Object recognition and pose estimation is a fundamental problem in computer vision and of utmost importance in robotic applications. Object recognition refers to the problem of recognizing certain object instances, or categorizing objects into specific classes. Pose estimation deals with estimating the exact position of the object in 3D space, usually expressed in Euler angles. There are generally two types of objects that require special care when designing solutions to the aforementioned problems: rigid and deformable. Dealing with deformable objects has been a much harder problem, and usually solutions that apply to rigid objects, fail when used for deformable objects due to the inherent assumptions made during the design. In this thesis we deal with object categorization, instance recognition and pose estimation of both rigid and deformable objects. In particular, we are interested in a special type of deformable objects, clothes. We tackle the problem of autonomously recognizing and unfolding articles of clothing using a dual manipulator. This problem consists of grasping an article from a random point, recognizing it and then bringing it into an unfolded state by a dual arm robot. We propose a data-driven method for clothes recognition from depth images using Random Decision Forests. We also propose a method for unfolding an article of clothing after estimating and grasping two key-points, using Hough Forests. Both methods are implemented into a POMDP framework allowing the robot to interact optimally with the garments, taking into account uncertainty in the recognition and point estimation process. This active recognition and unfolding makes our system very robust to noisy observations. Our methods were tested on regular-sized clothes using a dual-arm manipulator. Our systems perform better in both accuracy and speed compared to state-of-the-art approaches. In order to take advantage of the robotic manipulator and increase the accuracy of our system, we developed a novel approach to address generic active vision problems, called Active Random Forests. While state of the art focuses on best viewing parameters selection based on single view classifiers, we propose a multi-view classifier where the decision mechanism of optimally changing viewing parameters is inherent to the classification process. This has many advantages: a) the classifier exploits the entire set of captured images and does not simply aggregate probabilistically per view hypotheses; b) actions are based on learnt disambiguating features from all views and are optimally selected using the powerful voting scheme of Random Forests and c) the classifier can take into account the costs of actions. The proposed framework was applied to the same task of autonomously unfolding clothes by a robot, addressing the problem of best viewpoint selection in classification, grasp point and pose estimation of garments. We show great performance improvement compared to state of the art methods and our previous POMDP formulation. Moving from deformable to rigid objects while keeping our interest to domestic robotic applications, we focus on object instance recognition and 3D pose estimation of household objects. We are particularly interested in realistic scenes that are very crowded and objects can be perceived under severe occlusions. Single shot-based 6D pose estimators with manually designed features are still unable to tackle such difficult scenarios for a variety of objects, motivating the research towards unsupervised feature learning and next-best-view estimation. We present a complete framework for both single shot-based 6D object pose estimation and next-best-view prediction based on Hough Forests, the state of the art object pose estimator that performs classification and regression jointly. Rather than using manually designed features we propose an unsupervised feature learnt from depth-invariant patches using a Sparse Autoencoder. Furthermore, taking advantage of the clustering performed in the leaf nodes of Hough Forests, we learn to estimate the reduction of uncertainty in other views, formulating the problem of selecting the next-best-view. To further improve 6D object pose estimation, we propose an improved joint registration and hypotheses verification module as a final refinement step to reject false detections. We provide two additional challenging datasets inspired from realistic scenarios to extensively evaluate the state of the art and our framework. One is related to domestic environments and the other depicts a bin-picking scenario mostly found in industrial settings. We show that our framework significantly outperforms state of the art both on public and on our datasets. Unsupervised feature learning, although efficient, might produce sub-optimal features for our particular tast. Therefore in our last work, we leverage the power of Convolutional Neural Networks to tackled the problem of estimating the pose of rigid objects by an end-to-end deep regression network. To improve the moderate performance of the standard regression objective function, we introduce the Siamese Regression Network. For a given image pair, we enforce a similarity measure between the representation of the sample images in the feature and pose space respectively, that is shown to boost regression performance. Furthermore, we argue that our pose-guided feature learning using our Siamese Regression Network generates more discriminative features that outperform the state of the art. Last, our feature learning formulation provides the ability of learning features that can perform under severe occlusions, demonstrating high performance on our novel hand-object dataset. Concluding, this work is a research on the area of object detection and pose estimation in 3D space, on a variety of object types. Furthermore we investigate how accuracy can be further improved by applying active vision techniques to optimally move the camera view to minimize the detection error.

