James the PokerBot is a robot created to play a game of Texas Hold’em against human players using a normal deck of cards. The idea of making James came from a project in a computer vision course. The goal for this project was creating a computer vision program that would identify the playing card in a picture of a single card. As the project proved very interesting, it was decided to continue the work as a separate project after the computer vision course ended. The resulting program would be able to identify several playing cards from a picture, despite there being other non-card objects in this image. More about this previous project can be read in the report that was made for it, fall 2011 [1] that is included in the digital attachment.This report however, contains a description of the steps needed to go from having the computer vision program to make the beginning of a card playing robot. The first part of the report deals with the changes that needed to be done to the old computer vision program. These changes includes getting input from a live source rather than using static images, as well as speed optimization needed when running the program on a less powerful BeagleBoard. A video showing the raw output of the final program is included in the digital attachment under Result videos, Computer vision as well as being available online [2]. Next the report continues by describing the different modules used to play the game. First the main module is explained. This corresponds to the actual robot and is the module that runs the computer vision program as well as controls the progression of the game itself. The programs made to run on the main module were developed on a stationary computer. For the prototype made in this project, a development board called a BeagleBoard would work as the main module. The other modules made, were the player modules. These were small controllers used by the human players to interface with James when playing the game. Three different player modules have been made. First a simulator was made to use when developing the programs for the main module. Then two different hardware prototypes, both based on AVR microcontrollers, were made for the final setup.The report then describes the way communication was setup between the modules. For the simulator player modules, this was achieved using virtual COM ports. The hardware prototypes on the other hand, used XBee modules communicating wirelessly over ZigBee.Following this is a chapter describing how the actual Texas Hold’em application was designed. This application controls the game structure and handles input from the human players regarding the game. The application will also interpret the raw input from the computer vision program, regarding the current state of the game.Finally there is a chapter describing the robotic arm planned to use for picking up playing cards. Sadly there was never enough time to complete this arm, but ideas on how this would be done are discussed. There are also added pictures of an unfinished prototype that was made.Two videos demonstrating the final results both for the Texas Hold’em application running on the stationary computer, as well as for the hardware prototypes, are added in the digital attachment under Result videos as well as being available online [3] [4]. The first video demonstrates how the computer vision program can be used for a machine to effectively play a game of cards against human players. The second video demonstrates how these programs also can be used with relatively inexpensive hardware, making it possible to commercialize the product.

This report is written as a master’s thesis given at NTNU, and in collaboration with Kongsberg Defence System. This text investigates the use of L1 adaptive control for multi-variable systems with unknown nonlinear unmatched uncertainties and unmodeled actuator dynamics, particularly for the inner longitudinal and lateral loops of a nonlinear F-16 aircraft model. Chapter 1 gives an introduction to the report by introducing the basic principles of aircraft control and specifying the goals of the text. Chapter 2 introduces the theory of linear aircraft modelling, which specifies the starting point for the L1 adaptive control design. Chapter 3 presents the L1 adaptive control theory considered in this text, together with the main theoretical results. Chapter 4 presents the simulations results and the discussion, together with the specific L1 adaptive control design used in the simulations. Chapter 5 concludes this report and points toward possible future work.The L1 adaptive control theory has proven to give good results. This was also found in this report. The controller handles unmatched nonlinearities and disturbances very well and manages to track the reference signal. For the longitudinal controller, this has also been shown through simulations on a realistic nonlinear F-16 model. Compared to the MRACformulation we see that the introduction of the filter in the L1 adaptive control formulation gives great improvements with regards to performance. The results of this text shows how the L1 adaptive controller manages to separate adaptation from control and thus be able to introduce fast adaptation without introducing high gain feedback. This text also discusses how implementation issues like limited sampling rate affects the performance of the L1 adaptive controller, and shows how this can be handled by a proper redesign of the architecture. The results of the L1 adaptive controller for the longitudinal mode are compared to a simple PID-controller. We actually see that the PID-controller performs almost as good as the L1 adaptive controller for the simple longitudinal system.

Achieving robotic locomotion is in general a difficult task. When the system of concern is underactuated, i.e. it has more degrees of freedom than the number of control inputs available, dynamic constraints are imposed, further complicating the task. This is the case for the brachiation motion observed in the lesser apes, i.e. gibbons and siamangs, as the gait involves periods of time at which the ape is suspended by one arm with limited torques available to influence the rotation about the handhold. Earlier work has been concerned with modeling of a 24-degrees-of-freedom monkey robot and the design of a brachiation gait. In this thesis we develop a toolbox to facilitate the design of a controller based on transverse linearization for this brachiation gait. The main focus is to stabilize the single-support part of the gait, i.e. the part that is subjected to dynamic constraints due to the lack of torque about the handhold, as traditional control theory is unable to stabilize the desired motion in this case. The developed toolbox is used in designing a controller that orbitally stabilizes an inverted pendulum system. As an initial step in achieving orbital stabilization of the brachiating gait, asymptotic convergence to the virtual holonomic constraints is demonstrated for a simplified model of the 24 degrees-of-freedom monkey robot.

