Search

International audience

From the robot’s point of view, a major issue in human-robot collaboration is how to be robust against uncertain human objectives, and uncertain human behaviors given a known objective. A key preliminary question is then : How to derive realistic human behaviors given a known objective ? Indeed, to allow for collaboration, such behaviors should also account for the robot behavior, while it is not known in the first place. In this paper, we rely on Markov decision models, representing the uncertainty over the human objective as a probability distribution over a finite set of reward functions (what will induce a distribution over human behaviors). Based on this, we propose two contributions : 1. a robot planning algorithm that is robust to the uncertain human behavior and relies on solving a POMDP obtained by reasoning on the distribution over human behaviors ; and 2. an approach to automatically generate an uncertain human behavior (a policy) for each provided reward function while accounting for the possible robot behavior. A co-working scenario allows conducting experiments and presenting qualitative and quantitative results to evaluate our approach., Du point de vue du robot, une difficulté majeure de la collaboration homme-robot est d'être robuste face à des objectifs incertains de l'humain, et aux comportements incertains étant donné un objectif connu. Une question préliminaire clef est alors : Comment dériver des comportements humains réalistes étant donné un objectif connu ? En effet, pour rendre la collaboration possible, de tels comportements devraient aussi tenir compte du comportement du robot, alors que celui-ci n'est justement pas connu. Dans cet article, nous nous appuyons sur des modèles de décision markoviens, représentant l'incertitude sur l'objectif de l'humain par une distribution de probabilité sur un ensemble fini de fonctions de récompense (ce qui induira une distribution sur des comportements humains). Sur cette base, nous proposons deux contributions : 1. un algorithme de planification pour le robot qui est robuste au comportement incertain de l'humain et repose sur la résolution d'un POMDP obtenu en raisonnant sur la distribution sur les comportements humains ; et 2. une approche pour générer automatiquement un comportement humain incertain (une politique) pour chaque fonction de récompense fournie tout en tenant compte du comportement possible du robot. Un scénario de travail collaboratif permet de mener des expérimentations et de présenter des résultats qualitatifs et quantitatifs pour évaluer notre approche.

International audience

International audience; We address the problem of human pose and posture estimation without any high precision marker-based motion capture systems, by merging inertial data from wearable sensors and a single RGB camera. Our proposition is based on a biomechanical model of the human body and two coupled filters: the first filter takes advantage of the accurate posture observations provided by wearable sensors and a factorization of joints to estimate the human posture with a reduced number of particles while the second filter uses RGB camera observations to estimate the drift of the wearable sensor so as to estimate the global state (pose and posture). In order to combine those filters, the estimated human posture distribution of the first filter is used as a proposal distribution for the second fusion filter so as to focus on particles with an already highlikelihood posture and to improve the efficiency of the pose estimation. Results showed this approach can perform online estimation of the human posture and the human pose (through the drift of the wearable sensor) and performed better than techniques relying only on inertial sensor or on direct pose estimation.