Robust model predictive control (RMPC) is widely used in industry. However, the online computational burden of this algorithm restricts its development and application to systems with relatively slow dynamics. We investigate this problem in this thesis with the overall aim of reducing the online computational burden and improving the online efficiency. In RMPC schemes, robust control invariant (RCI) sets are vitally important in dealing with constraints and providing stability. They can be used as terminal (invariant) sets in RMPC schemes to reduce the online computational burden and ensure stability simultaneously. To this end, we present a novel algorithm for the computation of full-complexity polytopic RCI sets, and the corresponding feedback control law, for linear discrete-time systems subject to output and initial state constraints, performance bounds, and bounded additive disturbances. Two types of uncertainty, structured norm-bounded and polytopic uncertainty, are considered. These algorithms are then extended to deal with systems subject to asymmetric initial state and output constraints. Furthermore, the concept of RCI sets can be extended to invariant tubes, which are fundamental elements in tube based RMPC scheme. The online computational burden of tube based RMPC schemes is largely reduced to the same level as model predictive control for nominal systems. However, it is important that the constraint tightening that is needed is not excessive, otherwise the performance of the MPC design may deteriorate, and there may even not exist a feasible control law. Here, the algorithms we proposed for RCI set approximations are extended and applied to the problem of reducing the constraint tightening in tube based RMPC schemes. In order to ameliorate the computational complexity of the online RMPC algorithms, we propose an online-offline RMPC method, where a causal state feedback structure on the controller is considered. In order to improve the efficiency of the online computation, we calculate the state feedback gain offline using a semi-definite program (SDP). Then we propose a novel method to compute the control perturbation component online. The online optimization problem is derived using Farkas' Theorem, and then approximated by a quadratic program (QP) to reduce the online computational burden. A further approximation is made to derive a simplified online optimization problem, which results in a large reduction in the number of variables. Numerical examples are provided that demonstrate the advantages of all our proposed algorithms over current schemes.

The work presented in this thesis proposes a novel adaptive magnetic adhesion mechanism that can be implemented in most locomotion mechanisms employed in climbing robots for ferric structures. This novel mechanism has the capability to switch OFF and ON its magnetic adhesion with minimal power consumption, and remain at either state after the excitation is removed. Furthermore, the proposed adhesion mechanism has the ability to adapt the strength of the adhesive force to a desired magnitude. These capabilities make the proposed adhesion mechanism a potential solution in the field of wall climbing robots. The novel contributions of the proposed mechanism include the switching of the adhesive force, selectivity of the adhesive force magnitude; determination of the parameters that have an impact in the final adhesive force. Finally, a final prototype is constructed with customised components and it is subject to a set of simulations and experiments to validate its performance.

The problem of Simultaneous Localisation And Mapping (SLAM) has been widely researched and has been of particular interest in recent years, with robots and self driving cars becoming ubiquitous. SLAM solutions to date have aimed to produce faster, more robust solutions that yield consistent maps by improving the filtering algorithms used, introducing better sensors, more efficient map representations or improved motion estimates. Whilst performing well in simplified scenarios, many of these solutions perform poorly in challenging real life scenarios. It is therefore important to produce SLAM solutions that can perform well even when using limited computational resources and performing a quick exploration for time critical operations such as Urban Search And Rescue missions. In order to address this problem this thesis proposes the construction of informative Bayesian priors to improve performance without adding to the computational complexity of the SLAM algorithm. Indoors occupancy grid SLAM is used as a case study to demonstrate this concept and architectural drawings are used as a source of prior information. The use of prior information to improve the performance of robotics systems has been successful in applications such as visual odometry, self-driving car navigation and object recognition. However, none of these solutions leverage prior information to construct Bayesian priors that can be used in recursive map estimation. This thesis addresses this problem and proposes a novel method to process architectural drawings and floor plans to extract structural information. A study is then conducted to identify optimal prior values of occupancy to assign to extracted walls and empty space. A novel approach is proposed to assess the quality of maps produced using different priors and a multi-objective optimisation is used to identify Pareto optimal values. The proposed informative priors are found to perform better than the commonly used non-informative prior, yielding an increase of over 20% in the F2 metric, without adding to the computational complexity of the SLAM algorithm.