Radial Basis Function Neural Networks are well suited for learning the systemdynamics of a robot manipulator and implementation of these networks in thecontrol scheme for a manipulator is a good way to deal with the system uncertaintiesand modeling errors which often occur. The problem with RBF networkshowever is to nd a network with suitable size, not too computational demandingand able to give accurate approximations. In general two methods for creating anappropriate RBF network has been developed, 1) Growing and 2) Pruning.In this report two dierent pruning methods which are suitable for use in alearning controller for robot manipulators are proposed, Weight Magnitude Prun-ing and Neuron Output Pruning. Weight Magnitude Pruning is based on a pruningscheme in [8] while Neuron Output Pruning is based on [2]. Both pruning methodsare simple, have low computational cost and are able to remove several unitsin one pruning period. The thresholds used to nd which neurons to remove arespecied as a percent and hence less problem dependent to nd.Simulations with the two proposed pruning methods in a learning inverse kinematiccontroller for tracking a trajectory by using the three rst joints of the ABBIRB140 manipulator are conducted. The result was that implementing prunedRBF networks in the controller made it more robust towards system uncertaintiesdue to increased generalization ability. These pruned networks were found togive better tracking in the case of unmodeled dynamics compared to the incorrectsystem model, not pruning the RBFNNs and a type of growing network calledRANEKFs. Computational costs were also reduced when the pruning schemeswere implemented.NTNU has a manipulator of the type ABB IRB140 and the learning inversekinematic controller with pruning of RBF networks should be implemented andtested on this in real-life simulations.

Institutt for Teknisk Kybernetikk(ITK) ved NTNU har et programvarebibliotek for styring av proteser. Dette biblioteket består av et mønstergjenkjenningssystem som kan brukes til å styre en robothånd. Mønstergjenkjenning gir protesebrukeren mulighet til å styre en protese, men uten den grad av kontroll som er ønsket. Derfor er det i denne oppgaven blitt implementert et proporsjonalt styringssystem som skal gjøre at brukeren kan kontrollere en protese bedre. For å trene opp et slikt system er det utviklet metoder som skal brukes sammen med proteseguidet trening. Det er også utviklet treningssett som brukes for å lagre referanseverdier. For å vise at dette kan brukes på proteser, ble kontroll av en protese fra Motion Control implementert inn i ITKs bibliotek. Resultatene indikerer at dette fungerer bra, men at det å kontrollere flere frihetsgrader simultant er vanskelig. Derfor må det utvikles nye treningssett som kan forbedre kontroll av flere frihetsgrader simultant.

In this thesis, an experimental runtime system for utilizingthe computational power of a multi-computer environment is presented.Through simple benchmark tests it is shown how some taskswill have a considerate speed-up compared to running on a single computer.An outline for designing languages and compilers suited for the runtimeis also explored and discussed, and it is shown how the system,with some extensions, would be well suited for utilizing the sparecomputational power in a multi-computer environment.This also holds, with some extra considerations, for a real-time application.

