In recent years, learning and evolutionary methods have been proposed as methods for automatically designing robot controllers without the need for detailed human design effort. Unfortunately, the reality has been that these methods have only been successfully applied to relatively simple problems involving low-bandwidth sensors and actuators, and simple (often purely reactive) behaviours. Purely automated design methods seem unable to `scale up' to design controllers for the realistically complex tasks we wish to tackle. A promising compromise solution is the idea that the learning/evolutionary system can be left to do most of the work, but with a human providing some sort of high-level assistance to make the problem tractable. Designing robot controllers in this way is often called `robot shaping'. In this thesis I explore a number of dierent forms of shaping, focusing in particular on `black box' techniques which I suggest are more likely to scale up to complex problems than other shaping methods. I also propose a novel extension of Genetic Programming, for use with these shaping methods. Experiments are described in which controllers were evolved, both with and without shaping, for a range of complex tasks including getting a mobile camera to track a moving light in two dimensions, and the harder problem of visually tracking arbitrary moving objects. These controllers are evolved rst in simulation, and then the best ones, evolved using shaping, are transferred successfully to a real robot. I conclude that if used carefully, shaping can reduce learning time and improve nal controller performance. However, choosing an appropriate form of shaping still requires the designer to be very much aware of the underlying details of the evolutionary system.

This thesis proposes an efficient path planning technique for a non-contact optical “painting” system that produces surface images by moving a robot mounted laser across objects covered in photographic emulsion. In comparison to traditional 3D planning approaches (e.g. laminar slicing) the proposed algorithm dramatically reduces the overall path length by optimizing (i.e. minimizing) the amounts of movement between robot configurations required to position and orientate the laser. To do this the pixels of the image (i.e. points on the surface of the object) are sequenced using configuration space rather than Cartesian space. This technique extracts data from a CAD model and then calculates the configuration that the five degrees of freedom system needs to assume to expose individual pixels on the surface. The system then uses a closest point analysis on all the major joints to sequence the points and create an efficient path plan for the component. The implementation and testing of the algorithm demonstrates that sequencing points using a configuration based method tends to produce significantly shorter paths than other approaches to the sequencing problem. The path planner was tested with components ranging from simple to complex and the paths generated demonstrated both the versatility and feasibility of the approach.

No description available.

No description available.

No description available.

This work aims to improve corporate functional departments' confidence in adopting modern control approaches in new scenarios and thus presents control structure solutions based on model predictive control (MPC) for two control problems facing existing upstream oil and gas production plants. These are the disturbance growth in the series connected process and the control system dependency on operators. The suggested control solution integrates MPC as a master controller for the existing classical control of each subsystem, with a focus on those with high interaction phenomena. The proposed approach simply and inexpensively encompass MPC features such as predictions, optimizations, coordination and constraint handling as well as PID features like simplicity and ease of troubleshoot. In addition, the proposed control concept utilises the process safeguarding information and enhances the plant-wide optimal performance. The suggested control solution supports the role of control room operators, which is shown to reduce the growth in the impact of process disturbances. Compared with some alternative control structures (centralised MPC, decentralised MPC, distributed MPC (DMPC), and hierarchical DMPC) this proposal is simple, inexpensive to implement, and critically, builds on the local team operational experience and maintenance skills. Three process models were developed that representing the common gas treatment processes in upstream oil and gas plants, gas sweetening, gas dehydration and hydrocarbon dewopointing. The models were utilised to examine different control structures and proposals. These models are not only of benefit to studies on upstream oil and gas processes, but also to Large Scale Systems (LSS) in general. The models were used to analyse the disturbance impacts on a series connected processes, therefore to provide answers about how process malfunctions and different disturbances affect the processing operations. The proposed control system is designed on a cascade strategy and thus provides a flexible system control almost like a decentralised structure in dealing with disturbances and unit failures, and at the same time improves the closed loop performance and the plant-wide optimal operation. The control system contain MPC's that are designed to regulate the critical loops only while the rest of the uncritical loops will continue to function in a decentralised fashion under PID control algorithm. This minimises any design and set up costs, reduces demand on the communication network and simplifies any associated real time optimisations. The improved local control reduced the need for control room operator interactions with their associated weaknesses. The proposed control structure communicate with the process safeguarding system to enable prompt response to disturbances caused by unit failures, and shares critical information with adjacent MPC's, which indeed works as a feed-forward, to reduce the impact of process disturbances and enhance optimality. The control system design is simple, inexpensive to implement and significantly reduces the frequency of plant shut downs and saves on operating costs by properly controlling the disturbance growth in the process.

