ability to navigate is arguably the most fundamental competence of any mobile agent, besides the ability to avoid basic environmental hazards. Recent studies of insect behavior and navigation reveal a number of elegant strategies that can be valuable when applied to the design of autonomous robots, without the need for higher level cognitive processes such as object identification and labelling. These bio-mimetic approaches have been reviewed, focusing on a self-organizing cognitive model of mental development, that allows for a common description of biological map building behavior. The motivation came from the assumption that self organizing systems could be used to reduce the amount of predefinition put in by a human operator and the ability to address noisy, inconsistent or no meaningful information with respect to the task being performed. In addition, a visual interpretation scheme, mimicking simple cells in the primary visual cortex, have been examined and critically analysed.
The research undertaken resulted in the development of a novel rehearsal map building scheme that is proven to build a representation of the environment, sequentially, from acquired visual snapshots of physical locations. These results also demonstrate the ability of the scheme to efficiently address the plasticity elasticity dilema presented by various connectionist models such as the self organizing map (SOM), neural gas (NG) and growing neural gas algorithm (GNG). This is encouraging enough to prompt further research that could result in an autonomous agent capable of self-localizing with a satisfactory degree and reliability in unknown and dynamically changing environments.
This thesis also explores the advantages of evolutionary sub-goal robot navigation with a cognitive map architecture. Experiments in simulation show that an evolved robot, adapted to both exteroceptive and proprioceptive data, was able to successfully navigate through a list of sub-goals minimizing the problem of local minima in which evolutionary process can get trapped. The results demonstrate that a navigation behavior could be learned without the need for an in depth knowledge of the problems to be solved, especially in highly complex environments.
| Date of Award | Jan 2015 |
|---|---|
| Original language | English |
| Supervisor | Andrew Ware (Supervisor) |