The goal of this work is to build a cognitive system for the humanoid robot, Cog, that exploits human caregivers as catalysts to perceive and learn about actions, objects, scenes, people, and the robot itself. This thesis addresses a broad spectrum of machine learning problems across several categorization levels. Actions by embodied agents are used to automatically generate training data for the learning mechanisms, so that the robot develops categorization autonomously. Taking inspiration from the human brain, a framework of algorithms and methodologies was implemented to emulate different cognitive capabilities on the humanoid robot Cog. This framework is effectively applied to a collection of AI, computer vision, and signal processing problems. Cognitive capabilities of the humanoid robot are developmentally created, starting from infant-like abilities for detecting, segmenting, and recognizing percepts over multiple sensing modalities. Human caregivers provide a helping hand for communicating such information to the robot. This is done by actions that create meaningful events (by changing the world in which the robot is situated) thus inducing the "compliant perception" of objects from these human-robot interactions. Self-exploration of the world extends the robot's knowledge concerning object properties.This thesis argues for enculturating humanoid robots using infant development as a metaphor for building a humanoid robot's cognitive abilities. A human caregiver redesigns a humanoid's brain by teaching the humanoid robot as she would teach a child, using children's learning aids such as books, drawing boards, or other cognitive artifacts. Multi-modal object properties are learned using these tools and inserted into several recognition schemes, which are then applied to developmentally acquire new object representations. The humanoid robot therefore sees the world through the caregiver's eyes.Building an artificial humanoid robot's brain, even at an infant's cognitive level, has been a long quest which still lies only in the realm of our imagination. Our efforts towards such a dimly imaginable task are developed according to two alternate and complementary views: cognitive and developmental.

Synthesizing the realistic motion of a humanoid is a very sophisticated task, studied in different research areas. This work addresses the problem to synthetize realistic animations of 3D biped characters in a simulated environment, using genetic algorithms. Characters are represented as a structure of rigid bodies linked each other by 1DOF joints. Such joints are controlled by sinusoidal functions whose parameters are calculated by the genetic algorithm. Results, obtained by testing and comparing several different genetic operators, are presented. The system we have created allows the non-skilled user, to automatically create animations by setting only few key-poses of the characters.

