A central goal of the robotics community is to develop general optimization algorithms for producing high-performance dynamic behaviors in robot systems. This goal is challenging because many robot control tasks are characterized by significant stochasticity, high-dimensionality, expensive evaluations, and unknown or unreliable system models. Despite these challenges, a range of algorithms exist for performing efficient optimization of parameterized control policies with respect to average cost criteria. However, other statistics of the cost may also be important. In particular, for many stochastic control problems, it can be advantageous to select policies based not only on their average cost, but also their variance (or risk). In this thesis, I present new efficient global and local risk-sensitive stochastic optimization algorithms suitable for performing policy search in a wide variety of problems of interest to robotics researchers. These algorithms exploit new techniques in nonparameteric heteroscedastic regression to directly model the policy-dependent distribution of cost. For local search, learned cost models can be used as critics for performing risk-sensitive gradient descent. Alternatively, decision-theoretic criteria can be applied to globally select policies to balance exploration and exploitation in a principled way, or to perform greedy minimization with respect to various risk-sensitive criteria. This separation of learning and policy selection permits variable risk control, where risk sensitivity can be flexibly adjusted and appropriate policies can be selected at runtime without requiring additional policy executions. To evaluate these algorithms and highlight the importance of risk in dynamic control tasks, I describe several experiments with the UMass uBot-5 that include learning dynamic arm motions to stabilize after large impacts, lifting heavy objects while balancing, and developing safe fall bracing behaviors. The results of these experiments suggest that the ability to select policies based on risk-sensitive criteria can lead to greater flexibility in dynamic behavior generation.

<p> Traditionally, models of a robot's kinematics and sensors have been provided by designers through manual processes. Such models are used for sensorimotor tasks, such as manipulation and stereo vision. However, these techniques often yield static models based on one-time calibrations or ideal engineering drawings; models that often fail to represent the actual hardware, or in which individual unimodal models, such as those describing kinematics and vision, may disagree with each other.</p><p> Humans, on the other hand, are not so limited. One of the earliest forms of self-knowledge learned during infancy is knowledge of the body and senses. Infants learn about their bodies and senses through the experience of using them in conjunction with each other. Inspired by this early form of self-awareness, the research presented in this thesis attempts to enable robots to learn unified models of themselves through data sampled during operation. In the presented experiments, an upper torso humanoid robot, Nico, creates a highly-accurate self-representation through data sampled by its sensors while it operates. The power of this model is demonstrated through a novel robot vision task in which the robot infers the visual perspective representing reflections in a mirror by watching its own motion reflected therein.</p><p> In order to construct this self-model, the robot first infers the kinematic parameters describing its arm. This is first demonstrated using an external motion capture system, then implemented in the robot's stereo vision system. In a process inspired by infant development, the robot then mutually refines its kinematic and stereo vision calibrations, using its kinematic structure as the invariant against which the system is calibrated. The product of this procedure is a very precise mutual calibration between these two, traditionally separate, models, producing a single, unified self-model.</p><p> The robot then uses this self-model to perform a unique vision task. Knowledge of its body and senses enable the robot to infer the position of a mirror placed in its environment. From this, an estimate of the visual perspective describing reflections in the mirror is computed, which is subsequently refined over the expected position of images of the robot's end-effector as reflected in the mirror, and their real-world, imaged counterparts. The computed visual perspective enables the robot to use the mirror as an instrument for spacial reasoning, by viewing the world from its perspective. This test utilizes knowledge that the robot has inferred about itself through experience, and approximates tests of mirror use that are used as a benchmark of self-awareness in human infants and animals.</p>

Thèse (M.Sc.A.)--Université de Sherbrooke (Canada), 2007. / Titre de l'écran-titre (visionné le 1 février 2007). In ProQuest dissertations and theses. Publié aussi en version papier.

Thesis (Ph.D.)--University of Illinois at Urbana-Champaign, 2008. / Source: Dissertation Abstracts International, Volume: 69-11, Section: B, page: 6922. Adviser: Jean Ponce. Includes bibliographical references (leaves 127-134) Available on microfilm from Pro Quest Information and Learning.

Thesis (M.S.)--University of Nevada, Reno, 2006. / "December 2006." Includes bibliographical references (leaves 42-44). Online version available on the World Wide Web.

Thesis (Ph.D)--Electrical and Computer Engineering, Georgia Institute of Technology, 2010. / Committee Chair: Dr. Ayanna M. Howard; Committee Member: Dr. Charles Kemp; Committee Member: Dr. Magnus Egerstedt; Committee Member: Dr. Patricio Vela. Part of the SMARTech Electronic Thesis and Dissertation Collection.

<p> In this research project, a system is proposed to aid the visually impaired by providing partial contextual information of the surroundings using 360&deg; view camera combined with deep learning is proposed. The system uses a 360&deg; view camera with a mobile device to capture surrounding scene information and provide contextual information to the user in the form of audio. The system could also be used for other applications such as logo detection which visually impaired users can use for shopping assistance. </p><p> The scene information from the spherical camera feed is classified by identifying objects that contain contextual information of the scene. That is achieved using convolutional neural networks (CNN) for classification by leveraging CNN transfer learning properties using the pre-trained VGG-19 network. There are two challenges related to this paper, a classification and a segmentation challenge. As an initial prototype, we have experimented with general classes such restaurants, coffee shops and street signs. We have achieved a 92.8% classification accuracy in this research project.</p><p>

This thesis investigates several aspects of environment-driven adaptation in simulated evolutionary swarm robotics. It is centred around a specific algorithm for distributed embodied evolution called mEDEA. Firstly, mEDEA is extended with an explicit relative fitness measure while still maintaining the distributed nature of the algorithm. Two ways of using the relative fitness are investigated: influencing the spreading of genomes and performing an explicit genome selection. Both methods lead to an improvement in the swarm's abilityto maintain energy over longer periods. Secondly, a communication energy model is derived and introduced into the simulator to investigate the influence of accounting for the costs of communication in the distributed evolutionary algorithm where communication is a key component. Thirdly, a method is introduced that relates environmental conditions to a measure of the swarm's behaviour in a 3-dimensional map to study the environment's influence on the emergence of behaviours at the individual and swarm level. Interesting regions for further experimentation are identified in which algorithm specific characteristics show effect and can be explored. Finally, a novel individual learning method is developed and used to investigate how the most effective balance between evolutionary and lifetime-adaptation mechanisms is influenced by aspects of the environment a swarm operates in. The results show a clearlink between the effectiveness of different adaptation mechanisms and environmental conditions, specifically the rate of change and the availability of learning opportunities.

No description available.