The research work presented on this thesis provides an alternative tool for characterising oil fields under fluid injection by analysing historical production/injection rates. In particular polynomial and radial basis Non Linear Autoregressive with Exogenous Input Model (NARX) models were developed; these models were capable of capturing the dynamics of an operating field in the North Sea. A Greedy Randomised Adaptive Search Procedure (GRASP) heuristic optimisation method was applied for estimating a future injection strategy. This approach is combined with a risk analysis methodology, a popular approach in financial mathematics. As a result, it is possible to estimate how likely it is to reach a production goal. According to the simulations, it is possible to increase oil production by 10% in one year by implementing a smart injection strategy with low statistical uncertainty. Resulting from this research project, a computational tool was developed. It is now possible to estimate NARX models from any field under fluid injection as well as finding the best future injection scenario.

The primary purpose of this thesis is to present optimisation algorithms which facilitate the selection of weights in robust control techniques. The robust control paradigms investigated in this thesis are μ-synthesis and H_∞ loop-shaping, as these techniques offer a systematic framework for synthesising sensible controllers that meet performance objectives and guarantee robustness to model uncertainty and unmeasured disturbances. This dissertation is essentially composed of two parts. In the first part of the thesis, a conceptually new approach to the μ-synthesis robust performance problem is presented, whereby an optimisation problem is proposed whichmaximises the performance weights in the frequency ranges of interest subject to the existence of an internally stabilising controller that guarantees robust performance with respect to these maximised weights. Thus, performance weights and a robustly stabilising controller are simultaneously synthesised by one algorithm in a systematic way. Two solution algorithms are given for the posed optimisation problem - one being pointwise in frequency and the other using state-space techniques. The latter solution eliminates all of the disadvantages of the pointwise approach and considerably enhances the benefits of this type of optimisation based weight selection. The resulting conceptually novel method for performing μ-synthesis robust performance based design is a valuable alternative to the standard D-K iterative procedure. In the second part of the thesis, several steps of the standard H1 loop-shaping design procedure are combined into one optimisation problem that maximises the robust stability margin over the loop-shaping weights subject to constraints which ensure that the loop-shape and the singular values/ condition numbers of the weights lie in pre-specified regions. In this framework, loop-shaping weights, which can be required to have either a diagonal or a non-diagonal structure, and a robustly stabilising controller are simultaneously synthesised by one algorithm systematically. Correspondingly, the proposed algorithms greatly simplify the design of 'good' performance weights and loop-shaping weights, and hence allow the designer to concentrate on more fundamental design issues. These algorithms also give an indication of what performance is achievable, although further research is required in that direction.

