Using UAVs in everyday life has been increasing in recent years. UAV is an agile vehicle and often comes integrated with a camera and sensors which makes it suitable for object detection and tracking. In this thesis, we present a subsystem with a limited hardware setup only consisting of an on-board computer and a camera that is mounted on a UAV. The subsystem provides techniques to maneuver, detect and descend to an object, all executed autonomously. The system is implemented in Robotic Operating System (ROS). The object detection is implemented as a convolutional neural network provided by TensorFlow Object Detection API. This thesis covers the necessary steps to adopt a pre-trained TensorFlow model to specific needs and compares three different TensorFlow models considering accuracy, frames per second and energy efficiency. Additionally, methodologies to cover a predefined area and position an object in relation to the camera is proposed. Experiments are executed both in a real-world and simulated environment and the results are promising for the implemented system. / Användandet av UAVs i det vardagliga livet har ökat markant de senaste åren. En UAV är ett agilt fordon som ofta kommer integrerat med en kamera samt sensorer som gör det till ett lämpligt fordon för objektigenkänning och spårning. I den här avhandligen presenterar vi ett delsystem med en hårdvaruplattform endast bestående av en inbyggd dator och en kamera. Delsystemet tillhandahåller metoder som gör det möjligt för UAV:en att styras, känna igen objekt och landa på det detekterade objektet autonomt. Systemet implementeras i Robotic Operating System (ROS). Objektigenkänningen är implementerat som ett konvolutionellt neuralt nätverk tillhandahållt av TensorFlow Object Detection API. Avhandlingen omfattar stegen nödvändiga att ta för att anpassa en TensorFlow model till sina egna behov och gör jämförelser mellan tre olika Tensorflow modeller med avseende på precision, bildrutor per sekund och energi effektivitet. Dessutom presenteras metoder för att söka av ett fördefinierat område och positionering av ett objekt relativt komeran. Under experiment, både i simulering och verkliga världen, har lovande resultat framkommit.

The reported work lies in the area of automation in the construction industry, and involves multi-disciplinary engineering studies. In particular, sewer renovation methods, computer vision (CV) and robotics are all included. More specifically, the key objective of the research programme was to investigate the characteristics of retrofit components suited to mounting on an industrial / proprietary sewer tractor. The overall aim was the provision of a non-man-entry (NME) sewer renovation robot to undertake reconnection of lateral junctions, following a cured-in-place (CIP) relining process. The programme primarily involved theoretical studies of the requisite sensory and kinematic components, incorporation of a novel computer vision sensing system and production of a chainage measurement system and robotic drill task arm. The theory was supported by laboratory testing using a modified proprietary tractor, with emphasis placed on promoting applications of information technology driven systems (i.e. CV) to construction-industry tasks within hazardous environments involving significant health issues. The use of such techniques in the construction industry is rare. Chapter 1 reviews the context and history of sewer maintenance/dereliction in the UK. NME sewers are the most common type and are, by definition, difficult to maintain. Renovation, typically employing CIP liners, is therefore a cost-effective alternative to replacement. Lateral connections are, inevitably, blocked off during the relining process; it is suggested that application of a robust robotic system to the task of reconnecting them is novel and offers clear potential within such a hazardous environment. Chapters 2 and 3 develop the underlying theoretical models of the CV and kinematic systems respectively. The novel CV work (provided by third party specialists employing the TINA CV research environment) was incorporated by the author to provide detection and classification of lateral junctions, crucially noting the particular properties of direct and reflected illumination. Classification aspects include estimation of lateral/NME intersection angle and closure-to-target distance from the robot. The author proposes a separate procedure for estimating lateral diameter. A chainage measurement system, using a rotary encoder and inclinometer, was developed to determine invert path distance travelled. This allows for the inevitable wander and thereby gives the system robustness. The novel application of GRASP (a robotic modelling and simulation design tool) to NME environments, provided the ability to model arm designs without the need for the production of more than one expensive physical prototype. A mathematical solution for determining the requisite arm kinematics is presented. Chapter 4 details the hardware requirements of the robotic system components, whilst Chapters 5 and 6 present the laboratory evaluation results for the kinematic and CV systems respectively. The abilities of the CV system qualitatively to detect laterals under reflected illumination, and to provide quantitative classification data, are demonstrated. The chainage measurement system is assessed under a variety of initialisation conditions to determine suitability to task, and the ability of the robotic arm to physically simulate lateral reconnection is investigated. Chapter 7 discusses the specification for an industrially-applicable prototype, based on the findings herein. Appropriate comparisons with the pre-prototype system are made, including cost. Finally, Chapter 8 draws conclusions and makes suggestions for further work. Supporting documentation is provided in Chapter 9 and the Appendices.