Le chapitre 1 est pour l'essentiel une brève introduction générale qui donne le contexte générale de la planification et présente l'organisation du document dans son ensemble et quelques uns des points clés retenus : robot humanoïde, environnement non statique, perception par vision artificielle, et représentation de cet environnement par grilles d'occupation. Dans le chapitre 2, après une revue de littérature bien menée, l'auteur propose de considérer les points de repère de l'environnement dès la phase de planification de chemin afin de rendre plus robuste l'exécution des déplacements en cas d'évolution de l'environnement entre le moment où la planification est menée et celui où le robot se déplace ( évolution étant entendu comme liée à une amélioration de la connaissance par mise à jour, ou due à un changement de l'environnement lui-même). Le concept est décrit et une formalisation proposée. Le chapitre 3 s'intéresse en détail à la planification dans le cas d'environnements dynamiques. Les méthodes existantes, nombreuses, sont tout d'abord analysées et bien présentées. Le choix est fait ici de décrire l'environnement comme étant décomposé en cellules, regroupant elles-mêmes des voxels, éléments atomiques de la représentation. L'environnement étant changeant, l'auteur propose de réévaluer le plan préétabli à partir d'une bonne détection de la zone qui a pu se trouver modifiée dans l'environnement. L'approche est validée expérimentalement en utilisant une des plateformes robotiques du LAAS qui dispose de bonnes capacités de localisation : le manipulateur mobile Jido étant à ce jour plus performant sur ce plan que l'humanoïde HRP2, c'est lui qui a été utilisé. Ces expérimentations donnent des indications concordantes sur l'efficacité de l'approche retenue. Notons également que la planification s'appuie sur une boite englobante de l'humanoïde, et non pas sur une représentation plus riche (multi-degré-deliberté). En revanche, c'est bien de planification pour l'humanoïde considéré dans toute sa complexité qu'il s'agit au chapitre 4 : on s'intéresse ici à tous les degrés de liberté du robot. L'auteur propose des évolutions de méthodes existantes et en particulier sur la manière de tirer profit de la redondance cinématique. L'approche est bien décrite et permet d'inclure une phase d'optimisation de la posture globale du robot. Des exemples illustrent le propos et sont l'occasion de comparaison avec d'autres méthodes. Le chapitre 5 s'intéresse à la manière de modéliser l'environnement, sachant qu'on s'intéresse ici au cas d'une perception par vision artificielle, et précisément au cas de l'humanoïde, robot d'assurer lui-même cette perception au fur et à mesure de son avancée dans l'environnement. On est donc dans le cadre de la recherche de la meilleure vue suivante qui doit permettre d'enrichir au mieux la connaissance qu'a le robot de son environnement. L'approche retenue fait à nouveau appel à la boite englobante de l'humanoïde et non à sa représentation complète ; il sera intéressant de voir dans le futur ce que pourrait apporter la prise en compte des degrés de liberté de la tête ou du torse à la résolution de ce problème. Le chapitre 6 décrit la phase d'intégration de tous ces travaux sur la plateforme HRP2 du LAAS-CNRS, partie importante de tout travail de roboticien. / This thesis starts by proposing a new framework for motion planning using stochastic maps, such as occupancy-grid maps. In autonomous robotics applications, the robot's map of the environment is typically constructed online, using techniques from SLAM. These methods can construct a dense map of the environment, or a sparse map that contains a set of identifiable landmarks. In this situation, path planning would be performed using the dense map, and the path would be executed in a sensor-based fashion, using feedback control to track the reference path based on sensor information regarding landmark position. Maximum-likelihood estimation techniques are used to model the sensing process as well as to estimate the most likely nominal path that will be followed by the robot during execution of the plan. The proposed approach is potentially a practical way to plan under the specific sorts of uncertainty confronted by a humanoid robot. The next chapter, presents methods for constructing free paths in dynamic environments. The chapter begins with a comprehensive review of past methods, ranging from modifying sampling-based methods for the dynamic obstacle problem, to methods that were specifically designed for this problem. The thesis proposes to adapt a method reported originally by Leven et al.. so that it can be used to plan paths for humanoid robots in dynamic environments. The basic idea of this method is to construct a mapping from voxels in a discretized representation of the workspace to vertices and arcs in a configuration space network built using sampling-based planning methods. When an obstacle intersects a voxel in the workspace, the corresponding nodes and arcs in the configuration space roadmap are marked as invalid. The part of the network that remains comprises the set of valid candidate paths. The specific approach described here extends previous work by imposing a two-level hierarchical structure on the representation of the workspace. The methods described in Chapters 2 and 3 essentially deal with low-dimensional problems (e.g., moving a bounding box). The reduction in dimensionality is essential, since the path planning problem confronted in these chapters is complicated by uncertainty and dynamic obstacles, respectively. Chapter 4 addresses the problem of planning the full motion of a humanoid robot (whole-body task planning). The approach presented here is essentially a four-step approach. First, multiple viable goal configurations are generated using a local task solver, and these are used in a classical path planning approach with one initial condition and multiple goals. This classical problem is solved using an RRT-based method. Once a path is found, optimization methods are applied to the goal posture. Finally, classic path optimization algorithms are applied to the solution path and posture optimization. The fifth chapter describes algorithms for building a representation of the environment using stereo vision as the sensing modality. Such algorithms are necessary components of the autonomous system proposed in the first chapter of the thesis. A simple occupancy-grid based method is proposed, in which each voxel in the grid is assigned a number indicating the probability that it is occupied. The representation is updated during execution based on values received from the sensing system. The sensor model used is a simple Gaussian observation model in which measured distance is assumed to be true distance plus additive Gaussian noise. Sequential Bayes updating is then used to incrementally update occupancy values as new measurements are received. Finally, chapter 6 provides some details about the overall system architecture, and in particular, about those components of the architecture that have been taken from existing software (and therefore, do not themselves represent contributions of the thesis). Several software systems are described, including GIK, WorldModelGrid3D, HppDynamicObstacle, and GenoM.