So-called CF-FM bats are highly mobile creatures who emit long calls in which much of the energy is concentrated in a single frequency. These bats face sensor interpretation problems very similar to those of mobile robots provided with ultrasonic sensors, while navigating in cluttered environments. This dissertation presents biologically inspired engineering on the use of narrowband Sonar in mobile robotics. It replicates, using robotics as a modelling medium, how CF-FM bats process and use the constant frequency part of their emitted call for several tasks, aiming to improve the design and use of narrowband ultrasonic sensors for mobile robot navigation. The experimental platform for the work is RoBat, the biomimetic sonarhead designed by Peremans and Hallam, mounted on a commercial mobile platform as part of the work reported in this dissertation. System integration, including signal processing capabilities inspired by the bat’s auditory system and closed loop control of both sonarhead and mobile base movements, was designed and implemented. The result is a versatile tool for studying the relationship between environmental features, their acoustic correlates and the cues computable from them, in the context of both static, and dynamic real-time closed loop, behaviour. Two models of the signal processing performed by the bat’s cochlea were implemented, based on sets of bandpass filters followed by full-wave rectification and low-pass filtering. One filterbank uses Butterworth filters whose centre frequencies vary linearly across the set. The alternative filterbank uses gammatone filters, with centre frequencies varying non-linearly across the set. Two methods of estimating Doppler-shift from the return echoes after cochlear signal processing were implemented. The first was a simple energy-weighted average of filter centre frequencies. The second was a novel neural network-based technique. Each method was tested with each of the cochlear models, and evaluated in the context of several dynamic tasks in which RoBat was moved at different velocities towards stationary echo sources such as walls and posts. Overall, the performance of the linear filterbank was more consistent than the gammatone. The same applies to the ANN, with consistently better noise performance than the weighted average. The effect of multiple reflectors contained in a single echo was also analysed in terms of error in Doppler-shift estimation assuming a single wider reflector. Inspired by the Doppler-shift compensation and obstacle avoidance behaviours found in CF-FM bats, a Doppler-based controller suitable for collision detection and convoy navigation in robots was devised and implemented in RoBat. The performance of the controller is satisfactory despite low Doppler-shift resolution caused by lower velocity of the robot when compared to real bats. Barshan’s and Kuc’s 2D object localisation method was implemented and adapted to the geometry of RoBat’s sonarhead. Different TOF estimation methods were tested, the parabola fitting being the most accurate. Arc scanning, the ear movement technique to recover elevation cues proposed by Walker, and tested in simulation by her, Peremans and Hallam, was here implemented on RoBat, and integrated with Barshan’s and Kuc’s method in a preliminary narrowband 3D tracker. Finally, joint work with Kim, K¨ampchen and Hallam on designing optimal reflector surfaces inspired by the CF-FM bat’s large pinnae is presented. Genetic algorithms are used for improving the current echolocating capabilities of the sonarhead for both arc scanning and IID behaviours. Multiple reflectors around the transducer using a simple ray light-like model of sound propagation are evolved. Results show phase cancellation problems and the need of a more complete model of wave propagation. Inspired by a physical model of sound diffraction and reflections in the human concha a new model is devised and used to evolve pinnae surfaces made of finite elements. Some interesting paraboloid shapes are obtained, improving performance significantly with respect to the bare transducer.

A major mission driver for space exploration is to minimise groundbased human intervention and hence associated operations costs, thereby maximising science data return. Future robotic exploration such as the ESA ExoMars mission will require rovers to be equipped with greater autonomy. In line with such a requirement, a new autonomous system named Autonomous Rock Science Analysis System (ARSAS) is proposed in this thesis for the purpose of identifying and accessing scienti c rocks during exploration. ARSAS consists of three components: rock detection, rock science value evaluation and related executive agent. Three approaches are presented in the rock detection component. A number of image processing and machine learning techniques have been employed, including multispectral sampling, fuzzy-rough feature selection, classi cation, clustering, thresholding and saliency methods. The rock science evaluation component is primed by a human planetary geology expert. Some visual features are selected as the indicators of some geological attributes and then a fuzzy expert system is used to convert the rock attributes to corresponding science value. In contrast with previous works, the proposed science evaluation mechanism is more autonomous and geology-oriented. The executive agent mainly consists of a pair of cameras and a robotic arm, together with a series of algorithms for coordinate transformation. It serves as a platform to support the previous two components. Experiments have been conducted on this platform to demonstrate the usefulness, stability and repeatability of the proposed system. The details of design, implementation and experimentation of all components are elaborated in the thesis.