General autonomy is at the forefront of robotic research and practice. Earlier research has enabled robots to learn movement and manipulation within the context of a specific instance of a task and to learn from large quantities of empirical data and known dynamics. Reinforcement learning (RL) tackles generalisation, whereby a robot may be relied upon to perform its task with acceptable speed and fidelity in multiple---even arbitrary---task configurations. Recent research has advanced approximate policy search methods of RL, in which a function approximator is used to represent an optimal policy while avoiding calculation across the large dimensions of the state and action spaces of real robots. This thesis details the implementation and testing, on a lightweight industrial robot arm, of guided policy search (GPS), an RL algorithm that seeks to avoid the typical need, in machine learning, for lots of empirical behavioural samples, while maximising learning speed. GPS comprises a local optimal policy generator, here based on a linear-quadratic regulator, and an approximate general policy representation, here a feedforward neural network. A controller is written to interface an existing back-end implementation of GPS and the robot itself. Experimental results show that the GPS agent is able to perform basic reaching tasks across its configuration space with approximately 15 minutes of training, but that the local policies generated fail to be fully optimised within that timescale and that post-training operation suffers from oscillatory actions under perturbed initial joint positions. Further work is discussed and recommended for better training of GPS agents and making locally optimal policies more robust to disturbance while in operation.

No description available.

No description available.

The aim of this work is to introduce autonomous robots for collecting waste bins in an officeenvironment. Since the environment will contain humans, therefore, the robot-human interactionis essential so that neither the robot nor humans come in harm’s way. To prevent this, the robotneeds to be able to communicate with the humans, and it has to operate in a safe way in the officeenvironment. To facilitate this there is a need to investigate the state-of-the-art in various fieldssuch as: mapping, path planning and, machine-human interaction. The result is a robot that isable to map an indoor environment and use the map to its advantages for producing a suitable pathfrom any given position to any desired location. While traversing the path, sensor data is gatheredfor obstacle detection and avoidance in order to handle dynamic environments.

This thesis presents a robotic gripper with an intelligent sensor system to grasp objects with an adaptive grasp technique. Two techniques are used, one for small objects and one for large objects. The sensor system is able to detect the object and measure its size to adapt the grasp. Optical motion sensors are used to see when the object is slipping between the fingers which means that more force needs to be applied. This makes it possible to grasp rigid and soft objects without damaging them. The functionality of the gripper was tested on eight objects with various characteristics. The results show that it can adapt the grasp technique and grasping force to the objects’ size and softness. It also shows that adding one more grasp technique made it possible to grasp more objects, compared to a different gripper with a single grasp technique which were used as a foundation for this thesis.

No description available.

Testsystem är en stor del i dagen samhälle. Att testa en produkt är viktigt för att se hur länge den håller, extra viktigt att testa produkter är det när produkterna ska användas till välfärdsteknologi oche-hälsa då tekniken ska användas i människors vardag. Phoniro utvecklar produkter och tjänster inom välfärdsteknologi och e-hälsa. Exempel på produkt som tillverkas på Phoniro är digitala lås där testningen främst består av mjukvarutester under förhållanden där en verklig aspekt på användning av produkten är önskad men inte är medräknad. Därför vill Phoniro ha en automatiserad testmaskin som på ett mer korrekt sätt återspeglar verkligheten. Målet med projektet var att testa digitala lås där verklig aspekt på användning av produkten ärmedräknad. Det beslutades att göra en testmaskin som skulle innehålla en räknare för att se hur länge testet har körts och sensorer som ska känna av om något havererar. En testmaskin har tillverkats som kan testa det digitala låset. En räknare visar hur många gånger den har öppnat låset. Testmaskinen stannar om något havererat. Målen men projektet har uppnåtts, erfarenheter projektet har gett är praktiska och teoretiska kunskaper om processen att bygga upp ett testsystem.

Learning techniques are drawing extensive attention in the robotics community. Some reasons behind moving from traditional preprogrammed robots to more advanced human fashioned techniques are to save time and energy, and allow non-technical users to easily work with robots. Learning from Demonstration (LfD) and Imitation Learning (IL) are among the most popular learning techniques to teach robots new skills by observing a human or robot tutor. Flawlessly teaching robots new skills by LfD requires good understanding of all challenges in the field. Studies of imitation learning in humans and animals show that several cognitive abilities are engaged to correctly learn new skills. The most remarkable ones are the ability to direct attention to important aspects of demonstrations, and adapting observed actions to the agents own body. Moreover, a clear understanding of the demonstrator's intentions is essential for correctly and completely replicating the behavior with the same effects on the world. Once learning is accomplished, various stimuli may trigger the cognitive system to execute new skills that have become part of the repertoire. Considering identified main challenges, the current thesis attempts to model imitation learning in robots, mainly focusing on understanding the tutor's intentions and recognizing what elements of the demonstration need the robot's attention. Thereby, an architecture containing required cognitive functions for learning and reproducing high-level aspects of demonstrations is proposed. Several learning methods for directing the robot's attention and identifying relevant information are introduced. The architecture integrates motor actions with concepts, objects and environmental states to ensure correct reproduction of skills. This is further applied in learning object affordances, behavior arbitration and goal emulation. The architecture and learning methods are applied and evaluated in several real world scenarios that require clear understanding of goals and what to look for in the demonstrations. Finally, the developed learning methods are compared, and conditions where each of them has better applicability is specified.