107 results
Search results
Now showing 1 - 10 of 107
Publication Machine Learning For Robot Motion Planning(2021-01-01) Zhang, Clark JuneRobot motion planning is a field that encompasses many different problems and algorithms. From the traditional piano mover's problem to more complicated kinodynamic planning problems, motion planning requires a broad breadth of human expertise and time to design well functioning algorithms. A traditional motion planning pipeline consists of modeling a system and then designing a planner and planning heuristics. Each part of this pipeline can incorporate machine learning. Planners and planning heuristics can benefit from machine learned heuristics, while system modeling can benefit from model learning. Each aspect of the motion planning pipeline comes with trade offs between computational effort and human effort. This work explores algorithms that allow motion planning algorithms and frameworks to find a compromise between the two. First, a framework for learning heuristics for sampling-based planners is presented. The efficacy of the framework depends on human designed features and policy architecture. Next, a framework for learning system models is presented that incorporates human knowledge as constraints. The amount of human effort can be modulated by the quality of the constraints given. Lastly, semi-automatic constraint generation is explored to enable a larger range of trade-offs between human expert constraint generation and data driven constraint generation. We apply these techniques and show results in a variety of robotic systems.Publication Computer-Aided Clinical Trials For Medical Devices(2021-01-01) Jang, Kuk JinLife-critical medical devices require robust safety and efficacy to treat patient populations with potentially large patient heterogeneity. Today, the de facto standard for evaluating medical devices is the randomized controlled trial. However, even after years of device development many clinical trials fail. For example, in the Rhythm ID Goes Head to Head Trial (RIGHT) the risk for inappropriate therapy by implantable cardioverter defibrillators (ICDs) actually increased relative to control treatments. With recent advances in physiological modeling and devices incorporating more complex software components, population-level device outcomes can be obtained with scalable simulations. Consequently, there is a need for data-driven approaches to provide early insight prior to the trial, lowering the cost of trials using patient and device models, and quantifying the robustness of the outcome. This work presents a clinical trial modeling and statistical framework which utilizes simulation to improve the evaluation of medical device software, such as the algorithms in ICDs. First, a method for generating virtual cohorts using a physiological simulator is introduced. Next, we present our framework which combines virtual cohorts with real data to evaluate the efficacy and allows quantifying the uncertainty due to the use of simulation. Results predicting the outcome of RIGHT and improving statistical power while reducing the sample size are shown. Finally, we improve device performance with an approach using Bayesian optimization. Device performance can degrade when deployed to a general population despite success in clinical trials. Our approach improves the performance of the device with outcomes aligned with the MADIT-RIT clinical trial. This work provides a rigorous approach towards improving the development and evaluation of medical treatments.Publication Imu-Based State Estimation And Control Of Quadrotors Exploiting Aerodynamic Effects(2019-01-01) Svacha, James BairdQuadrotors and multirotors in general are common in many inspection and surveillance applications. For these applications, visual-inertial odometry is a common way to localize the vehicles and observe the environment. However, unlike with wheeled mobile robots, quadrotor localization algorithms often do not use knowledge of the control inputs and the full vehicle dynamics as a process model for localization. Rather, they use kinematic models, with the IMU providing acceleration and angular velocity. One of the reasons for avoiding the use of dynamics is that, until recently quadrotor aerodynamic effects have not been considered in the literature and hence the dynamic models for quadrotors have been less accurate than those for wheeled mobile robots. The main aerodynamic terms that are significant are first-order effects that are linear in velocity and angular velocity. They are predominantly caused by aerodynamic interaction with the spinning propellers. This work investigates the models for such effects, as well as what can be gained if such aerodynamic effects are incorporated into the dynamic model and the full dynamics are used for state estimation. We develop novel IMU-based filters, the end results of which are used to estimate the wind velocity of the quadrotor or, indoors, when the ambient wind is zero, the velocity of the quadrotor. In addition, these filters estimate the many aerodynamic parameters in the model online. They may also be used to estimate sensor biases and inertial parameters. We demonstrate the effectiveness of these filters through experiments. We also present nonlinear observability analyses that theoretically determine the observability properties of the systems.Publication Efficient Learning and Inference for High-dimensional Lagrangian Systems(2011-04-18) Vernaza, Paul NLearning the nature of a physical system is a problem that presents many challenges and opportunities owing to the unique structure associated with such systems. Many physical systems of practical interest in engineering are high-dimensional, which prohibits the application of standard learning methods to such problems. This first part of this work proposes therefore to solve learning problems associated with physical systems by identifying their low-dimensional Lagrangian structure. Algorithms are given to learn this structure in the case that it is obscured by a change of coordinates. The associated inference problem corresponds to solving a high-dimensional minimum-cost path problem, which can be solved by exploiting the symmetry of the problem. These techniques are demonstrated via an application to learning from high-dimensional human motion capture data. The