National audience; This paper looks at solving collaborative planning problems formalized as Decentralized POMDPs (DecPOMDPs) by searching for Nash equilibria, i.e., situations where each agent’s policy is a best response to the other agents’ (fixed) policies. While the joint equilibrium-based search for policies (JESP) algorithm does this in the finitehorizon setting relying on policy trees, we propose here to adapt JESP to infinite-horizon Dec-POMDPs by using Finite State Controller policy representations. In this article, we (1) explain how to turn a Dec-POMDP with N −1 fixed finite state controllers into an infinite-horizon POMDP whose solution is a best response of the Nth agent ; (2) propose a JESP variant based on this transformation, called inf-JESP, for solving infinite-horizon Dec-POMDPs ; (3) introduce heuristic initializations for JESP aiming at deterministically leading to good solutions ; and (4) conduct experiments on state-of-the-art benchmark problems to evaluate our approach.; Cet article s'intéresse à la résolution de problèmes de planification collaborative formalisés comme des POMDP décentralisés (Dec-POMDP) en cherchant des équilibres de Nash, c'est-à-dire des situations dans lesquelles la politique de chaque agent est une meilleure réponse aux politiques (fixes) des autres agents. Alors que l'algorithme joint equilibrium-based search for policies (JESP) fait ceci dans le cadre d'horizons finis en se reposant sur des arbres-politiques, nous proposons ici d'adapter JESP aux Dec-POMDP à horizon infini en représentant les politiques des agents par des contrôleurs à états finis. Dans cet article, nous (1) expliquons comment transformer un Dec-POMDP avec N − 1 contrôleurs à états finis fixés en un POMDP à horizon infini dont la solution est une meilleure réponse du N ième agent ; (2) proposons une variante de JESP, appelée inf-JESP, reposant sur cette transformation pour résoudre des Dec-POMDP à horizon infini ; (3) introduisons des initialisations heuristiques pour JESP visant à conduire à de bonnes solutions ; et (4) conduisons une évaluation empirique de notre approche sur des bancs d'essais de l'état de l'art.

International audience; We address the problem of recognizing the current activity performed by a human worker, providing an information useful for automatic ergonomic evaluation of workstations for industrial applications.Traditional ergonomic assessment methods rely on pen-and-paper worksheet, such as the Er-gonomic Assessment Worksheet (EAWS). Nowadays, there exists no tool to automatically estimate the ergonomics score from sensors (external cameras or wearable sensors). As the ergonomic evaluation depends of the activity that is being performed, the first step towards a fully automatic ergonomic assessment is to automatically identify the different activities within an industrial task. To address this problem, we propose a method based on wearable sensors and supervised learning based on Hidden Markov Model (HMM). The activity recognition module works in two steps. First, the parameters of the model are learned offline from observation based on both sensors, then in a second stage, the model can be used to recognize the activity offline and online. We apply our method to recognize the current activity of a worker during a series of tasks typical of the manufacturing industry. We recorded 6 participants performing a sequence of tasks using wearable sensors.Two systems were used: the MVN Link suit from Xsens and the e-glove from Emphasis Telematics (See Fig. 1). The first consists of 17 wireless inertial sensors embedded in a lycra suit, and is used to track the whole-body motion. The second is a glove that includes pressure sensors on fingertips, and finger flexion sensors. The motion capture data are combined with the one from the glove and fed to our activity recognition model. The tasks were designed to involve elements of EAWS such as load handling, screwing and manipulating objects while in different static postures. The data are labeled following the EAWS categories such as " standing bent forward " , " overhead work " or " kneeling ". In terms of performances, the model is able to recognize the activities related to EAWS with 91% of precision by using a small subset of features such as the vertical position of the center of mass, the velocity of the center of mass and the angle of the L5S1 joint.

International audience; In this paper, we address the problem of recognizing the current activity performed by a human operator, providing an information useful for automatic ergonomic evaluation for industrial applications. While the majority of research in activity recognition relies on cameras observing the human, here we explore the use of wearable sensors, which are more suitable in industrial environments. We use a wearable motion tracking suit and a sensorized glove. We describe our approach for activity recognition with a probabilistic model based on Hidden Markov Models, applied to the problem of recognizing elementary activities during a pick-and-place task inspired by a manufacturing scenario. We show that our model is able to correctly recognize the activities with 96% of precision if both sensors are used.

