Search

We propose a time-domain multiple sound source localization (SSL) method based on Bayesian inference. This method is specifically designed to run on the stochastic machines (SM) that we are currently developing to perform efficient low-level sensor signal processing with ultra-low power consumption. The proposed SSL method is divided into two main parts. First, a probabilistic model is run on 50 very short time frames (3. 75ms each) of multichannel recorded signals. Second, the results obtained on the different frames are fused to obtain a final localization map. Using the system in a supervised way allows to extract estimated source locations by selecting as many maxima as there are sources in the room. We explain how this method is implemented on a SM. Experiments are presented to illustrate the performance and robustness of the resulting system.

International audience; Bayesian models and stochastic computing form a promising paradigm for non-conventional, bio-inspired computation architectures. In particular, they are able to handle uncertainty and promise low power consumption. In this paper we study the application of such an architecture, the Sliced Bayesian Machine (SlicedBM) to a real-world problem, Sound Source Localization (SSL) for robots. We present an analysis of the quality of results and of computing time according to several parameters: sensor precision, result threshold, internal word length. Furthermore, we show that sensor data precision does not heavily influence the computation. On the opposite, the precision of the probability values plays an important role on result quality. This parameter also determines the circuit size. We also show that the higher the re-sampling threshold (RT), the better the distribution computed by the machine. Our results make it possible to choose optimal design parameters for a circuit along several trade-offs, and according to a given sensor fusion application.

International audience; Compared to conventional processors, stochastic computing architectures have strong potential to speed up computation time and to reduce power consumption. We present such an architecture, called Bayesian Machine (BM), dedicated to solving Bayesian inference problems. Given a set of noisy signals provided by low-level sensors, a BM estimates the posterior probability distribution of an unknown target information. In the present study, a BM is used to solve a sound source localization (SSL) problem: the BM computes the probability distribution of the position of a sound source given acoustic signals captured by a set of microphones. Assuming free field wave propagation (no reverberations), we express the SSL problem as the maximization of a likelihood function fed with audio features provided by the time-frequency (TF) analysis of the captured audio waves. The proposed BM uses bitwise parallel sampling to fuse the resulting multi-channel information. As the number of channels to fuse is large, the standard BM architecture encounters the so-called " time dilution problem " (long delays are necessary to obtain valid samples). We tackle this problem by using max-normalization of the distributions combined with a periodic re-sampling of the bit streams after processing a reasonably small subset of evidences. Finally, we compare the localization performance of the proposed machine with the results obtained using a standard version of the machine. The re-sampling leads to an impressive acceleration factor of 10³ in the computation.

International audience; —As the physical limits of Moore's law are being reached, a research effort is launched to achieve further performance improvements by exploring computation paradigms departing from standard approaches. The BAMBI project (Bottom-up Approaches to Machines dedicated to Bayesian Inference) aims at developing hardware dedicated to probabilistic computation , which extends logic computation realised by boolean gates in current computer chips. Such probabilistic computing devices would allow to solve faster and at a lower energy cost a wide range of Artificial Intelligence applications, especially when decisions need to be taken from incomplete data in an uncertain environment. This paper describes an architecture where very simple operators compute on a time coding of probability values as stochastic signals. Simulation tests and a reconfigurable logic hardware implementation demonstrated the feasibility and performances of the proposed inference machine. Hardware results show this architecture can quickly solve Bayesian sensor fusion problems and is very efficient in terms of energy consumption.

International audience; This work revisits the stochastic computationparadigm as a way to implement architectures dedicated toBayesian computation. It is assumed that Stochastic BayesianMachines (SMBs) are intrinsically tolerant to the effects ofradiation. However, practical assessment is mandatory beforeconsidering SBMs in hazardous environments. Results of faultinjectioncampaigns performed at the RTL level provide the firstevidences of SBMs robustness with respect to SEUs and SETs.

International audience; In the present paper we describe a bio-inspired non von Neumann controller for a simple sensorimotor robotic system. This controller uses a bitwise version of the Gibbs sampling algorithm to select commands so the robot can adapt its course of action and avoid perceived obstacles in the environment. The VHDL specification of the circuit implementation of this controller is based on stochastic computation to perform Bayesian inference at a low energy cost. We show that the proposed unconventional architecture allows to successfully carry out the obstacle avoidance task and to address scalability issues observed in previous works.