second part of this work is concerned with the application of these methods to high-dimensional motion planning. Algorithms are given to learn and exploit the struc- ture of holonomic motion planning problems effectively via spectral analysis and iterative dynamic programming, admitting solutions to problems of unprecedented dimension com- pared to known methods for optimal motion planning. The quality of solutions found is also demonstrated to be much superior in practice to those obtained via sampling-based planning and smoothing, in both simulated problems and experiments with a robot arm. This work therefore provides strong validation of the idea that learning low-dimensional structure is the key to future advances in this field.Publication Autonomous Behaviors With A Legged Robot(2018-01-01) Ilhan, Berkay DenizOver the last ten years, technological advancements in sensory, motor, and computational capabilities have made it a real possibility for a legged robotic platform to traverse a diverse set of terrains and execute a variety of tasks on its own, with little to no outside intervention. However, there are still several technical challenges to be addressed in order to reach complete autonomy, where such a platform operates as an independent entity that communicates and cooperates with other intelligent systems, including humans. A central limitation for reaching this ultimate goal is modeling the world in which the robot is operating, the tasks it needs to execute, the sensors it is equipped with, and its level of mobility, all in a unified setting. This thesis presents a simple approach resulting in control strategies that are backed by a suite of formal correctness guarantees. We showcase the virtues of this approach via implementation of two behaviors on a legged mobile platform, autonomous natural terrain ascent and indoor multi-flight stairwell ascent, where we report on an extensive set of experiments demonstrating their empirical success. Lastly, we explore how to deal with violations to these models, specifically the robot's environment, where we present two possible extensions with potential performance improvements under such conditions.Publication Quantitative Mapping of Lung Ventilation Using Hyperpolarized Gas Magnetic Resonance Imaging(2011-05-16) Emami, KiarashThe main objective of this project was to develop and implement techniques for high-resolution quantitative imaging of ventilation in lungs using hyperpolarized gas magnetic resonance imaging (MRI). Pulmonary ventilation is an important aspect of lung function and is frequently compromised through several different mechanisms and at varying degrees in presence of certain lung conditions, such as chronic obstructive pulmonary diseases. The primary focus of this development is on large mammalian species as a steppingstone towards translation to human subjects. The key deliverables of this project are a device for real-time mixing and delivery of hyperpolarized gases such as 3He and 129Xe in combination with O2, an MRI acquisition scheme for practical imaging of ventilation signal build-up in the lungs, and a robust mathematical model for estimation of regional fractional ventilation values at a high resolution. A theoretical framework for fractional gas replacement in the lungs is presented to describe MRI signal dynamics during continuous breathing of a mixture of hyperpolarized gases in presence of several depolarization mechanisms. A hybrid ventilation and imaging acquisition scheme is proposed to acquire a series of images during short end-inspiratory breath-holds over several breaths. The sensitivity of the estimation algorithm is assessed with respect to noise, model uncertainty and acquisition parameters, and subsequently an optimal set of acquisition parameters is proposed to minimize the fractional ventilation estimation error. This framework is then augmented by an undersampled parallel MRI scheme to accelerate image acquisition to enable fractional ventilation imaging over the entire lung volume in a single pass. The image undersampling was also leveraged to minimize the coupling associated with signal buildup in the airways and the irreversible effect of RF pulses. The proposed technique was successfully implemented in pigs under mechanical ventilation, and preliminary measurements were performed in an adult human subject under voluntary breathing.Publication Generative-Discriminative Low Rank Decomposition for Medical Imaging Applications(2012-01-01) Batmanghelich, Nematollah KayhanIn this thesis, we propose a method that can be used to extract biomarkers from medical images toward early diagnosis of abnormalities. Surge of demand for biomarkers and availability of medical images in the recent years call for accurate, repeatable, and interpretable approaches for extracting meaningful imaging features. However, extracting such information from medical images is a challenging task because the number of pixels (voxels) in a typical image is in order of millions while even a large sample-size in medical image dataset does not usually exceed a few hundred. Nevertheless, depending on the nature of an abnormality, only a parsimonious subset of voxels is typically relevant to the disease; therefore various notions of sparsity are exploited in this thesis to improve the generalization performance of the prediction task. We propose a novel discriminative dimensionality reduction method that yields good classification performance on various datasets without compromising the clinical interpretability of the results. This is achieved by combining the modelling strength of generative learning framework and the classification performance of discriminative learning paradigm. Clinical interpretability can be viewed as an additional measure of evaluation and is also helpful in designing methods that account for the clinical prior such as association of certain areas in a brain to a particular cognitive task or connectivity of some brain regions via neural fibres. We formulate our method as a large-scale optimization problem to solve a constrained matrix factorization. Finding an optimal solution of the large-scale matrix factorization renders off-the-shelf solver computationally prohibitive; therefore, we designed an efficient algorithm based on the proximal method to address the computational bottle-neck