Demand for autonomous multi-robot systems, where robots can cooperate with each other without human intervention, is set to grow rapidly in the next decade. Today, technologies such as self-driving cars and fleets of robotic assistants in hospitals and warehouses are being developed and used. In the future, robot swarms could be deployed in retrieval, reconnaissance and construction missions. Distributed collective systems have desirable properties, such as low cost of individual robots, robustness, fault tolerance and scalability. One of the main challenges in swarm robotics is that `bottom-up' approach to behaviour design is required. While the swarm performance is specied on the collective level of the swarm, robot designers need to program control algorithms of individual robots, while taking into account complex robot-robot interactions that allow emergence of collective intelligence. In order to be able to develop such systems, we need a methodology that aligns bottom-up design decisions with top-down design specifications. In this thesis, a novel approach to understanding and designing robot swarms that perform foraging and task allocation is proposed. Based on thousands of different simulation experiments, the Information-Cost-Reward framework is formulated, that relates the way in which a swarm obtains and uses information, to its ability to use that information in order to obtain reward efficiently. Secondly, based on the information-based understanding of swarm performance, design patterns for robot swarms are formalised. The design patterns are modular aspects of robot behaviour that dene when and how information should be obtained, exchanged or updated by robots, given particular swarm mission characteristics. Multiple design patterns can be unambiguously combined together in order to create a suitable robot control strategy. The design patterns specify robot behaviour in a newly developed Behaviour-Data Relations Modeling Language, where relationships between robot behaviour and data stored in and outside of robots are explicitly defined. This allows the design patterns to define behaviour of robots that cooperate and share information.

In recent years, professional first responders have started to use novel technologies at the scene of disasters in order to save more lives. Increasingly, they use robots to search disaster sites. One of the most widely and successfully used robot platforms in the disaster response domain are unmanned aerial vehicles (UAVs). UAVs allow remote inspection and mapping. They are able to provide high resolution imagery and often need minimal infrastructure to fly. This allows settings where multiple UAVs are airborne accelerating the information gathering from the disaster site. However, current deployments use labour intensive, individually teleoperated UAVs. Given this, there is a drive toward using multiple robots operating with a certain level of autonomy, in order to decrease the operators' workload. One approach for utilising multiple robots in this way is semi-autonomous operation supervised by a small number of professionals; only requiring human experts for crucial decisions. Current commercial UAV platforms also allow the deployment of a diverse group of robots, allowing them to combine their individual capabilities to be more ecient. For example, xed-wing UAVs are capable of flying faster and carry larger payload, but when they do so, they should be deployed with higher safety measures (safety pilots are required for non-lightweight aircraft). On the other hand, small rotary-wing UAVs are more agile and can approach and provide imagery about objects on the ground. To this end, this thesis develops a number of new approaches for the collaboration of a heterogeneous group of robots in disaster response. More specifically, the problem of collaborative planning with robots operating in an uncertain workflow based setting is investigated by solving the search and rescue (SAR) collaboration problem. Of course, the problem complexity increases when collaborating with dierent robots. It is not different in this setting, the actions of dierent types of robots need to be planned with dependencies between their actions under uncertainty. To date, research on collaboration between multiple robots has typically focused on known settings, where the possible robot actions are dened as a set of tasks. However, in most real world settings, there is a signicant amount of uncertainty present. For ii example, information about a disaster site develops gradually during disaster relief, thus initially there is often very little certainty about the locations of people requiring assistance (e.g. damaged buildings, trapped victims, or supply shortages). Existing solutions that tackle collaboration in the face of uncertain information are typically limited to simple exploration or target search problems. Moreover, the use of generic temporal planners rapidly becomes intractable for such problems unless applied in a domain-specific manner. Finally, domain specific approaches rarely involve complex action relations, such as task dependencies where the actions of some robots are built on the actions of others. When they do so, decomposition techniques are applied to decrease the problem complexity, or simple heuristics are applied to enhance similar collaboration. Such approaches often lead to low quality solutions, because vital action dependencies across different roles are not taken into account during the optimisation. Against this background, we oer novel online planning approaches for heterogeneous multi-robot collaboration under uncertainty. First, we provide a negotiation-based bidirectional collaborative planning approach that exploits the potential in determinisation via hindsight optimisation (HOP) combined with long-term planning. Second, we extend this approach to create an anytime Monte Carlo tree search planner that also utilises HOP combined with long-term planning. In online planning settings, such as SAR, anytime planners are benecial to ensure the ability of providing a feasible plan within the given computational budget. Third, we construct a scenario close to physical deployment that allows us to show how our long-term collaborative planning outperforms the current state of the art path-planning approaches by 25 %. We conclude that long-term collaborative planning under uncertainty provides an improvement when planning in SAR settings. When combined, the contributions presented in this thesis represent an advancement in the state of the art in the eld of online planning under uncertainty. The approaches and methods presented can be applied in collaborative settings when uncertainty plays an important role for defining dependencies between partial planning problems.