This dissertation presents the results of a series of studies on dynamic stabilisation of CoMan, which is actuated by series elastic actuators. The main goal of this dissertation is to dynamically stabilise the humanoid robot on the floor by the simplest multivariate feedback control for the purpose of walking. The multivariable scheme is chosen to take into account the joints' interactions, as well as providing a systematic way of designing the feedback system to improve the bandwidth and tracking performance of CoMan's existing PID control. A detailed model is derived which includes all the motors and joints state variables and their multibody interactions which are often ignored in the previous studies on bipedal robots in the literature. The derived dynamic model is then used to design multivariable optimal control feedback and observers with a mathematical proof for the relative stability and robustness of the closed loop system in face of model uncertainties and disturbances. In addition, two decentralized optimal feedback design algorithms are presented that explicitly take the compliant dynamics and the multibody interactions into account while providing the mathematical proof for the stability of the overall system. The purpose of the proposed decentralized control methods is to provide a systematic model based PDPID design to replace the existing PID controllers which are derived by a trial and error process. Moreover, the challenging constrained and compliant motion of the robot in double support is studied where a novel constrained feedback design is proposed which directly takes the compliance dynamics, interactions and the constraints into account to provide a closed loop feedback tracking system that drives the robot inside the constrained subspace. This method of control is particularly interesting since most control methods applied to closed kinematic chains (such as the double support phase) are over complicated for implementation purposes or have an ad-hoc approach to controller design. In terms of walking trajectory generation, an extension to the ZMP walking trajectory generation is proposed to utilise the CoMan's upper body to tackle the non-minimum phase behaviour that is faced in trajectory generation. Simple inverted pendulum models of walking are then used to study the maximum feasible walking speed and step size where parameters of CoMan are used to provide numerical upperbounds on the step size and walking speed. Use of straight knee and toe push-off during walking is shown to be beneficial for taking larger step lengths and hence achieving faster walking speeds. Subsequently, the designed tracking systems are then applied to a dynamic walking simulator which is developed during this PhD project to accurately model the compliant walking behaviour of the CoMan. A walking gait is simulated and visualized to show the effectiveness of the developed walking simulator. Moreover, the experimental results and challenges faced during the implementation of the designed tracking control systems are discussed where it is shown that the LQR feedback results in 50% less control effort and tracking errors in comparison with CoMan's existing independent PID control. This advantage directly affects the feasible walking speed. In addition, a set of standard and repeatable tests for CoMan are designed to quantify and compare the performance of various control system designs. Finally, the conclusions and future directions are pointed out.

Afin que les robots humanoïdes puissent travailler avec les humains et être en mesure de résoudre des tâches répétitives, nous devons leur permettre de planifier leurs mouvements de façon autonome. La planification de mouvement est un problème de longue date en robotique, et tandis que sa fondation algorithmique a été étudiée en profondeur, la planification de mouvement est encore un problème NP-difficile et qui manque de solutions efficaces. Nous souhaitons ouvrir une nouvelle perspective sur le problème en mettant en évidence sa structure: le comportement du robot, le système mécanique du robot et l’environnement du robot. Nous allons nous intéresser à l’hypothèse que chaque composante structurelle peut être exploitée pour créer des algorithmes de planification de mouvement plus efficaces. Nous présentons trois algorithmes exploitant la structure, basés sur des arguments géométriques et topologiques: d’abord, nous exploitons le comportement d’un robot de marche en étudiant la faisabilité des transitions des traces de pas. L’algorithme qui en résulte est capable de planifier des traces de pas tout en évitant jusqu’à 60 objets situés sur une surface plane 6 mètres carrés. Deuxièmement, nous exploitons le système mécanique d’un robot humanoïde en étudiant les structures des liaisons linéaires de ses bras et de ses jambes. Nous introduisons le concept d’une trajectoire irréductible, qui est une technique de réduction de dimension préservant la complétude. L’algorithme résultant est capable de trouver des mouvements dans des environnements étroits, où les méthodes d’échantillonnage ne pouvaient pas être appliquées. Troisièmement, nous exploitons l’environnement en raisonnant sur la structure topologique des transitions de contact. Nous montrons que l’analyse de l’environnement est une méthode efficace pour pré-calculer les informations pertinentes pour une planification de mouvement efficace. En s’appuyant sur ces résultats, nous arrivons à la conclusion que l’exploitation de la structure est une composante essentielle de la planification de mouvement efficace. Il en résulte que tout robot humanoïde, qui veut agir efficacement dans le monde réel, doit être capable de comprendre et d’exploiter la structure. / If humanoid robots should work along with humans and should be able to solve repetitive tasks, we need to enable them with a skill to autonomously plan motions. Motion planning is a longstanding core problem in robotics, and while its algorithmic foundation has been studied in depth, motion planning is still an NP-hard problem lacking efficient solutions. We want to open up a new perspective on the problem by highlighting its structure: the behavior of the robot, the mechanical system of the robot, and the environment of the robot. We will investigate the hypothesis that each structural component can be exploited to create more efficient motion planning algorithms. We present three algorithms exploiting structure, based on geometrical and topological arguments: first, we exploit the behavior of a walking robot by studying the feasibility of footstep transitions. The resulting algorithm is able to plan footsteps avoiding up to 60 objects on a 6 square meters planar surface. Second, we exploit the mechanical system of a humanoid robot by studying the linear linkage structures of its arms and legs. We introduce the concept of an irreducible motion, which is a completeness-preserving dimensionality reduction technique. The resulting algorithm is able to find motions in narrow environments, where previous sampling-based methods could not be applied. Third, we exploit the environment by reasoning about the topological structure of contact transitions. We show that analyzing the environment is an efficient method to precompute relevant information for efficient motion planning. Based on those results, we come to the conclusion that exploiting structure is an essential component of efficient motion planning. It follows that any humanoid robot, who wants to act efficiently in the real world, needs to be able to understand and to exploit structure.

