TECHNOLOGIES FOR FAST RECONFIGURATION OF ADAPTIVE ROBOTIC WORKCELLSGAŠPAR, TIMOTEJ (Avtor) Ude, Aleš (Mentor) reconfigurable manufacturing systemsflexible fixturesStewart platformlayout optimizationprogramming by demonstrationdynamic movement primitivesstatistical learning of robot skillsRobots have become a crucial component in contemporary manufacturing. Due to their large workspace, high precision and repeatability, they are used in a broad spectrum of automation. They especially excel in industrial environments, replaying the same motions and performing the same tasks repeatedly over extended periods of time. The usefulness of robots, however, still falls short in industries where the production demands change frequently. In such production environments, the traditional automation approaches and the corresponding robot workcells still do not offer a suitable solution. It is therefore necessary to explore the available options and progress the scientific field into more adaptable robotic workcells in order to bring automation to these types of industries. In this thesis, we present novel technologies and methods aimed at increasing the ease of reconfiguration and shortening setup times of adaptive robot workcells. The first chapter of this thesis provides an insight to the main topic addressed in this thesis and presents the current state of the art. We begin the chapter with a broader presentation of the traditional manufacturing systems and how they fall short when greater flexibility is needed. We continue by introducing the paradigm of Reconfigurable Manufacturing Systems (RMS), which states that manufacturing systems should be adaptable in a quick and efficient manner to unexpected changes in market demands and consequently production specifications. One of the components that are often needed in a manufacturing system are fixtures, which should also conform to the RMS characteristics. We present the concept of passive flexible fixtures and explain the challenge of their placement so that a set of workpieces can be mounted onto them. Another aspect that needs to be addressed in order to shorten setup and reconfiguration times is programming of robot motions. We present how methods for statistical learning in conjunction with Programming by Demonstration (PbD) can be used to increase the efficiency of robot programming. However, the results of standard PbD methods are susceptible to variations in the speed profiles of user demonstrations. We tackle this issue by introducing a statistical learning method based on arc-length dynamic movement primitives (AL-DMP). We conclude the chapter with a summary of novel contributions of this thesis. The second chapter presents the software and hardware design paradigms for building adaptive robotic cells and how they were utilized in a prototype cell. We first present the hardware aspects of the cell and the novel approach to build reconfigurable hardware by using reconfigurable components with passive degrees of freedom. This type of components are built without actuators or sensors but can be reconfigured by a robot arm. We then provide a detailed description of the software system architecture. The software system is based on the Robot Operating System (ROS), which allowed us to make it reconfigurable and to support the reconfigurable hardware. Additionally, we describe how the robot programming process was enhanced in the developed cell by making use of PbD and a robot skill database. The chapter is concluded with a summary of the novel workcell design paradigms. The third chapter provides a detailed description of 1. the developed methodology for the reconfiguration of fixtures with passive degrees of freedom, called hexapods in what follows, and 2. the new method for statistical learning of robot skills based on kinesthethic guidance. We start by describing a new optimization method that can be used to automatically determine the layout of the hexapods in order to make it possible to mount a set of workpieces without the need to re-position their bases. The method considers the kinematic limitations of the fixturing system and the physical limits of the cell layout, including collisions. We also provide a strategy for how to generate a robot trajectory to reconfigure the fixtures in order to avoid the kinematic limits during the reconfiguration process. In the second part of this chapter we present a novel method for statistical learning of robot skills that uses arc-length dynamics movement primitives – AL-DMPs – to represent robot skills. We provide its mathematical formulation and explain its benefits over the standard DMP formulation. We conclude the chapter by explaining how all the developed methods and paradigms can be used together to achieve fast setup and reconfiguration of adaptive robot workcells. The experimental evaluation of the methods and paradigms proposed in this thesis are presented in the fourth chapter. We first explain the results of the implementation of various industrial use-cases in the developed prototype workcell. This was done to evaluate the proposed adaptive cell design paradigms. The results show the industrial readiness of the system and that it can achieve the desired performance in terms of setup and reconfiguration times. Next we show how the proposed optimization method to determine the layout of the hexapods assures that the kinematic limitations and other physical constraints of the workspace are respected. Finally, we present the performance of the AL-DMP representation for statistical learning and action recognition. The fifth and final chapter of the thesis contains the discussion and final remarks. Each of the scientific contributions is briefly summarized and discussed. The possibilities for future work are also laid out. At the end of the chapter, we delineate the contributions in peer-reviewed journals and conferences that support the scientific relevance of the presented research results.20202020-12-16 10:20:01Doktorsko delo/naloga122919sl