International audience; We present an architecture and a compilation toolchain for stochastic machines dedicated to Bayesian inferences. These machines are not Von Neumann and code information with stochastic bitstreams instead of using floating point representations. They only rely on stochastic arithmetic and on Gibbs sampling to perform approximate inferences. They use banks of binary random generators which capture the prior knowledge on which the inference is built. The output of the machine is devised to continuously sample the joint probability distribution of interest. While the method is explained on a simple example, we show that our machine computes a good approximation of the solution to a problem intractable in exact inference.

International audience; Features Presents a new modeling methodology and inference algorithms for Bayesian programming Explains how to build efficient Bayesian models Addresses controversies, historical notes, epistemological debates, and tricky technical questions in a dedicated chapter separate from the main text Encourages further research on new programming languages and specialized hardware for computing large-scale Bayesian inference problems Offers an online Python package for running and modifying the Python program examples in the book Summary Probability as an Alternative to Boolean Logic While logic is the mathematical foundation of rational reasoning and the fundamental principle of computing, it is restricted to problems where information is both complete and certain. However, many real-world problems, from financial investments to email filtering, are incomplete or uncertain in nature. Probability theory and Bayesian computing together provide an alternative framework to deal with incomplete and uncertain data. Decision-Making Tools and Methods for Incomplete and Uncertain Data Emphasizing probability as an alternative to Boolean logic, Bayesian Programming covers new methods to build probabilistic programs for real-world applications. Written by the team who designed and implemented an efficient probabilistic inference engine to interpret Bayesian programs, the book offers many Python examples that are also available on a supplementary website together with an interpreter that allows readers to experiment with this new approach to programming. Principles and Modeling Only requiring a basic foundation in mathematics, the first two parts of the book present a new methodology for building subjective probabilistic models. The authors introduce the principles of Bayesian programming and discuss good practices for probabilistic modeling. Numerous simple examples highlight the application of Bayesian modeling in different fields. Formalism and Algorithms The third part synthesizes existing work on Bayesian inference algorithms since an efficient Bayesian inference engine is needed to automate the probabilistic calculus in Bayesian programs. Many bibliographic references are included for readers who would like more details on the formalism of Bayesian programming, the main probabilistic models, general purpose algorithms for Bayesian inference, and learning problems. FAQs Along with a glossary, the fourth part contains answers to frequently asked questions. The authors compare Bayesian programming and possibility theories, discuss the computational complexity of Bayesian inference, cover the irreducibility of incompleteness, and address the subjectivist versus objectivist epistemology of probability. The First Steps toward a Bayesian Computer A new modeling methodology, new inference algorithms, new programming languages, and new hardware are all needed to create a complete Bayesian computing framework. Focusing on the methodology and algorithms, this book describes the first steps toward reaching that goal. It encourages readers to explore emerging areas, such as bio-inspired computing, and develop new programming languages and hardware architectures.

International audience; In this paper, we consider the application of a naïve Bayes model for the evaluation of corruption risk associated with government agencies. This model applies probabilistic classifiers to support a generic risk assessment model, allowing for more efficient and effective use of resources for the detection of corruption in government transactions, and assisting audit agencies in becoming more proactive regarding corruption detection and prevention.

voir basilic : http://emotion.inrialpes.fr/bibemotion/2005/DBM05/; International audience; This paper deals with the probabilistic modeling of space, in the context of mobile robot navigation. We define a formalism called the Bayesian Map, which allows incremental building of models, thanks to the Superposition operator, which is a formally well-defined operator. Firstly, we present a syntactic version of this operator, and secondly, a version where the previously obtained model is enriched by experimental learning. In the resulting map, locations are the conjunction of underlying possible locations, which allows for more precise localization and more complex tasks. A theoretical example validates the concept, and hints at its usefulness for realistic robotic scenarios.

voir basilic : http://emotion.inrialpes.fr/bibemotion/2004/DBM04/ address: New Orleans, LA (US); This paper presents a new method for probabilistic modeling of space, called the Bayesian Map formalism. It offers a generalization of some common approaches found in the literature, as it does not constrain the dependency structure of the probabilistic model. The formalism allows incremental building of hierarchies of models, by the use of the Abstraction operator. In the resulting hierarchy, localization in the high level model is based on probabilistic competition of the lower level models. Experimental results validate the concept, and hint at its usefulness for large scale scenarios.

voir basilic : http://emotion.inrialpes.fr/bibemotion/2004/DBM04a/ address: New Orleans, LA (US); This paper deals with the domain of space modeling for mobile robotics. It offers a comparison of probabilistic and biomimetic models of navigation. Both approaches are shown to be quite complementary: while the probabilistic methods exploit sound theoretical grounds, they lack the modularity and, as a consequence, flexibility, of their biomimetic counterparts. We propose a new formalism, called the Bayesian Map formalism, that attempts to bridge the gap between the two domains: it is based on Bayesian modeling and inference for defining the building blocks, and uses operators for building hierarchies of models.