This document describes my research around Bayesian modeling and robotics. My work started with the modeling of biological processes before evolving towards robotics. In both cases, I was interested in both perception and action. I first proposed a model of human perception of planar surfaces with optic flow which fuses in a single framework two concurrent hypotheses of the literature. I also proposed and compared several models of eye movement selection in a Multiple Object Tracking task. I was able to show that the model with explicit uncertainty was the closest to the subjects eye movements.In robotics, I worked on the state estimation of several robots with classical filtering techniques but also including fusion of multiple sources of information of various nature and characteristics. I also discuss the Iterative Closest Point algorithm for which we proposed a more rigorous method for evaluating the different variants. The last piece of work I present deals with online three-dimensional path planning and execution of a tracked robot with significant climbing capabilities.I conclude this document with perspectives on what I call situated robotics, that is robots not taken in isolation but embedded in a sensorized environment shared with humans.; Ce document décrit mes travaux de recherche autour de la modélisation bayésienne et de la robotique. Mon travail a commencé par la modélisation de processus biologiques avant, dans un deuxième temps, d'évoluer vers la robotique. Dans les deux cas, je me suis intéressé à la fois à la perception et à l'action. J'ai donc proposé un modèle de la perception humaine de plans par le flux optique qui réunit deux hypothèses de la littérature dans un cadre unique. J'ai aussi proposé et comparé différents modèle de la sélection de mouvement oculaire dans une tâche de suivi multi-cibles, et montré que le modèle prenant en compte explicitement l'incertitude proposait des mouvements plus proches de ceux des sujets.Du côté robotique, j'ai travaillé sur l'estimation d'état de plusieurs robots avec des techniques classiques de filtrage mais en incluant la fusion de plusieurs sources d'informations de nature et caractéristiques différentes. Je discute aussi de l'algorithme d'Iterative Closest Point pour proposer une méthode plus rigoureuse d'évaluation des différentes variantes. Le dernier travail que je présente concerne la planification en ligne et l'exécution de chemin pour un robot à chenille avec des capacités de franchissement importantes.Je conclus ce document par des perspectives de travail sur ce que j'appelle la robotique située, c'est-à-dire des robots non plus isolés mais plongés dans un environnement équipé de capteurs et partagé avec des humains.

International audience; Urban Search and Rescue missions for mobile robots require reliable state estimation systems resilient to conditions given by the dynamically changing environment. We design and evaluate a data fusion system for localization of a mobile skid-steer robot intended for USAR missions. We exploit a rich sensor suite including both proprioceptive (inertial measurement unit and tracks odometry) and exteroceptive sensors (omnidirectional camera and rotating laser rangefinder). To cope with the specificities of each sensing modality (such as significantly differing sampling frequencies), we introduce a novel fusion scheme based on Extended Kalman filter for 6DOF orientation and position estimation. We demonstrate the performance on field tests of more than 4.4 km driven under standard USAR conditions. Part of our datasets include ground truth positioning; indoor with a Vicon motion capture system and outdoor with a Leica theodolite tracker. The overall median accuracy of localization—achieved by combining all the four modalities—was 1.2 % and 1.4 % of the total distance traveled, for indoor and outdoor environments respectively. To identify the true limits of the proposed data fusion we propose and employ a novel experimental evaluation procedure based on failure case scenarios. This way we address the common issues like: slippage, reduced camera field of view, limited laser rangefinder range, together with moving obstacles spoiling the metric map. We believe such characterization of the failure cases is a first step towards identifying the behavior of state estimation under such conditions. We release all our datasets to the robotics community for possible benchmarking.

International audience; A large variety of mobile robots has blossomed recently, made possible by progresses in energystorage, electronics, processing speed, etc. We have seen these robots performing exploration andnavigation tasks in different environments, and demonstrating impressive autonomy. However, theactual programming of complex behaviors, in particular when physical interaction takes place be-tween the robot and its environment, has remained tedious. This work often involves writing andtuning complicated state machines and requires a lot of engineering resources. Machine learning,which aims at specifying complex algorithms by leveraging data instead of just expert knowledge,proposes an alternative path. It appears especially useful for robotic tasks, as these often deal withhigh-dimensional input and output spaces while relying on a lot of contingent parameter values.Our aim is to understand the limits of capabilities that can be achieved with a simple, non-parametricsystem that learns from demonstration. We want to reduce the number of meta parameters to theminimum, because these ultimately have to be given to the system, while still handling tasks con-sisting of several steps. This paper proposes a system with only ns+ 2 meta parameters, where nsis the number of sensor dimensions. All these parameters are related to the robotic platform and itsapplication, but do not depend on the task. We believe that such a system, should it perform wellenough, would be of tremendous help for developing and deploying robotic applications. Indeed,the current requirement of choosing by hand many parameters is a major obstacle to the deploymentof programming by demonstration.