Inflatables have considerable packing efficiencies and allow simple deployment. This is due to their lack of stiffness when deflated, offering low cost solutions for space applications. After inflation, these typically soft systems could be improved using tape springs as structural stiffeners along the length of the boom, creating hybrid structures. These simple low cost single element components are easily stowed gaining strain energy in their collapsed state without permanent deformation providing additional potential to drive boom deployment. Combining both inflatable and tape spring components could create a superior hybrid boom with significant structural performance, whilst maintaining the advantages of gossamer structures. This research focuses on the structural performance improvement of adding tape springs to cantilever inflatable booms in over 40 experimental permutations. Applied tip loads identify the deflection response of these inflatable and hybrid booms, allowing a comparison between the two technologies. A computational hybrid boom model is developed alongside the experimental analysis using detailed material testing data of the inflatable fabric boom allowing an increased range of permutations and greater detail. The structural analysis has demonstrated the performance flexibility of hybrid booms where specific peak moment and rigidity requirements can be tailored through two key configurations; 2 opposed tape springs vertically aligned to the applied load and 4 tape springs in a cross formation square to the applied load respectively. A performance evaluation between the inflatable and hybrid booms shows significant potential whilst reducing the operational importance of maintaining pressurised systems. The greatest structural performance improvement is at 2.5 PSI with an increase of over 8 and 10 times for peak moment and boom rigidity respectively. This is achieved when adding 4 tapes in a cross formation to the inflatable boom with an added mass of 105%. This research has also highlighted the importance of the attachment method between the tape springs and inflatable boom with respect to packing efficiency, parasitic mass and structural performance trade offs.

This thesis describes new developments in nonlinear controllers for industrial applications. It first introduces the Nonlinear Generalised Minimum Variance (NGMV) control algorithm, for Linear Parameter Varying systems (LPV). This combines the benefits of the basic NGMV algorithm in dealing with nonlinearities, where a black box input model can be used, and adds an option to also approximate a nonlinear system with an LPV output subsystem. The models can therefore represent LPV systems and characteristics including saturation, discontinuities and time-varying dynamics. The next major contribution is in the nonlinear predictive control algorithms proposed that are also using the LPV model structure. The simplest is the Nonlinear Generalized Predictive Control (NGPC) algorithm that relates to the best known model predictive control law for linear systems. The final predictive control solution is one that may be specialized to either the NGMV or NGPC cases and is therefore the most general. This is referred to as a Nonlinear Predictive Generalized Minimum Variance Controller (NPGMV). When the algorithms use only the LPV structure to approximate the nonlinear system the solutions are particularly simple in unconstrained and constrained versions, and are relatively light computationally for implementation. Three representative industrial design examples have been chosen to validate the algorithms for different Bandwidth (BW) and nonlinear characteristics. All three examples were based on real application problems with company interest. In the first example (small BW) the basic state-space and LPV versions of the algorithm are used for the auto-manoeuvring and dynamic positioning of marine vessel. In this application the parameter variations were representative of wave disturbance changes with sea state, rather than due to approximating nonlinear behaviour. Actuator constraints were considered in the design. In the second industrial example (medium BW) the LPV-NPGMV was implemented for controlling the blade pitch and generator torque of a 5MW offshore wind turbine. The main objective here was to maintain the power produced at the rated value which requires compensation against wind disturbances, so that wind speed is the varying parameter. The LPV-NPGMV controller produced here used a parameterised system model involving the wind speed so that the controller performance changed with wind conditions. Actuator constraints were included and statistical performance assessed. The third example (fast BW) explores the stabilisation of a 2-axis gyroscopic electro-optical turret used in surveillance applications. This application was designed and employed on a real system. Because of the limitations imposed by BW requirements and the memory of the digital controller, only the basic state-space version of the algorithm was possible to implement. The main objective in this problem was to improve the tracking performance around the NADIR singularity point (a discontinuity) in the trajectory. In all three examples the NGMV controllers showed notable improvement in comparison to the baseline controllers without the need for scheduled gains or re-configuration when moving across different operating points.

