This thesis work investigated the advantages and disadvantages of using superquadrics (SQ) to do the collision-checking part of the Rapidly-exploring Random Trees (RRT) motion planning algorithm for higher Degree of Freedom (DoF) motion planning, comparing it with an established proximity querying method known as the Gilbert-Johnson-Keerthi (GJK) algorithm. In the RRT algorithm, collision detection is the main bottleneck, making this topic interesting to research. The SQ-based collision detection method was compared to the GJK algorithm both qualitatively and quantitatively, comparing computational speed, memory requirements, as well as the ability to handle arbitrary shapes. Furthermore, how appropriate they are in modelling a 6 DoF arm was analyzed. A qualitative comparison between the RRT algorithm and the A* algorithm was also provided, comparing their suitability for searching in higher dimensional spaces. When there were no collisions the SQ-based algorithms performed roughly at parity with the GJK algorithm in terms of computational speed. However, when a collision had occurred, the SQ-based algorithms were able to return a positive faster than the GJK algorithm, outperforming it. From a memory standpoint the SQ-based algorithms required less memory as they could leverage the explicit and implicit representations of the SQ objects, whereas the GJK algorithm requires both objects being checked for collision to be explicitly represented as convex sets of points. Regarding handling arbitrary shapes, the SQ-based algorithms have an advantage in that they can allow for certain non-convex shapes to be. Conversely, the GJK algorithm is limited to convex shapes. The GJK algorithm would thus require more geometric primitives to accurately capture the same non-convex shape. Thus, it can be concluded that the SQ-based method is more suitable for modelling a 6 DoF arm. However, a GJK-based collision detection module would in most cases be a lot more straightforward than the alternative to set up, as it is very simple to collect a set of points. Finally, both collision detection method types were implemented with the RRT algorithm. Due to the inherently random nature of the RRT algorithm the results of this set of tests could not be used to make any further conclusions beyond showing that it is possible to combine the SQbased algorithm with the RRT algorithm. Instead, one should see the RRT algorithm as a multiplicative factor applied to the inherent properties of the previously examined collision detection methods. / Detta examensarbete undersökte fördelarna och nackdelarna med att använda n:tegradsytor (NY) för att utföra kollisionsdetektion i algoritmen Raskt-utforskande Slumpmässiga Träd (RST). RST används typiskt för planeringen av system med relativt många frihetsgrader. En etablerad metod för kollisionsdetektion, Gilbert-Johnson-Keerthi-algoritmen (GJK), implementerades även i jämförelsesyfte. Då GJK-algoritmens största flaskhals ligger i kollisionsdetektionen är detta ett intressant ämne att efterforska. Den NY-baserade kollinsdetektionsmetoden jämfördes med den GJK-baserade metoden både kvantitativt och kvalitativt. Kvalitativt jämfördes beräkningshastighet och minnesåtagande, medan de kvalitativt jämfördes i deras förmåga att representera godtyckliga geometriska former. På ett högre plan diskuterades det även hur lämpliga de är för att modellera en robotarm med 6 stycken frihetsgrader. RST-algoritmen jämfördes även med en annan planeringsalgoritm, A*. Framförallt fokuserade diskussionen kring planering av system med relativt många frihetsgrader. I det fall inga kollisioner fanns presterade GJK-algoritmen ungefär lika bra som NY algoritmerna i att fastslå detta, utifrån beräkningshastighet. Men när det kom till att upptäcka existerande kollisioner presterade GJK-algoritmen sämre. Minnesmässigt använder GJK-algoritmen mer minne, då den kräver att båda objekten är explicitrepresenterade (dvs, som ett punktmoln), medan man med en NY-metod endast behöver representera ena objektet explicit och den andra implicit. Gällande förmågan att representera godtyckliga geometriska former är NY-baserade metoder bättre. Till skillnad från GJK som är begränsad till konvexa mängder kan NY uppta ickekonvexa former, exempelvis flottyrmunkformade supertoroider. En metod som använder GJKalgoritmen skulle behöva bygga upp icke-konvexa former med flera mindre konvexa komponenter. NY-metoden är således bättre för att modellera robotarmar med 6 frihetsgrader. Det är dock i praktiken lättare att implementera GJK-metoden då den endast kräver punktmoln, medan NY kräver parametrar som måste bestämmas eller finjusteras. RST-algoritmen implementerades sist, utformad så att kollisionsdetektionsmetoderna är utbytbara. Det var dock inte möjligt att dra slutsatser utifrån det testdata som erhölls, ty RSTalgoritmens slumpmässiga karaktär. RST-algoritmen kan ses som en multiplikator som endast förstorar de inneboende egenskaperna hos kollisionsdetektionsmetoderna.