of the optimization problem. Our formulation is readily extended for different scenarios such as cases where a large cohort of subjects has uncertain or no class labels (semi-supervised learning) or a case where each subject has a battery of imaging channels (multi-channel), \etc. We show that by using various notions of sparsity as feasible sets of the optimization problem, we can encode different forms of prior knowledge ranging from brain parcellation to brain connectivity.Publication Self-Manipulation and Dynamic Transitions for a Legged Robot(2014-01-01) Johnson, Aaron M.How can we make a robot that can go anywhere on its own? This thesis presents several new behaviors on the RHex robot that greatly increase the variety of obstacles that it can overcome, including vertical jumps, flips, leaps onto and across ledges, aerial reorientations, and proprioceptively-aware behaviors. These behaviors inspire new tools to model and understand their transitional nature, wherein it is no longer useful to think of each step as being an equal part of a steady state gait. Legged robots will necessarily experience a variety of changing contact conditions as they locomote in complex environments epitomized by the rocky, sandy desert. Drawing on the much more mature literature of robot manipulation, this thesis presents the new modeling paradigm of "self-manipulation" that formally generates analytical equations of motion across all contact states. The framework is amenable to many ubiquitous simplifying assumptions (such as rigid bodies, plastic impact, persistent contact, Coulomb friction, and massless limbs) to reduce the complexity of these models despite the obvious physical inaccuracies that each incurs. Nevertheless the models capture enough of the physical world to represent the challenges confronting interesting behaviors in a qualitatively correct manor, including the effects of impulsive transitions between the various contact modes. More than numerical simulation, our goal is the distillation of these physically parametrized models into formal design insights (platform design, behavior design, and controller design), utilizing a variety of analytical and numerical methods. These behaviors are only possible with a robot designed to be both robust and powerful, and they make use of the unique capability of legged machines to interact with the environment in varied and, possibly, unpredictable ways. Careful actuator modeling is needed to achieve such acrobatic results, and so this thesis presents a spectrum of motor sizing tasks to ensure that the platform is up to the task. These tools are used to gain insight into various dynamic transitions for RHex, and we conjecture that their generalization will be of importance for a broad class of legged robots operating in remote and unstructured terrain.Publication Communication Aware Mobile Robot Teams(2015-01-01) Stephan, James EdwardThe type of scenarios that could benefit from a team of robots that are able to self configure into an ad-hoc multi-hop mobile communication network while completing a task in an unknown environment, range from search and rescue in a partially collapsed building to providing a security perimeter around a region of interest. In this thesis, we present a hybrid system that enables a team of robots to maintain a prescribed end-to-end data rate while moving through a complex unknown environment, in a distributed manner, to complete a specific task. This is achieved by a systematic decomposition of the real-time situational awareness problem into subproblems that can be efficiently solved by distributed optimization. The validity of this approach is demonstrated through multiple simulations and experiments in which the a team of robots is able to accurately map an unknown environment and then transition to complete a traditional situational awareness task. We also present MCTP, a lightweight communication protocol that is specifically designed for use in ad-hoc multi-hop wireless networks composed of low-cost low-power transceivers. This protocol leverages the spatial diversity found in mobile robot teams as well as recently developed robust routing systems designed to minimize the variance of the end-to-end communication link. The combination of the hybrid system and MCTP results in a system that is able to complete a task, with minimal global coordination, while providing near loss-less communication over an ad-hoc multi-hop network created by the members of the team in unknown environments.Publication Wave Interaction With Epsilon-znd-Mu-Near-Zero (emnz) Platforms and Nonreciprocal Metastructures(2016-01-01) Mahmoud, Ahmed Mohamed AbdelwahabThe concept of metamaterials has offered platforms for unconventional tailoring and manipulation of the light-matter interaction. In this dissertation, we explore several concepts and designs within this scope. We investigate some of the electromagnetic characteristics of the concept of “static optics”, i.e., wave interaction with structures in which both the relative effective permittivity and permeability attain near-zero values at a given operating frequency and thus the spatial distributions of the electric and magnetic fields exhibit curl-free features, while the fields are temporally dynamic. Using such structures, one might in principle ‘open up’ and ‘stretch’ the space, and have regions behaving electromagnetically as ‘single points’ despite being electrically large. We study some of the wave-matter interaction in these platforms and suggest possible designs for implementation of such structures in different frequency regimes and experimentally verify our findings in the microwave regime. Another research direction that is explored in this dissertation is the development of some nonreciprocal metaplatforms. We investigate theoretically an approach through which one-way electromagnetic wave flow can be achieved using properly designed nonlinearity combined with structural asymmetry. The approach is rather general and applicable for any desired frequency regime and opens doors for high performance “electromagnetic diodes” and nonreciprocal metasurfaces and metastructures. We also theoretically study the usage of time-dependent materials in achieving wave flow isolation within plasmonic waveguides environments. We also provide physical remarks on our various findings.