International audience; We propose a new method to program robots based on Bayesian inference and learning. It is called BRP for Bayesian Robot Programming. The capacities of this programming method are demonstrated through a succession of increasingly complex experiments. Starting from the learning of simple reactive behaviors, we present instances of behavior combinations, sensor fusion, hierarchical behavior composition, situation recognition and temporal sequencing. This series of experiments comprises the steps in the incremental development of a complex robot program. The advantages and drawbacks of BRP are discussed along with these different experiments and summed up as a conclusion. These different robotics programs may be seen as an illustration of probabilistic programming applicable whenever one must deal with problems based on uncertain or incomplete knowledge. The scope of possible applications is obviously much broader than robotics.

voir basilic : http://emotion.inrialpes.fr/bibemotion/2004/LBDM04a/; Cet article propose une mthode originale de programmation des robots fonde sur linfrence et lapprentissage baysien. Cette mthode traite formellement des problmes dincertitude et dincompltude inhrents au domaine considr. La principale difficult de la programmation des robots vient de linvitable incompltude des modles utiliss. Nous exposons le formalisme de description dune tche robotique ainsi que les mthodes de rsolution. Nous lillustrons en utilisant ce systme pour programmer une application de surveillance pour un robot mobile : le Khepera. Pour cela, nous utilisons des ressources gnriques de programmation appeles descriptions . Nous montrons comment dfinir et utiliser de manire incrmentale ces ressources (comportements ractifs, fusion capteur, reconnaissance de situations et squences de comportements) dans un cadre systmatique et unifi

voir basilic : http://emotion.inrialpes.fr/bibemotion/2003/DBM03a/ note: Int. Workshop on Service, Assistive and Personal Robots. Technical Challenges and Real World Application Perspectives address: Madrid (ES) editor: Armada, M. and {Gonzalez de Santos}, Pa; This paper deals with the probabilistic modeling of an environment that a robot has to navigate in. We use a method for the probabilistic modeling of space cal led the Bayesian Map formalism. This formalism al lows incremental building of models: we define the Super- position operator, which is a formal ly wel l-defined op- erator. We present first a syntactic version of this operator, and second, a version where the previously obtained model is refined and enriched by experimen- tal learning. In the resulting superposed map, loca- tions are the conjunction of underlying possible loca- tions, which al lows for more precise localization and more complex tasks. A theoretical example validates the concept, and hints at its usefulness for realistic robotic scenarios.

voir basilic : http://emotion.inrialpes.fr/bibemotion/2003/DBM03b/ address: Acapulco (MX); This paper presents a new method for probabilistic modelling of space, called the Bayesian Map for- malism. It offers a generalization of some com- mon approaches found in the literature, as it does not constrain the dependency structure of the prob- abilistic model. The formalism allows incremental building of hierarchies of models, by the use of the Abstraction Operator. In the resulting hierarchy, lo- calization in the high level model is based on prob- abilistic competition of the lower level models. Ex- perimental results validate the concept, and hint at its usefulness for large scale scenarios.

International audience; A prerequisite to the design of future Advanced Driver Assistance Systems for cars is a sensing sytem providing all the information required for high-level driving assistance tasks. Carsense is a European project whose purpose is to develop such a new sensing system. It will combine different sensors (laser, radar and video) and will rely on the fusion of the information coming from these sensors in order to achieve better accuracy, robustness and an increase of the information content. This paper demonstrates the interest of using probabilistic reasoning techniques to address this challenging multi-sensor data fusion problem. The approach used is called Bayesian Programming. It is a general approach based on an implementation of the Bayesian theory. It was introduced first to design robot control programs but its scope of application is much broader and it can be used whenever one has to deal with problems involving uncertain or incomplete knowledge.

International audience; We present in this paper a Bayesian CAD system for robotic applications. We address the problem of the propagation of geometric uncertainties, and how to take into account this propagation when solving inverse problems. The methodology used to represent and handle uncertainties using conditional probability distributions on the system's parameters and the sensor measurements is presented. It may be seen as a generalization of constraint-based approaches in which we explicitly model geometric uncertainties. Using this methodology, a constraint is represented by a probability distribution instead of a simple equality or inequality. Numerical algorithms used to apply this methodology are also described. Using an example, we show how to apply our approach by providing simulation results using our CAD system.

