Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.809251
Title: Calibration of sound source localisation for robots using multiple adaptive filter models of the cerebellum
Author: Baxendale, Mark David
Awarding Body: University of the West of England
Current Institution: University of the West of England, Bristol
Date of Award: 2020
Availability of Full Text:
Access from EThOS:
Access from Institution:
Abstract:
The aim of this research was to investigate the calibration of Sound Source Localisation (SSL) for robots using the adaptive filter model of the cerebellum and how this could be automatically adapted for multiple acoustic environments. The role of the cerebellum has mainly been identified in the context of motor control, and only in recent years has it been recognised that it has a wider role to play in the senses and cognition. The adaptive filter model of the cerebellum has been successfully applied to a number of robotics applications but so far none involving auditory sense. Multiple models frameworks such as MOdular Selection And Identification for Control (MOSAIC) have also been developed in the context of motor control, and this has been the inspiration for adaptation of audio calibration in multiple acoustic environments; again, application of this approach in the area of auditory sense is completely new. The thesis showed that it was possible to calibrate the output of an SSL algorithm using the adaptive filter model of the cerebellum, improving the performance compared to the uncalibrated SSL. Using an adaptation of the MOSAIC framework, and specifically using responsibility estimation, a system was developed that was able to select an appropriate set of cerebellar calibration models and to combine their outputs in proportion to how well each was able to calibrate, to improve the SSL estimate in multiple acoustic contexts, including novel contexts. The thesis also developed a responsibility predictor, also part of the MOSAIC framework, and this improved the robustness of the system to abrupt changes in context which could otherwise have resulted in a large performance error. Responsibility prediction also improved robustness to missing ground truth, which could occur in challenging environments where sensory feedback of ground truth may become impaired, which has not been addressed in the MOSAIC literature, adding to the novelty of the thesis. The utility of the so-called cerebellar chip has been further demonstrated through the development of a responsibility predictor that is based on the adaptive filter model of the cerebellum, rather than the more conventional function fitting neural network used in the literature. Lastly, it was demonstrated that the multiple cerebellar calibration architecture is capable of limited self-organising from a de-novo state, with a predetermined number of models. It was also demonstrated that the responsibility predictor could learn against its model after self-organisation, and to a limited extent, during self-organisation. The thesis addresses an important question of how a robot could improve its ability to listen in multiple, challenging acoustic environments, and recommends future work to develop this ability.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.809251  DOI: Not available
Share: