@mastersthesis{Gal2017,

title = {EEG Signal Processing and Classification for Advanced Human-Machine Interfaces},

author = {Gorjup, Gal},

year  = {2017},

date = {2017-09-10},

school = {University of Ljubljana},

abstract = {EEG-based human-machine interfaces oer an alternative means of interaction with

the environment that relies solely on interpreting brain activity. They can not only

signicantly improve the life quality of people with neuromuscular disabilities, but also

present a wide range of opportunities for industrial and commercial adoption. This thesis

focuses on processing and classication of motor imagery EEG recordings. The used

data consisted of three data sets, two of which were recorded within this project. A software

framework that supports EEG signal ltering, feature extraction and classication

was developed and successfully used. Selected instances of FIR and IIR digital lters

were implemented and compared, showing that the latter was more appropriate for the

current application. Several feature extraction methods were implemented, including

band power, autoregressive modelling, Hjorth parameters and FFT-based features. An

LDA-based linear classication method was implemented and tests have shown that

it performs best with the band power features. Additionally, an LSTM-based neural

classication method was implemented and optimised in terms of architecture shape,

learning rate and weight decay parameters. Through optimisation, it was found that

this method also performs best with band power features. The implemented classiers

were compared based on the band power feature, using the available data sets recorded

with wet and dry electrodes, with monopolar and bipolar montage. The two methods

achieved similar performance in terms of prediction accuracy, although the linear classi

er was for the given data and training approach found to be favourable due to its

robustness and low complexity.},

note = {Co-supervision between SDU and University of Ljubljana},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{Adit2017,

title = {Modeling and Control of Dung Beetle-Like Robots},

author = {Kapilavai, Aditya},

year  = {2017},

date = {2017-06-01},

school = {University of Southern Denmark},

abstract = {African Dung beetles that belong to the species of Scarabaeus zambesianus possess a

distinctive feature in each of its leg structures. They can use their legs not only to walk

but also to manipulate objects. This thesis presents the design and development of real

dung beetle-like leg by conducting biological experiments on real insect front leg using

microCT scans. Using these data, a robot leg kinematic model is developed and kinematic

control is applied. Due to novel design with additional degrees of freedom controlling

orientation of the second joint, virtual and physical prototype kinematics is able to emulate

real beetle’s behavior and successfully mimics its degrees of freedom. Furthermore,

simulations were conducted to analyse workspace of the leg, the performance of the kinematic

control and the trajectory planning of the leg for locomotion and object manipulation

based on video recordings of the African dung beetle. Experiments were conducted on

the developed prototype to test the kinematic algorithm. Robotic design process based on

biological experiments and modeling corrected by feedback from biologists offered novel

methodology to improve the workspace of insect-like robots. This may increase both the

energy efficiency for bio-inspired robots and their potential for complex manipulation tasks,

such as demanding search and rescue missions in the wake of natural disasters.},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{Kristoffer2017,

title = {A Brain-Computer Interface with Neural Networks for Device Control},

author = {Honore, Kristoffer},

year  = {2017},

date = {2017-06-01},

school = {University of Southern Denmark},

abstract = {Brain-Control Interfaces (BCI) is a research subject, that has been around for a long time,

but it seams that a new surge of interest into this area is under way, which is understandable

as BCI has main useful applications. This thesis will use the BCI framework, to

investigate the hypothesis that Artificial Neural Networks (ANN) with memory capabilities,

performs better when classifying EEG signals of either right or left hand movement,

then ANNs that does not have memory capabilities. The success will be seen, by how well

the Radial Basis Function (RBF) network, performs compared to the Echo State Network

(ESN).},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{Malte2017,

title = {Implementation and Verification of an Adaptive Neural Control Oscillator for Robot Control on an FPGA},

author = {Carstensen, Malte },

year  = {2017},

date = {2017-06-01},

school = {University of Southern Denmark},

abstract = {This report describes the implementation of an existing and pre-trained neural network for

robotic legged locomotion control on a FPGA. The used neural networks follow a modular

approach consisting of relative simple elements who are capable to produce complex

control behavior when connected together. The system is currently based on a microcontroller

motivating this work to present the first step of investigation to convert the modules

into hardware based neural network. The designed elements were first tested using the

simulation environment from Xilinx Vivado Design Suite. After successful validation of the