A computational approach to direct and generalized inverse model acquisition is presented. The approach is based on a proposed method to direct model acquisition from partial information. The method decomposes a hyper-space function in one variable functions, simplifying the learning problem. The acquired direct model is then implemented in a tree-like structure that can be used in the inverse sense without additional learning effort. The authors' approach is able to acquire complete models in hyper-spaces requiring only selected data focused in one-dimension sub-spaces, strongly reducing the data acquisition effort. The authors' approach is particularly interesting for applications in robotics. The acquisition of direct models in robotics frequently takes place in high dimension phase spaces. When traditional approximation methods are used, enormous data bases, containing the examples to be interpolated, are required.

International audience

This paper presents a very simple robotic experiment to illustrate how probabilistic reasoning may be used for sensory-motor systems. We will show how our robot may learn internal representations of its interactions with the environment, how it may predict the sensory result of a given action, how it may generate motor command to reach a wished sensory situation, how it may recognize different status and novel conditions and finally how it may behave consistently to explore and exploit simultaneously its environment.

International audience

We describe an implementation of a real time path planner for a robot arm with six degrees of freedom moving among dynamical obstacles. The planner is based on a novel technique called the Ariadne's Clew Algorithm. A brief description of this algorithm and parallel implementation of it are presented. Finally we analyze experiments made with this planner.

The goal of the work described in this paper is to build a path planner able to drive a robot in a dynamic environment where the obstacles are moving. In order to do so, we propose a method, called "Ariadne's clew algorithm", to build a global path planner based on the combination of two local planning algorithms: an EXPLORE algorithm and a SEARCH algorithm. The purpose of the EXPLORE algorithm is to collect information about the environment with an increasingly fine resolution by placing landmarks in the searched space. The goal of the SEARCH algorithm is to opportunistically check if the target can be easily reached from any given placed landmark. The Ariadne's clew algorithm is shown to be very fast in most cases allowing planning in dynamic environments. Hence, it is shown to be complete, which means that it is sure to find a path when one exists. Finally, we describe a massively parallel implementation of this algorithm.