This thesis considers two main issues: the theoretical aspects of stationkeeping capability analysis in general, and dynamic capability (DynCap) analysis in particular; further there has been conducted capability analyses of the model vessel, CyberShip III, including an experimental validation of the results.This thesis presents the most comprehensive investigation into the differences between the DynCap analysis and the industry standard DP capability analysis (DPCap), published to date, as far as both Marine Cybernetics AS and the author are aware of. The mathematical foundation of DynCap has been investigated in detail, and it has been shown that by applying the simplications and assumptions proposed in the industry used IMCA M140 specications for DP capability plots to the equations of DynCap, it is possible to arrive at the DPCap equations. The review of the mathematical foundation of capability analysis, highlights the benefits of employing the DynCap analysis in favor of the DPCap analysis, and provides the necessary mathematical motivation for choosing the DynCap analysis in favor of the DPCap analysis.This thesis also presents multiple stationkeeping capability analyses of CyberShip III. The stationkeeping capability of CyberShip III was obtained through both DPCap analysis and DynCap analysis. In addition, an experiment using the model vessel CyberShip III was conducted, in order to obtain the true stationkeeping capability of CyberShip III. The results obtained revealed that the traditional DPCap analysis, seems repeatedly to be too optimistic, and that the DynCap results seems to be more realistic. The experimental results show that the relative reduction in depicted capability between the result obtained through DynCap analysis and experimental data obtained, is approximately 10%. In contrast, the relative reduction from DPCap results, compared to the experimental data, where found to be approximately 60%.By combining the theoretical foundation of DynCap, the simulations conducted and the experimental data obtained, the thesis is able to validate the DynCap stationkeeping capability analysis. The superior performance of the DynCap analysis, relative the DPCap analysis, provides additional arguments for employing the DynCap method in favor of the industry standard DPCap. Further the thesis highlights possible shortcomings of the DynCap analysis, including remarks in regard to the complexity and accuracy of the models used during the DynCap simulations. The implications of the results presented in this thesis may be regarded as considerable. The results show that without a well defined standard for DPCap analysis, the validity of the results are uncertain. The uncertainty in turn limits the use of the obtained results. This thesis proposes that by employing the DynCap analysis, the operators, ship owners and oil companies, will be able to make decisions based on more accurate data, which will enhance security, and reduce non-productive time. With well defined limits of operation for the vessels, it will be possible to utilize theresources in a more productive manner. By basing decisions on accurate data, the selected vessel may be able to operate closer to the limit, without compromising security. This will in turn reduce costs and possibly increase profits.Keywords: Dynamic Positioning, DP, DP Capability, DPCap, Dynamic Stationkeeping Capability, DynCap

The velocity observer in Wondergem et al. [2011] ("the Wondergem Observer") and the velocity observer in Stamnes et al. [2011] ("the Stamnes Observer") are evaluated and compared for a ship model given in Wondergem et al. [2011]. The former observer has shown to be uniformly semi-globally exponentially stable. The latter observer has proven to be uniformly globally asymptotically stable and semi-globally exponentially stable when certain requirements are satisfied. However, due to a non-linear damping term in the ship model, the Stamnes Observer only meets the conditions for uniformly semi-globally exponentially stability.The two observers have shown different convergence properties. The Wondergem Observer has a large overshoot during the settling period, then follows the reference robustly. The Stamnes Observer has a first order system behaviour and can become underdamped in the position convergence. Both observers show similar behaviour when considering a bias in the system equations. However, the run time for the Stamnes observer when simulating on a desktop computer is one to two orders of magnitude longer than for the Wondergem Observer.From the evaluations in this thesis, the Wondergem Observer has shown to be the best choice for the given ship model. It is more robust than the Stamnes Observer, it does not have a lower degree of stability as first implied, and it has a much simpler structure.

Spectroscopy is a powerful tool for analysis in a widerange of scientific and industrial fields. Spectroscopic measurements obtaininformation about the interaction between matter and radiated energy.This information may be utilized for detection and classification of objectsand substances. There is a wealth of different methods and technologiesdeveloped for different applications, from space- and air-borne remotesensing to medical sample analysis.In this report we present a spectral imaging system consisting of a digitalcamera with an acousto-optical tunable filter (AOTF) as the spectralband selective element. An evaluation of the properties of the imager ispresented, and we investigate the feasibility of real-time spectral imagingvideo based on this system. A comparison with other spectral imagingtechnologies is made. Custom application software for camera control andimage acquisition is presented. A special optical layout that minimizes theoptical aberrations produced by the diffraction process in the AOTF is alsopresented.The system produce a stack of monochrome images at different wavelength bandsforming an "image cube" containing a combinationof spatial an spectral information. The system uses the visible part ofthe electromagnetic spectrum to collect spectral information at every pointin an image plane. Our results show that the camera have good lightsensitivity and a high signal-to-noise ratio, but that the image throughputand spatial resolution is not optimal for a high image rate. The AOTFis a highly flexible device with random access to wavelength bands andselectable band-pass shape. But the many configurable parameters of thesystem also complicates usage. Many parameters need to be calibrated,and we present some methods on this. A major challenge in an spectralimaging system is to get enough light through to the sensor, and we findthat there is high loss of light caused by the diffraction process in the filter.It will be shown that the system is not suitable for real-time spectral videoimaging because of the low frame rate. Suggestions for improvements aregiven.

In this thesis, the development of a web application for designing electronic circuits has been initiated and documented.The application will feature some unique features regarding the design process of electronic circuits.Among them are interface based routing, a plugin-friendly environment and a collaborative resource database.At the start of working on this thesis, there were no known web-based EDA software available.This provided an unique opportunity to fill this gap.The application has been implemented using HTML5 and JavaScript for the interactive front-end (The web browser),and Google Go and MongoDB for the backend (The Server).The basic building blocks of this application has been implemented, and together serves as an tech demo, available under a GPL licence.