simulated signals the modular hardware neural networks were converted and tested in the

analog domain using an oscilloscope. Thus, the validation and verification part consists

of two different methods. The developed neural units were packaged into an IP repository

that might be used for further neural FPGA implementations. All experiments are based

on the same methods that were already used to validate the functionality of the software

methods. All previous results could be successfully repeated by FPGA simulations and

the actual chip implementation.},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{Nikolaj2017,

title = {Evolving Robot Control for Object Transportation},

author = {Schaldemose Reibke, Nikolaj },

year  = {2017},

date = {2017-06-01},

school = {University of Southern Denmark},

abstract = {This Master Thesis explores the possibilities of phonotaxis controls using behavior based

controls. It focuses on the task of seeking towards a sound source while keeping another

sound source in a parallel alignment as a means for ground work distributed phonotaxis

individual agent formation tasks. The sound is processed in a lizard peripheral auditory

system which allows the agent to subtract valuable directional information from the sources.

The thesis explores the possibilities in a simulation environment which allows for precise

measures of different metrics as the distance between agents and sound sources where the

task explored has been the formative movement between 2 agents moving towards a goal

using behaviors built from the basic building blocks from the Braitenberg vehicle models

fed into the agent being a small robot with 2 wheels for moving around.},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{Aitor2017,

title = {Locomotion control for complex behaviour of bio-inspired multi-legged robotic systems},

author = {Aitor Miguel Blanco},

year  = {2017},

date = {2017-05-17},

school = {University of Southern Denmark},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{Setaro2017,

title = {Affordance Learning Applied on Bio-inspired Artificial Agents},

author = {Setaro, Michelangelo},

year  = {2017},

date = {2017-01-20},

school = {University of Southern Denmark},

abstract = {The field of robotics is in continuous evolution. handling objects affordances represents a further step towards the creation of intelligent agents. This Master Thesis 

focuses on implementing a representation of affordance making use of limited set of sensory-motor skills and Neural Networks. The design of a controller takes inspiration 

from biology and ecological psychology. A simulated Dung-Beetle hexapod robot is employed, it is capable of insect-like locomotion and dung beetle-like object 

manipulation. This thesis has resulted in the successful implementation of an affordance neural controller, that is eventually combined with the dung-beetle object transportation capabilities.},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{Dominik2016,

title = {Action-Sequence Learning in Mobile Robotics},

author = {Weickgenannt, Dominik Steven},

year  = {2016},

date = {2016-06-15},

school = {University of Southern Denmark},

abstract = {This master thesis explores a number of different deep and shallow Artificial Neural Network 

architectures to solve the problem of Action-Sequence-Learning on a mobile robot 

platform. The architectures are divided into two archetypes: direct Sensor-to-Behaviour 

mapping and a Trigger Network combined with a Neural Memory Module. The tested architectures include Feed-Forward, Recurrent, Echo-State and Long-Short-Term-Memory 

networks as well as Dynamic Neural Fields. The performances are compared against each 

other with and without noise. Findings indicate that a number of structures are capable 

of learning the predefined sequence in a robot simulation environment.},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{carlos2016,

title = {Multiple Time Scales of Learning for Adaptive Behavior of Mobile Robots},

author = {Moro García, Carlos},

year  = {2016},

date = {2016-06-15},

school = {University of Southern Denmark},

abstract = {Different time scales of learning can be found within the artificial intelligence algorithms. These time scales have been traditionally used separately except for some recent papers that started combining two of them. In fact, each of them has advantages and disadvantages that could be easily compensated by following the approach of using them at the same time as it 

is seen in nature, where different species evolve while organisms of each generation learn new behaviours. Therefore, this thesis tries to research the viability of using a short time scale of learning such as input correlation learning (ICO) with a long time scale of learning in the shape of a genetic algorithm (GA). These two systems together were tested with a mobile robot in a simulated environment relatively complex to navigate. The idea behind the simulation was to see how the evolutionary algorithms are able to obtain a specific morphology of the robot that is adapted to the environment while the correlation-based learning allows the system to navigate it. 