Autonomous vehicles navigation implies that decisions are taken continuously based on a partial and uncertain knowledge of their environment. This is constrained by limited digital representations of the environment and uncertainty associated with the perception process. Further, it is difficult to predict the behaviour of the perceived entities. This is highlighted at crossroad intersections, where most road accidents occur. This thesis proposes a decision-making process that reasons with different types of uncertainties including the behaviour of the observed drivers to plan the vehicle motion.To understand the context and the behaviours of the observed driver's behaviour, a machine learning approach is proposed. The result is feed to the decision-making process to build probabilistic estimation of the environment. The vehicle motion is planned taking into account the effect this might have when the vehicle interact with other entities. Our approach rewards actions that promote interaction and reduces risk. The system behaviour is analysed by using a set of metrics derived from the scenario analysis as well as safety and operational constraints.To infer road context, Gaussian Processes are applied to learn motion patterns from simulated trajectories, which included the effect of vehicle interactions with other entities. The resulting patterns are segmented into areas and used to understand the behaviour of a vehicle approaching an intersection. Then, Random Forest Classifiers are applied to estimate the driver manoeuvre in each area. The dataset used for this training is built using data recorded from road trials and simulations. These classifiers infer lateral and longitudinal manoeuvre by extracting features from the vehicle trajectories. This approach shows that the road context improves the manoeuvre classification and that by mixing few real and simulated trajectories, it is possible to classify the manoeuvres of drivers arriving at an intersection.The decision-making process is built upon Partially Observable Markov Decision Process and uses the output of the manoeuvre understanding. This probabilistic framework reasons including perception and behaviour uncertainty in the environment models. These are used to predict the likely consequences of the autonomous vehicle actions on its immediate environment. However, the evaluation of all combinations of actions and state estimations is complex, therefore, an online solver is used to obtain an approximation of each action value. The vehicle actions are evaluated using a set of rewards, namely: collision risk, behavioural risk, comfort and traffic rules. A weighted sum of linear functions is used to balance each component of the reward function with respect to the vehicle distance to the intersection. It allows to adapt the behaviour of the automated vehicle to different scenarios.To validate the system performances and to determine causes of failure and success, Key Performance Indicators associated to the scenario are proposed. These are part of a generic testing architectures. The approach is applied to cross-cutting scenarios at road intersections, considered very complex and hazardous. Simulation techniques have been used to evaluate the proposed framework, to examine the largest number of scenarios and to be tested in safe conditions.The result of this thesis shows that it is possible to reason with other vehicle behaviour in the decision-making process while approaching a road intersection crossing. While classical methods fail to evaluate the system behaviour, the proposed validation method gives more insights. It will allow to test the system on the real road as well as using more advance data driven methods in the decision-making models.; La navigation pour les véhicules autonomes implique que des décisions soient prises en permanence, sur la base d'une connaissance partielle et incertaine de leur environnement. Ceci est limité par les représentations numériques de l'environnement et par l'incertitude associée au processus de perception. De plus, il est difficile de prédire le comportement des entités perçues, particulièrement aux intersections. Cette thèse propose un processus de prise de décision prenant en compte différents types d'incertitudes, notamment le comportement des conducteurs observés, pour planifier le mouvement du véhicule.Pour comprendre le contexte et les comportements des conducteurs observés, une approche par apprentissage automatique est proposée. Le résultat est utilisé par le processus de prise de décision pour construire une estimation probabiliste de l'environnement. Le mouvement du véhicule est planifié en tenant compte l'effet que celui-ci pourrait avoir lorsque le véhicule interagit avec d'autres entités. Notre approche récompense les actions favorisant l'interaction et réduisant les risques. Le comportement du système est analysé à l'aide d'un ensemble de métriques dérivées de l'analyse du scénario.Pour déduire le contexte routier, des processus gaussiens apprennent des modèles de mouvement à partir de trajectoires simulées, qui incluent l'effet des interactions entre les véhicules. Les modèles sont segmentés en zones et utilisés pour comprendre le comportement d'un véhicule approchant d'une intersection. Ensuite, des forêts d'arbres décisionnels sont appliqués pour estimer la manœuvre du conducteur dans chaque zone. L'ensemble de données utilisé pour l'entrainement est construit à l'aide de données enregistrées lors d'essais sur route et de simulations. Ces classificateurs infèrent les manœuvres latérales et longitudinales en extrayant des caractéristiques des trajectoires du véhicule. Cette approche montre que le contexte routier améliore la classification des manœuvres et qu'en mélangeant quelques trajectoires réelles et simulées, il est possible de différencier les manœuvres de conducteurs arrivant à une intersection.Le processus de prise de décision repose sur un Processus de décision markovien partiellement observable et les résultats de la compréhension de la manœuvre. Ce cadre probabiliste permet de prendre en compte l'incertitude de la perception et du comportement. Ceux-ci sont utilisés pour prédire les conséquences probables des actions des véhicules autonomes sur son environnement immédiat. L’évaluation de toutes les combinaisons d'actions et d'estimations d'état est complexe. Par conséquent, un solveur online est utilisé pour obtenir une approximation de la valeur de chaque action. Elles sont évaluées à l'aide d'un ensemble de récompenses, à savoir : risque de collision, risque comportemental, confort et règles de circulation. Une somme pondérée de fonctions linéaires est utilisée pour équilibrer chaque composant de la fonction de récompense par rapport à la distance du véhicule à l'intersection.Pour valider les performances du système et déterminer les causes d'échec et de succès, des indicateurs de performance associés au scénario sont proposés. Ceux-ci font partie d'une architecture de tests génériques pour la validation de système autonome. Des techniques de simulation ont été utilisées pour évaluer le cadre proposé, pour examiner le plus grand nombre de scénarios et pour être testé sans danger.Le résultat de cette thèse montre qu'il est possible de raisonner avec le comportement d'autres véhicules dans le processus de prise de décision à l'approche d'une intersection. Alors que les méthodes classiques ne permettent pas d’évaluer le comportement du système, la méthode de validation proposée donne davantage d’informations. Cela permettra de tester le système sur la route ouverte et d’analyser plus en profondeur des méthodes basées sur l’intelligence artificielle.