Cette thèse traite de l'estimation et de la stabilisation de l'état des compliances passives présentes dans les chevilles du robot humanoïde HRP-2. Ces compliances peuvent être vues comme un degré de liberté unique et observable, sous quelques hypothèses qui sont explicitées. L'estimateur utilise des mesures provenant de la centrale inertielle située dans le torse du robot et éventuellement des capteurs de forces situés dans ses pieds. Un filtre de Kalman étendu est utilisé pour l'estimation d'état. Ce filtre utilise un modèle complet de la dynamique du robot, pour lequel la dynamique interne du robot, considérée comme parfaitement connue et contrôlée, a été découplée de la dynamique de la compliance passive du robot. L'observabilité locale de l'état a été montrée en considérant ce modèle et les mesures provenant de la centrale inertielle seule. Il a de plus été montré que l'ajout des mesures des capteurs de forces dans les pieds du robot permet de compléter l'état avec des mesures d'erreurs dans le modèle dynamique du robot. L'estimateur a été validé expérimentalement sur le robot humanoïde HRP-2. Sur cet estimateur a été construit un stabilisateur de l'état de la compliance d'HRP-2. L'état commandé est la position et vitesse du centre de masse (contrôle indirecte de la quantité de mouvement) du robot, l'orientation et la vitesse angulaire de son tronc (contrôle indirecte du moment cinétique), ainsi que l'orientation et la vitesse angulaire de la compliance. Les grandeurs de commande sont l'accélération du centre de masse du robot et l'accélération angulaire de son tronc. Un régulateur quadratique linéaire (LQR) a été utilisé pour calculer les gains du retour d'état, basé sur un modèle appelé "pendule inverse flexible à roue d'inertie" qui consiste en un pendule inverse dont la base est flexible et où une répartition de masse en rotation autour du centre de masse du robot représente le tronc du robot. Des tests ont été effectués sur le robot HRP-2 en double support, utilisant l'estimateur décrit précédemment avec ou sans les capteurs de forces. / This PhD thesis covers the estimation and stabilization of the passive compliances state wich are located in the HRP-2 humanoid robot ankles. These compliances can be seen as a unique compliance under some assumptions that are presented. The estimator uses measurements coming from an Inertial Measurement Unit (IMU) located in the robot's chest. It also uses measurements coming from forces sensors located in its feet. An Extended Kalman Filter (EKF) is used for state estimation. This filter uses a complete model of the robot dynamics, in which the internal dynamics of the robot, considered as known, is decoupled from the dynamics of its passive compliance. The local observability of the state is shown by considering this model and the measurements coming from the IMU only. Furthermore, it has been shown that, by adding the measurements coming from the forces sensors in the robot's feet, we are able to complete the state with some errors measurements in the dynamical model of the robot. The estimator was validated experimentaly on the HRP-2 humanoid robot. Based on this estimator, a stabilizer of the compliance state of the HRP-2 robot was build. The control state is the position and velocity of the center of mass of the robot, the orientation and angular velocity of its trunk, and the orientation and the angular velocity of the compliance. The control values are the acceleration of the robot's center of mass and the angular acceleration of its trunk. A Linear Quadratic Regulator (LQR) is used to compute the feed-back gains, based on a Viscoelastic Reaction Mass Pendulum model which consist in an inverse pendulum whith a flexible base and where a mass repartition rotating around the center of mass is modeling the robot's trunk. Some tests were made on the HRP-2 robot in double support, using the previous estimator with and without the use of forces sensors

No description available.