As future robotic surface exploration missions to other planets, moons and asteroids become more ambitious in their science goals, there is a rapidly growing need to significantly enhance the capabilities of entry, descent and landing technology such that landings can be carried out with pin-point accuracy at previously inaccessible sites of high scientific value. As a consequence of the extreme uncertainty in touch-down locations of current missions and the absence of any effective hazard detection and avoidance capabilities, mission designers must exercise extreme caution when selecting candidate landing sites. The entire landing uncertainty footprint must be placed completely within a region of relatively flat and hazard free terrain in order to minimise the risk of mission ending damage to the spacecraft at touchdown. Consequently, vast numbers of scientifically rich landing sites must be rejected in favour of safer alternatives that may not offer the same level of scientific opportunity. The majority of truly scientifically interesting locations on planetary surfaces are rarely found in such hazard free and easily accessible locations, and so goals have been set for a number of advanced capabilities of future entry, descent and landing technology. Key amongst these is the ability to reliably detect and safely avoid all mission critical surface hazards in the area surrounding a pre-selected landing location. This thesis investigates techniques for the use of a single camera system as the primary sensor in the preliminary development of a hazard detection system that is capable of supporting pin-point landing operations for next generation robotic planetary landing craft. The requirements for such a system have been stated as the ability to detect slopes greater than 5 degrees and surface objects greater than 30cm in diameter. The primary contribution in this thesis, aimed at achieving these goals, is the development of a feature-based,self-initialising, fully adaptive structure from motion (SFM) algorithm based on a robust square-root unscented Kalman filtering framework and the fusion of the resulting SFM scene structure estimates with a sophisticated shape from shading (SFS) algorithm that has the potential to produce very dense and highly accurate digital elevation models (DEMs) that possess sufficient resolution to achieve the sensing accuracy required by next generation landers. Such a system is capable of adapting to potential changes in the external noise environment that may result from intermittent and varying rocket motor thrust and/or sudden turbulence during descent, which may translate to variations in the vibrations experienced by the platform and introduce varying levels of motion blur that will affect the accuracy of image feature tracking algorithms. Accurate scene structure estimates have been obtained using this system from both real and synthetic descent imagery, allowing for the production of accurate DEMs. While some further work would be required in order to produce DEMs that possess the resolution and accuracy needed to determine slopes and the presence of small objects such as rocks at the levels of accuracy required, this thesis presents a very strong foundation upon which to build and goes a long way towards developing a highly robust and accurate solution.

This thesis studies the design and implementation of anti windup compensators for UAVs with magnitude and rate saturated actuators. The focus is on two types of UAVs; a Quadrotor UAV and a Fixed wing UAV. Decentralized anti-windup compensators are designed to address the problem of magnitude saturation in Quadrotor UAVs. The developed anti-windup compensators are founded on an LMI-based approach previously used in literature to provide global stabilty guarantees with some level of performance guarantees. The work on the decentralized anti-windup compensators for Quadrotor UAVs are further improved on by replacing the use of LMIs in the determination of the anti windup compensator parameters with approximate linear based guidelines after a Lure-Postinikov Lyapunov function is used to provide global stability guarantees. This approach applies not only to Quadrotor UAVs but also to a wide class of systems that contain double integrators. The developed anti-windup compensators were designed and implemented for an experimental Quadrotor UAV where both simulation results and flight test results clearly show the ability of the anti-windup compensators to reduce the effect of magnitude saturation in Quadrotor UAVs. Finally, the thesis describes the design of decoupled multivariable anti-windup compensators to tackle the problem of rate saturation on a fixed wing UAVs. Simulation results obtained demonstrate that these anti-windup compensators are capable of managing the system responses during periods of rate saturation.