Computers are without doubt one of the most important invention of the last century, whose impact cannot be overestimated. Over the years they became powerfull, due to the constant optimization of their processors. With the growing need of computing power due to AI, processors have become faster then ever. However, since we are reaching the power wall, Moore’s law is coming to an end. Therefore, a young research community called rebooting computing is looking for alternative computation architectures. In this work, we propose to use stochastic computing to build architectures dedicated to Bayesian inference aiming low-power consumption. We develop two machines, namely the Bayesian machine (BM) and the Bayesian sampling machine (BSM). In this thesis, we look at two signal processing applications: Sound Source Localization (SSL) and Source Separation (SS). For SSL, we introduce three methods using the BM. The first one is working in the time-frequency domain and hence uses the Fourier transform. The second one is running entirely in the temporal domain. The third one is a multi-source localization approach based on the previous method. We present a technique to speed up the stochastic computation by a factor of up to 10 3 . Moreover, we designed an on-chip likelihoods computation mechanism to reduce the memory needs of our machine. Furthermore, we ran simulations and real world experiments to validate our methods. We made ASIC simulations to evaluate the power consumption. For the second problem, the source separation, we introduce a more general machine, the Bayesian sampling machine, which is based on the Gibbs sampling approach. We present a sampling method to solve source separation and run simulations to show the effectiveness of this technique.; L’ordinateur est sans aucun doute l’une des inventions les plus importantes du siècle dernier, dont l’impact ne peut être surestimé. Au fil des années, ils sont devenus de plus en plus puissants grâce à l’optimisation constante des processeurs. Avec un besoin croissant en puissance de calcul, et notamment à cause de l’IA, les processeurs sont devenus plus rapides que jamais. Cependant, à cause des limites physiques, la loi de Moore touche à sa fin. Par conséquent, il est nécessaire de proposer des alternatives. C’est le but de la communauté rebooting computing. Dans ce travail, nous nous proposons d’utiliser le calcul stochastique pour construire des architectures dédiées à l’inférence bayésienne visant une faible consommation d’énergie. Nous avons développé deux machines, à savoir la Bayesian machine (BM) et la Bayesian sampling machine (BSM). Dans cette thèse, nous nous intéresserons à deux applications de traitement du signal : la localisation de sources sonores (SSL) et la séparation de source. Pour la SSL, nous présentons trois méthodes utilisant la Bayesian machine. La première méthode fonctionne dans le domaine temps-fréquence, nécessitant le calcul de la transformée de Fourier. La deuxième est entièrement dans le domaine temporel. La troisième approche est une méthode de localisation multi-sources qui est basée sur la seconde. De plus, nous proposons une technique permettant d’accélérer le calcul stochastique d’un facteur 10 3 . Nous avons également développé une méthode de calcul des vraissemblances afin de réduire la mémoire de notre machine. Nous avons simulé les trois méthodes et fait des expérimentations en environnement réel. Nous présentons la consommation d’énergie obtenue via des simulations ASIC. Pour la seconde application, la séparation de source, nous introduisons une machine plus générale, la Bayesian sampling machine, qui est basée sur l’échantillonnage de Gibbs. Nous présentons une méthode basée sur l’échantillonnage pour séparer des source sonores. Cette méthode a été validée en simulation.