In order for assistive robots to collaborate effectively with humans for completing everyday tasks, they must be endowed with the ability to effectively perceive scenes and more importantly, recognize human intentions. As a result, we present in this dissertation a novel scene-dependent human-robot collaborative system capable of recognizing and learning human intentions based on scene objects, the actions that can be performed on them, and human interaction history. The aim of this system is to reduce the amount of human interactions necessary for communicating tasks to a robot. Accordingly, the system is partitioned into scene understanding and intention recognition modules. For scene understanding, the system is responsible for segmenting objects from captured RGB-D data, determining their positions and orientations in space, and acquiring their category labels. This information is fed into our intention recognition component where the most likely object and action pair that the user desires is determined. Our contributions to the state of the art are manifold. We propose an intention recognition framework that is appropriate for persons with limited physical capabilities, whereby we do not observe human physical actions for inferring intentions as is commonplace, but rather we only observe the scene. At the core of this framework is our novel probabilistic graphical model formulation entitled Object-Action Intention Networks. These networks are undirected graphical models where the nodes are comprised of object, action, and object feature variables, and the links between them indicate some form of direct probabilistic interaction. This setup, in tandem with a recursive Bayesian learning paradigm, enables our system to adapt to a user's preferences. We also propose an algorithm for the rapid estimation of position and orientation values of scene objects from single-view 3D point cloud data using a multi-scale superquadric fitting approach. Additionally, we leverage recent advances in computer vision for an RGB-D object categorization procedure that balances discrimination and generalization as well as a depth segmentation procedure that acquires candidate objects from tabletops. We demonstrate the feasibility of the collaborative system presented herein by conducting evaluations on multiple scenes comprised of objects from 11 categories, along with 7 possible actions, and 36 possible intentions. We achieve approximately 81% reduction in interactions overall after learning despite changes to scene structure.

Cette thèse porte sur le développement d'une loi de guidage autonome par platitude pour les véhicules de rentrée atmosphérique. La problématique associée au développement d'une loi de guidage autonome porte sur l'organisation globale, l'intégration et la gestion de l'information pertinente jusqu'à la maîtrise du système spatial durant la phase de rentrée. La loi de guidage autonome proposée dans ce mémoire s'appuie sur le concept de platitude, afin d'effectuer un traitement des informations à bord, dans le but double d'attribuer un niveau de responsabilité et d'autonomie au véhicule, déchargeant ainsi le segment sol de tâches opérationnelles "bas niveau", pour lui permettre de mieux assumer son rôle de coordination globale. La première partie de ce mémoire traite de la caractérisation formelle de sorties plates pour les systèmes non linéaires régis par des équations différentielles ordinaires, ainsi que pour les systèmes linéaires à retards. Des algorithmes constructifs sont proposés afin de calculer des sorties plates candidates sous un environnement de calcul formel standard. Dans la seconde partie, une méthodologie complète et générique de replanification de trajectoires de rentrée atmosphérique est proposée, afin de doter la loi de guidage d'un certain niveau de tolérance à des pannes actionneur simple/multiples pouvant survenir lors des phases critiques d'une mission de rentrée atmosphérique. En outre, une méthodologie d'annexation superellipsoidale est proposée afin de convexifier le problème de commande optimale décrit dans l'espace des sorties plates. La loi de guidage proposée est ensuite appliquée étape par étape à une mission de rentrée atmosphérique pour la navette spatiale américaine STS-1. / This thesis deals with the design of an autonomous guidance law based on flatness approach for atmospheric reentry vehicles. The problematic involved by the design of an autonomous guidance law relates to the global organization, the integration and the management of relevant data up to the mastering of the spacecraft during the re-entry mission. The autonomous guidance law proposed in this dissertation is based on flatness concept, in order to perform onboard processing so as to locally assign autonomy and responsibility to the vehicle, thus exempting the ground segment from "low level" operational tasks, so that it can ensure more efficiently its mission of global coordination. The first part of the manuscript deals with the formal characterization of flat outputs for nonlinear systems governed by ordinary differential equations, as well as for linear time-delay systems. Constructive algorithms are proposed in order to compute candidate flat outputs within a standard formal computing environment. In the second part of the manuscript, a global and generic reentry trajectory replanning methodology is proposed in order to provide a fault-tolerance capability to the guidance law, when facing single/multiple control surface failures that could occur during the critical phases of an atmospheric reentry mission. In addition, a superellipsoidal annexion method is proposed so as to convexify the optimal control problem described in the flat outputs space. The proposed guidance law is then applied step by step to an atmospheric reentry mission for the US Space Shuttle orbiter STS-1.