Mimicking human motion with a humanoid robot can prove to be useful for studying gaits, designing better prostheses, or assisting the elderly or disabled. Directly mimicking and implementing a motion of a human on a humanoid robot may not be successful because of the different dynamic characteristics between them, which may cause the robot to fall down due to instability. Using the Zero Moment Point as the stability criteria, this work proposes an Analytical Motion Filter (AMF), which stabilizes a reference motion that can come from human motion capture data, gait synthesis using kinematics, or animation software, while satisfying common constraints. In order to determine how the AMF stabilized a motion, the different kinds of instabilities were identified and classified when examining the reference motions. The different cases of instability gave more insight as to why a particular motion was unstable: the motion was too fast, too slow, or inherently unstable. In order to stabilize the gait two primary methods were utilized: time and spatial scaling. Spatial scaling scaled the COM trajectory down towards a known stable trajectory. Time scaling worked similarly by changing the speed of the motion, but was limited in effectiveness based on the types of instabilities in the motion and the coupling of the spatial directions. Other constraints applied to the AMF and combinations of the different methods produced interesting results that gave more insight into the stability of the gait. The AMF was tested using both simulations and physical experiments using the DARwIn miniature humanoid robot developed by RoMeLa at Virginia Tech as the test platform. The simulations proved successful and provided more insight to understanding instabilities that can occur for different gait generation methods. The physical experiments worked well for non-walking motions, but because of insufficient controllability in the joint actuators of the humanoid robot used for the experiment, the high loads during walking motions prevented them from proper testing. The algorithms used in this work could also be expanded to legged robots or entirely different platforms that depend on stability and can use the ZMP as a stability criterion. One of the primary contributions of this work was showing that an entire reference motion could be stabilized using a single set of closed form solutions and equations. Previous work by others considered optimization functions and numeric schemes to stabilize all or a portion of a gait. Instead, the Analytical Motion Filter gives a direct relationship between the input reference motion and the resulting filtered output motion. / Ph. D.

This study focuses on the design and implementation of a humanoid robot for disaster response. In particular, this thesis investigates the lower body design in detail with the upper body discussed at a higher level. The Tactical Hazardous Operations Robot (THOR) was designed to compete in the DARPA Robotics Challenge where it needs to complete tasks based on first-responder operations. These tasks, ranging from traversing rough terrain through driving a utility vehicle, suggest a versatile platform in a human sized form factor. A physical experiment of the proposed tasks generated a set of joint range of motions (RoM). Desired limb lengths were determined by comparing existing robots, the test subject in the experiment of proposed tasks, and an average human. Simulations using the desired RoM and limb lengths were used to calculate baseline joint torques. Based on the generated design constraints, THOR is a 34 degree of freedom humanoid that stands 1.78 [m] tall and weighs 65 [kg]. The 12 lower body joints are driven by series elastic linear actuators with multiple joints actuated in parallel. The parallel actuation mimics the human body, where multiple muscles pull on the same joint cooperatively. The legs retain high joint torques throughout their large RoM with some joints achieving torques as high as 289 [Nm]. The upper body uses traditional rotary actuators to drive the waist, arms, and head. The proprioceptive sensor selection was influenced by past experience on humanoid platforms, and perception sensors were selected to match the competition. / Master of Science

As robotic systems grow in complexity, they inevitably undergo a process of specialization whereby they separate into an array of interconnected subsystems and individual processes. In order to function as a unified system, these processes rely heavily on interprocess communications (IPC) to transfer information between subsystems and various execution loops. This thesis presents the design, implementation, and validation of the Valor ROS Controller, a hybrid IPC interface layer and robot controller. The Valor ROS Controller connects the motion control system, implemented with the internally created Bifrost IPC, developed by Team VALOR for the DARPA Robotics Challenge (DRC) with the high level software developed by Team ViGIR that uses the Robot Operating System (ROS) IPC framework. The Valor ROS Controller also acts as a robot controller designed to run on THOR and ESCHER, and is configurable to use different control modes and controller implementations. By combining an IPC interface layer with controllers, the Valor ROS Controller enabled Team VALOR to use Team ViGIR's software capabilities at the DRC Finals. In addition to the qualitative validation of Team VALOR competing at the DRC Finals, this thesis studies the efficiency of the Valor ROS Controller by quantifying its computational resourceful utilization, message pathway latency, and joint controller tracking. Another contribution of this thesis is the quantification of end-effector pose error incurred by whole-body motions. This phenomenon has been observed on both THOR and ESCHER as one of their arms moves through a trajectory, however, it has never been studied in depth on either robot. The results demonstrate that the Valor ROS Controller adequately uses computational resources and has message latencies in the order of 50 ms. The results also indicate several avenues to improve arm tracking in Team VALOR's system. Whole-body motions account for approximately 5 cm of the end-effector pose error observed on hardware when an arm is at near full extension. / Master of Science