During the past few decades automakers have consistently introduced technological innovations aimed to make road vehicles safer. The level of sophistication of these advanced driver assistance systems has increased parallel to developments in sensor technology and embedded computing power. More recently, a lot of the research made both by industry and institutions has concentrated on achieving fully automated driving. The potential societal benefits of this technology are numerous, including safer roads, improved traffic flows, increased mobility for the elderly and the disabled, and optimized human productivity. However, before autonomous vehicles can be commercialized they should be able to safely share the road with human drivers. In other words, they should be capable of inferring the state and intentions of surrounding traffic from the raw data provided by a variety of onboard sensors, and to use this information to make safe navigation decisions. Moreover, in order to truly navigate safely they should also consider potential obstacles not observed by the sensors (such as occluded vehicles or pedestrians). Despite the apparent complexity of the task, humans are extremely good at predicting the development of traffic situations. After all, the actions of any traffic participant are constrained by the road network, by the traffic rules, and by a risk-aversive common sense. The lack of this ability to naturally understand a traffic scene constitutes perhaps the major challenge holding back the large-scale deployment of truly autonomous vehicles in the roads.In this thesis, we address the full pipeline from driver behavior modeling and inference to decision-making for navigation. In the first place, we model the behavior of a generic driver automatically from demonstrated driving data, avoiding thus the traditional hand-tuning of the model parameters. This model encodes the preferences of a driver with respect to the road network (e.g. preferred lane or speed) and also with respect to other road users (e.g. preferred distance to the leading vehicle). Secondly, we describe a method that exploits the learned model to predict the future sequence of actions of any driver in a traffic scene up to the distant future. This model-based prediction method assumes that all traffic participants behave in a risk-aware manner and can therefore fail to predict dangerous maneuvers or accidents. To be able to handle such cases, we propose a more sophisticated probabilistic model that estimates the state and intentions of surrounding traffic by combining the model-based prediction with the dynamic evidence provided by the sensors. In a way, the proposed model mimics the reasoning process of human drivers: we know what a given vehicle is likely to do given the situation (this is given by the model), but we closely monitor its dynamics to detect deviations from the expected behavior. In practice, combining both sources of information results in an increased robustness of the intention estimates in comparison with approaches relying only on dynamic evidence. Finally, the learned driver behavioral model and the prediction model are integrated within a probabilistic decision-making framework. The proposed methods are validated with real-world data collected with an instrumented vehicle. Although focused on highway environments, this work could be easily adapted to handle alternative traffic scenarios.; Au cours des dernières décennies, les constructeurs automobiles ont constamment introduit des innovations technologiques visant à rendre les véhicules plus sûrs. Le niveau de sophistication de ces systèmes avancés d’aide à la conduite s’est accru parallèlement aux progrès de la technologie des capteurs et de la puissance informatique intégrée. Plus récemment, une grande partie de la recherche effectuée par l'industrie et les institutions s'est concentrée sur l'obtention d'une conduite entièrement automatisée. Les avantages sociétaux potentiels de cette technologie sont nombreux, notamment des routes plus sûres, des flux de trafic améliorés et une mobilité accrue pour les personnes âgées et les handicapés. Toutefois, avant que les véhicules autonomes puissent être commercialisés, ils doivent pouvoir partager la route en toute sécurité avec d’autres véhicules conduits par des conducteurs humains. En d'autres termes, ils doivent pouvoir déduire l'état et les intentions du trafic environnant à partir des données brutes fournies par divers capteurs embarqués, et les utiliser afin de pouvoir prendre les bonnes décisions de conduite sécurisée. Malgré la complexité apparente de cette tâche, les conducteurs humains ont la capacité de prédire correctement l’évolution du trafic environnant dans la plupart des situations. Cette capacité de prédiction est rendu plus simple grâce aux règles imposées par le code de la route qui limitent le nombre d’hypothèses; elle repose aussi sur l’expérience du conducteur en matière d’évaluation et de réduction du risque. L'absence de cette capacité à comprendre naturellement une scène de trafic constitue peut-être, le principal défi qui freine le déploiement à grande échelle de véhicules véritablement autonomes sur les routes.Dans cette thèse, nous abordons les problèmes de modélisation du comportement du conducteur, d'inférence sur le comportement des autres véhicules, et de la prise de décision pour la navigation sûre. En premier lieu, nous modélisons automatiquement le comportement d'un conducteur générique à partir de données de conduite démontrées, évitant ainsi le réglage manuel traditionnel des paramètres du modèle. Ce modèle codant les préférences d’un conducteur par rapport au réseau routier (par exemple, voie ou vitesse préférées) et aux autres usagers de la route (par exemple, distance préférée au véhicule de devant). Deuxièmement, nous décrivons une méthode qui utilise le modèle appris pour prédire la séquence des actions à long terme de tout conducteur dans une scène de trafic. Cette méthode de prédiction suppose que tous les acteurs du trafic se comportent de manière aversive au risque, et donc ne peut pas prévoir les manœuvres dangereux ou les accidents. Pour pouvoir traiter de tels cas, nous proposons un modèle probabiliste plus sophistiqué, qui estime l'état et les intentions du trafic environnant en combinant la prédiction basée sur le modèle avec les preuves dynamiques fournies par les capteurs. Le modèle proposé imite ainsi en quelque sorte le processus de raisonnement des humains. Nous humains, savons ce qu’un véhicule est susceptible de faire compte tenu de la situation (ceci est donné par le modèle), mais nous surveillerons sa dynamique pour en détecter les écarts par rapport au comportement attendu. En pratique, la combinaison de ces deux sources d’informations se traduit par une robustesse accrue des estimations de l’intention par rapport aux approches reposant uniquement sur des preuves dynamiques. En dernière partie, les deux modèles présentés (comportemental et prédictif) sont intégrés dans le cadre d´une approche décisionnel probabiliste. Les méthodes proposées se sont vues évalués avec des données réelles collectées avec un véhicule instrumenté, attestant de leur efficacité dans le cadre de la conduite autonome sur autoroute. Bien que centré sur les autoroutes, ce travail pourrait être facilement adapté pour gérer des scénarios de trafic alternatifs.