International audience; In this paper we examine the case of a mobile robot that is part of a human-robot urban search and rescue (USAR) team. During USAR scenarios, we would like the robot to have a geometrical-functional understanding of space, using which it can infer where to perform planned tasks in a manner that mimics human behavior. We assess the situation awareness of rescue workers during a simulated USAR scenario and use this as an empirical basis to build our robot's spatial model. Based upon this spatial model, we present " functional mapping " as an approach to identify regions in the USAR environment where planned tasks are likely to be optimally achievable. The system is deployed and evaluated in a simulated rescue scenario.

Perception can be seen as collecting and confronting various pieces of information in order to understand the environment. One uses many semsory modalities each one consisting of many kind of information. One of these is the optic flow, the displacement of the projected image on the retina.Optic flow allows for the extraction of part of the geometry of the observed scene. However, it is a very complex problem due to its inverse, ill-posed and uncertain nature.These issues can be dealt with using the Bayesian programming formalism. It is based on the probabilities, to handle uncertainty and the ill-posed problem, and on Bayes rule, to solve the inversion.We propose a Bayesian model of the perception of planes from optic flow based on few explicit assumptions. These assumptions are mainly the rigidity, which assumes the observation of a rigid plane is more probable than of a collection of independently-moving points, and the stationarity, which assumes the motion of the plane is more probably small.We validate our model using five experiments from the literature. For each one, we explain in details the inner workings of our model and the assumptions conditionning a good reproduction of the experimental results. Finally, we show how to adapt our model to an experiment of perception of a corrgated surface by optic flow. We also present a Bayesian model to generate the input of our perception model and generalise the integration of our model into a more complete system., La perception est le regroupement d'informations en vue de comprendre son environnement. Pour cela, l'homme utilise plusieurs modalités sensorielles consistuées de plusieurs types d'information. L'une d'entre elles est le flux optique, le déplacement de l'image sur la rétine.Le flux optique permet de recouvrer une partie de la géometrie de la scène observée, mais il s'agit d'un problème très complexe car inverse, mal-posé et incertain.Ces difficultés peuvent être surmontées à l'aide du formalisme de la programmation bayésienne qui s'appuie sur l'usage des probabilités, pour manipuler l'incertitude et le caractère mal-posé, et de la règle de Bayes pour résoudre l'inversion.Nous proposons un modèle bayésien de la perception de plans par le flux optique basé sur quelques hypothèses explicites. Ces hypothèses sont principalement la rigidité, qui considère que l'on observe plus probablement un plan rigide qu'une collection de points de mouvements indépendants et la sationnarité, qui suppose que le mouvement du plan est plus probablement petit.Nous validons notre modèle en nous basant sur cinq expériences issues de la littérature. Pour chacune d'elles, nous expliquons les détails du fonctionnement du modèle et les hypothèses qui conditionnent la bonne reproduction des résultats. Enfin, nous montrons d'une part comment intégrer notre modele au sein de systèmes plus complexes et, d'autre part, l'applicabilité du modèle bayésien à des expériences de perception du flux optique différentes de part leurs protocoles, mais similaires dans leurs ambiguités.

International audience; The paper describes experience with applying a user-centric design methodology in developing systems for human-robot teaming in Urban Search & Rescue. A human-robot team consists of several semi-autonomous robots (rovers/UGVs, microcopter/UAVs), several humans at an off-site command post (mission commander, UGV operators) and one on-site human (UAV operator). This system has been developed in close cooperation with several rescue organizations, and has been deployed in a real-life tunnel accident use case. The human-robot team jointly explores an accident site, communicating using a multi-modal team interface, and spoken dialogue. The paper describes the development of this complex socio-technical system per se, as well as recent experience in evaluating the performance of this system.