Experiments were done combining the algorithms in a sequential way, and 

a proof that this lead to an improvement was gotten. Moreover, there is 

evidence in the results obtained which suggests that a parallel combination 

of the algorithms would be able to perform even better.},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{Dinten2016,

title = {EEG Pattern Recognition},

author = {Dinten, Imara van},

year  = {2016},

date = {2016-06-15},

school = {University of Southern Denmark},

abstract = {During this master thesis, research regarding recurring patterns in EEG was done. Different 

methods on to how to detect these recurring patterns were compared. The recurring 

pattern chosen to detect in the EEG signal were 2D images that were shown to a 

test subject. It was then concluded that it is possible to see a change as to where in the 

moment the image was shown. After this, research was conducted to see which method 

could best automatically detect the difference in the EEG signal. Even make a distinction 

between the different images shown. The final method chosen for this application 

was an Artificial Neural Network.},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{Timon2016,

title = {Artificial Neural Control of a Prosthetic},

author = {Tomás, Timon},

year  = {2016},

date = {2016-06-15},

school = {University of Southern Denmark},

abstract = {Artificial limb prosthetics have come a long way from being crude, unsophisticated 

devices they once were. In the past couple of years, advanced actuation 

systems have been developed which now allow even relatively precise movements 

to be performed under laboratory conditions. Although these systems would require 

continuous and stable control signals to operate, in real-life scenarios these conditions 

are rarely met, which poses a significant engineering problem needing to 

be solved. This is especially true in the case of myoelectric prosthetic hands, since 

they possess many degrees of freedom, and the body produced signals — mainly 

do to their biological nature — are inherently unstable. One possible solution to this 

problem is to translate the incoming electromyographic (EMG) signals to stable control 

signals with the help of an intermediating artificial neural network (ANN), that 

has been trained to overcome the aforementioned errors present in the system. In 

this thesis work, I will investigate the potential usefulness of the application of ANNbased 

filtering and conditioning techniques for myoelectric signals in order to provide 

cleaner inputs to systems used to control assistive devices in the medical domain. 

Moreover, the thesis also aims to verify that ANN with memory capabilities 

are better suited for this given task than those without one. First a method will be 

developed to synthesize EMG signals to help in their analysis, then the capabilities 

of a radial basis function (RBF) network will be compared to that of an echo state 

network (ESN) using the synthetic data. In the next phase, real EMG signals will be 

measured and errors will be introduced to them. Finally, the contaminated signals 

will be cleaned, and the process tested through the control of a prosthetic hand. It 

will be shown, that neural networks are suitable to mitigate — and in certain cases 

eliminate — the artifacts present in these signals, and that the ESN — relying on its 

reservoir of interconnected neurons providing it with memory — can outperform the 

— memoryless — RBF network in multiple scenarios.},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{Martin2016,

title = {Classification of Terrain based on Proprioception and Tactile Sensing for Multi-legged Walking Robot},

author = {Bulin, Martin},

year  = {2016},

date = {2016-06-15},

school = {University of Southern Denmark},

abstract = {Proprioception and tactile sensing in insect-like legged robots is a fast, illumination insensitive and biologically inspired way of ground perception. In this thesis, 14 virtually generated terrains are classified based on the mentioned sensor types for a simulated version of hexapod robot AMOS II. A feedforward neural network framework equipped with a novel network pruning algorithm has been developed for classification. We observe over 

92% classification accuracy on deterministic terrain data and 72% on manually noised data. The pruning algorithm removes unimportant synapses (generally more than 90%) from a fully-connected network, while the classification accuracy does not drop significantly. The number of input neurons is reduced by 65%, resulting in the minimal network structure for the classification problem. A theory of using minimal structures for feature selection is proposed. The thesis outcome consists of a minimal neural network capable of terrain classification based on selected features of proprioceptive and tactile sensory signals.},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{Chris2016b,

title = {Object Manipulation in a Hexapod Robot},

author = {Lund Sørensen, Chris Tryk},

year  = {2016},

date = {2016-06-13},

school = {University of Southern Denmark},

abstract = {Insects, like dung beetles, can perform versatile motor behaviors including walking, climbing an object (i.e., dung ball), as well as manipulating and transporting it. To achieve such complex 

behaviors for artificial legged systems, this work presents modular neural controllers of a bio- 

inspired hexapod robot. The controllers utilize discrete-time neurodynamics and consists of 

seven modules based on three generic neural networks. One is a neural oscillator network 