Proteins are macromolecules participating in important biophysical processes of living organisms. It has been shown that changes in protein structures can lead to changes in their functions and are found linked to some diseases such as those related to neurodegenerative processes. Hence, an understanding of their structures and interactions with other molecules such as ligands is of major concern for the scientific community and the medical industry for inventing and assessing new drugs.In this dissertation, we are particularly interested in developing new methods to find for a system made of a single protein or a protein and a ligand, the pathways that allow changing from one state to another. During past decade, a vast amount of computational methods has been proposed to address this problem. However, these methods still have to face two challenges: the high dimensionality of the representation space, associated to the large number of atoms in these systems, and the complexity of the interactions between these atoms.This dissertation proposes two novel methods to efficiently find relevant pathways for such biomolecular systems. The methods are fast and their solutions can be used, analyzed or improved with more specialized methods. The first proposed method generates interpolation pathways for biomolecular systems using the As-Rigid-As-Possible (ARAP) principle from Computer Graphics. The method is robust and the generated solutions preserve at best the local rigidity of the original system. An energy-based extension of the method is also proposed, which significantly improves the solution paths. However, in scenarios requiring complex deformations, this geometric approach may still generate unnatural paths. Therefore, we propose a second method called ART-RRT, which combines the ARAP principle for reducing the dimensionality, with the Rapidly-exploring Random Trees from Robotics for efficiently exploring possible pathways. This method not only gives a variety of pathways in reasonable time but the pathways are also low-energy and clash-free, with the local rigidity preserved as much as possible. The mono-directional and bi-directional versions of the ART-RRT method were applied for finding ligand-unbinding and protein conformational transition pathways, respectively. The results are found to be in good agreement with experimental data and other state-of-the-art solutions., Les protéines sont des macromolécules participant à d’importants processus biophysiques de la vie des organismes. Or, il a été démontré que des variations de leur structure peuvent conduire à des changements de fonction en lien avec certaines maladies telles que celles associées à des processus neurodégénératifs. Ainsi, tant pour la communauté scientifique que pour l’industrie médicale, il est capital d’avoir une meilleure compréhension de la structure de ces protéines, ainsi que de leurs interactions avec d’autres molécules, ce en vue d’inventer et d’évaluer de nouveaux médicaments.Au cours de ces travaux de thèse, nous nous sommes particulièrement intéressés au développement de nouvelles méthodes de recherche de chemins biologiquement faisables entre deux états connus pour un système composé d’une protéine ou d’une protéine et d’un ligand. Au cours des dernières décennies, une grande quantité d’approches algorithmiques ont été proposées pour faire face à ce problème. Pourtant, les méthodes développées sont encore aujourd’hui confrontées à deux grands défis : d’une part la haute dimension des espaces de recherche, associée au grand nombre d’atomes impliqués, d’autre part la complexité des interactions entre ces atomes.Cette dissertation propose deux nouvelles méthodes pour obtenir de manière efficace des chemins pertinents pour des systèmes moléculaires. Ces méthodes sont rapides et génèrent des solutions qui peuvent ensuite être analysées ou améliorées à l’aide de méthodes d’avantage spécialisées. La première approche proposée produit des chemins d’interpolation pour systèmes biomoléculaires, à l’aide des approches dites aussi-rigides-que-possible, (ARAP) utilisées en animation graphique. Cette méthode est robuste et génère des solutions préservant au mieux la rigidité du système d’origine. Une extension de cette méthode basée sur des critères énergétiques a également été proposée et s’est avérée capable d’améliorer de manière significative les chemins solution. Cependant, pour les scénarios nécessitant de complexes déformations, cette approche géométrique peut conduire à des chemins solution non naturels. Nous avons donc proposé une seconde méthode appelée ART-RRT, qui utilise l’approche ARAP pour réduire la dimensionalité de l’espace et la combine avec les arbres d’exploration RRT (Rapidely-exploring Random Tree) issus de la Robotique, afin d’explorer efficacement les chemins possibles de l’espace.En plus de fournir une variété de solutions en temps raisonnable, cette ART-RRT produit des chemins de faible énergie, sans collision et dont la rigidité est préservée autant que possible. Des versions monodirectionnelles de bidirectionelles de cette méthode ont été proposées et appliquées respectivement à la recherche de chemin d’extraction d’un ligand hors du site actif d’une protéine, ainsi qu’a la recherche de chemin de transition conformationnelle pour protéine seule. Les solutions trouvées se sont avérées être en accord avec les données expérimentales ainsi qu’avec les solutions issues de l’état de l’art.