Biologically inspired goal directed navigation for mobile robots

Abstract

This project involved an investigation into low-cost navigation of mobile robots with the aim of creating and adaptive navigation system inspired by behaviour seen in animals. The navigation module developed here would need to be able to successfully localise a robot and navigate it to a defined target. A critical literature review was carried out of current localisation and path-planning architectures and a bio-inspired approach using an Echo State Network and Liquid State Machine architecture was chosen as the base for the navigation modules. The navigation module implemented in this work is trained to navigate and localise itself in different environments drawing its inspiration from the behaviour of small rodents. These architectures were adapted for use by a robot with a view on the physical implementation of these architectures on an embedded low-cost robot using a Raspberry Pi computer. This robot was then built using low-cost, noisy proximity sensors which formed the inputs to the navigation modules. Before the deployment on the embedded robot the system was tested and validated in a full physics simulator. While the training of the Echo State Networks and Liquid State Machine has been carried out in the literature by the offline method of linear regression, in this work we introduce a novel way of training these networks that is online using concepts from adaptive filters. This online method increases the adaptability of this system while significantly decreasing its memory requirements making it very attractive for low-cost embedded robots. The end result from the project was a functioning navigation module using an Echo State Network that was able to navigate the robot to a target position as well as learn new paths, either using offline or online methods. The results showed that the Echo State Network approach was valid both in simulation and practically as a base for creating navigation modules for low-cost robots and could also lead to more efficient and adaptable robots being developed if the training was carried out in an online manner. The increased computational complexity of implementing the liquid State machine on analytical machines however made it unsuitable for deployment on robots using embedded micro-controllers.