serving as a central pattern generator (CPG) which generates basic rhythmic patterns. The 

other two networks are so-called velocity regulating and phase switching networks. They are 

used for regulating the rhythmic patterns and changing their phase. As a result, the modular 

neural controllers enable the hexapod robot to walk and climb a cylindrical object with a 

diameter of 18 cm (i.e.,  2.8 times the robot's body height) and a spherical object with a 

diameter of 30 cm (i.e.,  4.6 times the robot's body height). Additionally, they can also 

generate different hind leg movements for different object manipulation modes, like soft push, boxing-like motion and hard push. The manipulation behaviors makes it possible for the robot to transport the objects, and for the spherical object, across rough terrain. Combining these pushing modes, the robot can quickly transport the cylindrical object across an obstacle with a height up to 14 cm (i.e.,  2.2 times the robot's body height) and the spherical object up to 16 cm (i.e.,  2.2 times the robot's body height). The controllers was developed and evaluated using a physical simulation environment.},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{Torroba2016,

title = {Design, Simulation and Development of a Bipedal Locomotion Platform for Human-like Gaits Studies},

author = {Torroba Balmori, Ignacio; Rodriguez Marin, Jorge},

year  = {2016},

date = {2016-06-13},

school = {University of Southern Denmark},

abstract = {The modern field of robotics, in its attempt to understand and mimic the underlying principles 

of legged locomotion, has in the past decades given birth to some of the most advanced 

existing bipedal creatures, finding inspiration in nature and the human being. This thesis 

contributes the biped robot RuBi and its development environment to the advancement of 

this area of the engineering. RuBi is a human-inspired, compliance-reconfigurable and 

low-cost bipedal platform for research in human-like walking and running gaits control. 

The framework presented consists of the lower body frame of the robot and ROS enabled 

control architecture, a simulation model and environment and a test bench for 2D motion. 

The controller interface with the robot and the simulation environment in Gazebo 

has been successfully implemented tested. A set of experiments for the robot prototype is 

presented here although they have not been conducted due to time constraints and eventualities. Furthermore, everything has been set so that the utilization of the framework 

can be easily learned and mastered by new users.},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{Lasse_Brresen,

title = {Separation of Multiple Sound Sources Using Directional Stereo},

author = {Børresen, Lasse Simmelsgaard},

year  = {2015},

date = {2015-12-03},

address = {CBR Embodied AI and Neurorobotics Lab, The Mærsk Mc-Kinney Møller Institute},

school = {University of Southern Denmark},

abstract = {Humans primarily interpret sound as combinations of frequencies, this is the way useful 

information in sound is encoded, not in waveform. By converting sound to spectrograms, the change of frequency spectrum over timer can be represented as an image, 

eeffectively representing sound information over time. Using machine learning, otherwise 

unobtainable knowledge can be extracted from complex data. ML algorithms create 

models directly based on the data unlike expert systems. Performance of machine learning algorithms and Neural Networks applied to sound and sound direction are explored in this thesis. 

 

A data set of spectrograms with direction labels was recorded using two directional 

microphones placed at the centre of a semicircle of 37 speakers (5 deg interval). Both 

single sinusoids and complex speech is in the data set. 

 

Two experiments have been executed: (1) Estimating direction of a single source sound 

using a neural network. (2) Filtering a multi source sound by target direction. For both 

purposes, the input is the right and left spectrograms of the sound.For estimating source 

direction, a mean error of  2.01 deg was achieved. However, precision is direction 

dependent, with the best resolution around 0.0 deg and worst at the extremities. This 

is an attribute of the specific microphone setup. For filtering sounds by direction a 

convolutional neural network was trained on 10000 samples of mixed source sounds. A 

mean MSE of 0:093 on filtering standardized spectrograms was achieved, resulting in 

remarkable filtering of off-axis frequencies.},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{Deniel2015,

title = {Efficient Reconstruction of Neural Interconnections in Simulated Spiking Neurons},

author = {Horvatic, Deniel},

year  = {2015},

date = {2015-06-10},

school = {University of Southern Denmark},

abstract = {Insects, like dung beetles, can perform versatile motor behaviors including walking, climbing an object (i.e., dung ball), as well as manipulating and transporting it. To achieve such complex 

behaviors for articial legged systems, this work presents modular neural controllers of a bio- 

