Spiking neural networks for robot locomotion control

Wiklendt, Lukasz

Title: Spiking neural networks for robot locomotion control
Creator: Wiklendt, Lukasz
Relation: University of Newcastle Research Higher Degree Thesis
Resource Type: thesis
Date: 2014
Description: Research Doctorate - Doctor of Philosophy (PhD)
Description: Spiking neural networks (SNNs) are computational models of biological neurons and the synapses that connect them. They are chosen for their characteristic property of information exchange via the timing of events called spikes, in contrast to earlier-developed models such as sigmoid neural networks which have no explicit timing component. SNNs are often applied to tasks in artificial intelligence by using existing models of biological neural networks that were used in neuroscience in the past, and that are detailed enough to contain the timing-property. Neurons can be modelled at many levels of detail, and often a neuron model is chosen with scant consideration of the most appropriate level of detail for the given task. This thesis presents a novel spiking neuron model developed to retain the timing-property, including proposed favourable characteristics for application to artificial intelligence tasks, while removing the unnecessary detail for achieving those characteristics that current SNN models contain. The result is a computationally powerful neuron model with an analytically solvable spiking-time calculation. While SNNs have been applied to various tasks in artificial intelligence, including robot control, the types of control problems faced have been primarily of a stable nature. This thesis focuses on unstable control problems, that is, problems where the dynamics governing the motion of the robot under control are such that small disturbances, inaccuracies, or pauses in control can lead to a rapid acceleration away from a desired state. Concretely, simulation experiments are conducted (i) on a planar underactuated inverted double-pendulum called the Acrobot for the swing-up and balance task which, combined with linear quadratic regulation (LQR) control for balance, was able to achieve the task, and (ii) to a 1.5m tall biped for the distance locomotion task, where it walked 16m without collapsing. In the interests of automatically developing bipedal dynamic walking behaviour, via the stochastic tuning of spiking neural network parameters, a new spherical-foot model is presented that exhibits favourable dynamical properties. Existing physical biped robot morphologies can be clustered into three main groups based on their feet and ankle configurations. One group contains large flat feet with actuated ankles, and is most often seen in environments and tasks requiring moving in both sagittal (forward-backward) and coronal (left-right) planes, such as robotic soccer. The second group contains point feet with no ankles, and finds success in fast locomotion such as running, where coronal motion is limited. The third group consists of passive-dynamic walkers, that contain rounded feet and are able to walk in the sagittal plane along a slight decline without any control input. In this thesis a new biped feet-angle configuration is proposed which is a marriage of these groups, with relatively small (second group) rounded feet capable of smooth continuous ground contact (third group), and actuated ankles (first group) that aid in standing balance control. An analysis of this novel type of foot configuration is presented here for the planar case, and a controller for standing balance is included.
Subject: spiking neural networks; robot locomotion; evolutionary computation; acrobot; biped
Identifier: http://hdl.handle.net/1959.13/1049196
Identifier: uon:15006
Language: eng
Full Text

Hits: 1877
Visitors: 2032
Downloads: 584

		Thumbnail	File	Description	Size	Format
View Details Download			ATTACHMENT01	Abstract	211 KB	Adobe Acrobat PDF	View Details Download
View Details Download			ATTACHMENT02	Thesis	3 MB	Adobe Acrobat PDF	View Details Download