inspired hexapod robot. The controllers utilize discrete-time neurodynamics and consists of 

seven modules based on three generic neural networks. One is a neural oscillator network 

serving as a central pattern generator (CPG) which generates basic rhythmic patterns. The 

other two networks are so-called velocity regulating and phase switching networks. They are 

used for regulating the rhythmic patterns and changing their phase. As a result, the modular 

neural controllers enable the hexapod robot to walk and climb a cylindrical object with a 

diameter of 18 cm (i.e.,  2.8 times the robot's body height) and a spherical object with a 

diameter of 30 cm (i.e.,  4.6 times the robot's body height). Additionally, they can also 

generate dierent hind leg movements for dierent object manipulation modes, like soft push, 

boxing-like motion and hard push. The manipulation behaviors makes it possible for the robot to transport the objects, and for the spherical object, across rough terrain. Combining these pushing modes, the robot can quickly transport the cylindrical object across an obstacle with a height up to 14 cm (i.e.,  2.2 times the robot's body height) and the spherical object up to 16 cm (i.e.,  2.2 times the robot's body height). The controllers was developed and evaluated using a physical simulation environment.},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{SwarmKrauch,

title = {Exploration and Object Retrieval with Robot Swarms},

author = {Hans-Joachim Krauch},

year  = {2015},

date = {2015-06-01},

address = {CBR Embodied AI and Neurorobotics Lab, The Mærsk Mc-Kinney Møller Institute},

school = {University of Southern Denmark},

abstract = {Swarm robotics is a new approach to the coordination of large robot populations. This 

thesis focuses on coordination strategies for exploration and foraging tasks. Dierent 

agent strategies are implemented and tested in a simulation environment. Key parts of 

the implemented design include the modular agent design and the probabilistic way in 

which agents decide where to go next. 

Each agent strategy is simulated on exploration and foraging tasks and rated based 

on its performance, scalability, robustness against agent failure and computational complexity. 

The experiments show that the most sophisticated strategies have the best performance 

for small swarms. For larger swarms however, the strategies that perform the 

best, are the most simple ones.},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{Patrick_Giuliano,

title = {Adaptive Locomotion Control of Embodied Legged Systems},

author = {Giuliano Di Canio and Patrick Stolc},

year  = {2015},

date = {2015-06-01},

address = {CBR Embodied AI and Neurorobotics Lab, The Mærsk Mc-Kinney Møller Institute},

school = {University of Southern Denmark},

abstract = {The eld of robotics is evolving very fast. In the eld of biped robotics research is 

focused on making robots walk and adapt to the environment. This master thesis 

focuses on locomotion and balance control of the DACBot, a small planar robot. Both 

locomotion and balance control of the DACBot is inspired by biology. By using biology 

as an inspiration for the development of the controllers, this thesis has shown that these 

principles can be used for controlling a biped robot. This thesis has resulted in the 

successful implementation of an adaptive locomotion controller for the DACBot and a 

two stage balance controller.},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{Stoyan,

title = {Brain - Computer Interface for Robot Control},

author = {Stoyanov, Stoyan},

year  = {2015},

date = {2015-06-01},

address = {CBR Embodied AI and Neurorobotics Lab, The Mærsk Mc-Kinney Møller Institute},

school = {University of Southern Denmark},

abstract = {This thesis presents a framework for brain - computer interface for robot control. The 

list of functionalities of the framework includes data acquisition, data processing, feature 

extraction and classication. The list is further extended with options for robot control 

and simulation due to the framework integration with the lpzrobots and gorobots software. 

The framework also can save, visualise and stream the acquired data because of 

its integration with ROS. As a proof of concept several experiments were performed with 

a single channel EEG device and a real hexapod robot. The experiments showed that 

the framework is able to demonstrate real robot control based on real EEG recordings 

as well as recording of neuron synchronisation and de synchronisation over the motor 

cortex area.},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}

@mastersthesis{Ditlev,

title = {Visual Segmentation of Potatoes in Cluttered Environments},

author = {Ditlev Andersen and Anders Prier Lindvig},

year  = {2015},

date = {2015-06-01},

address = {The Mærsk Mc-Kinney Møller Institute},

school = {University of Southern Denmark},

keywords = {},

pubstate = {published},

tppubtype = {mastersthesis}

}