Autonomous robots are intelligent machines capable of performing tasks in the real world without explicit human control for extended periods of time. A high degree of autonomy is particularly desirable in fields where robots can replace human workers, such as state-of-the-practice video surveillance system and space exploration. However, not having humans sophisticated sensing and control system, two broad open problems in autonomous robot systems are the perceptual discrepancy problem, that is, there is no guarantee that the robot sensing system can recognize or detect objects defined by a human designer, and the autonomous control problem, that is, how the robots can operate in unstructured environments without continuous human guidance. As a result, autonomous robot systems should have their own ways to acquire percepts and control by learning. In this work, a computer vision system is used for visual percept acquisition and a working memory toolkit is used for robot autonomous control. Natural images contain statistical regularities which can set objects apart from each other and from random noise. For an object to be recognized in a given image, it is often necessary to segment the image into nonoverlapping but meaningful regions whose union is the entire image. Therefore, a biologically based percept acquisition system is developed to build an efficient low-level abstraction of real-world data into percepts. Perception in animals is strongly related to the type of behavior they perform. Learning plays a major part in this process. To solve how the robots can learn to autonomously control their behavior based on percepts theyve acquired, the computer vision system is integrated with a software package called the Working Memory Toolkit (WMtk) for decision making and learning. The WMtk was developed by Joshua L. Phillips & David C. Noelle based on a neuron computational model of primate working memory system. The success of the whole system is demonstrated by its application to a navigation task.

Accurate analysis of the hepatic vasculature is of great importance for many medical applications, such as liver surgical planning and diagnosis of tumors and/or vascular diseases. Vessel segmentation is a pivotal step for the morphological and topological analysis of the vascular systems. Physical imaging limitations together with the inherent geometrical complexity of the vessels make the problem challenging. In this work, we present an automatic method to separate and render the portal and hepatic veins in multi-phase CT images. This methods involve image enhancement using an Hessian filter, automatic selection of thresholds to separate vasculature and liver parenchyma, and an iterative classification step for vessel separation. We also propose a series of methods and techniques that segment the portal vein and the hepatic vein, and extract the centerlines of both vessel trees. The approaches have been tested with success on clinical multiphase CT data sets acquired as part of the standard delivery of care.

A key goal in machine vision is to understand how the actions of sentient agents such as humans are processed, identified, and understood. The most apparent challenge is the need to segment a continuous set of visual movements into meaningful discrete actions. Part of the work of the intentional vision research was to detect a set of determining features exhibited by human participants that account for the selection of significant action boundaries as judged by human raters. They found that action boundaries could be identified by a set of sub-actions such as hand-to-object contacts, object-to-object contacts, occlusions, and eye movements. Our goal was to create a cost effective vision system to be an easy-to-use tool for training and tracking to aid in analysis of video recordings of experiments for non-vision specialists. The system was validated for human motion analysis by applying it in conjunction with psychological studies performed with the intentional vision research. The results show correlation with the human rater data gathered from the intentional vision research showing that the cues observed in the intentional vision research are captured in our behavior feature vector. The system was extended to perform autonomous segmentation and analysis for motion studies to expand the possibility of interdisciplinary use. Of the 100 videos collected, 84 were successfully segmented and analyzed without intervention. The autonomous system was also shown to yield good results in natural scene segmentation.

In this thesis, we examine the total dose response of planar fully depleted planar SOI MOSFETs fabricated in a FinFET technology as functions of both drain bias and gate length. The ID for negative Vgf increases with the drain bias and decreases with the gate length. The mechanisms that are involved include: band-to-band tunneling (BBT), positive charge trapping in the BOX, direct tunneling through the thin gate oxide, and short-channel effects. In order to extend our TID understanding to more advanced FinFETs, devices with narrower fins (40 nm and 80 nm) and shorter gate lengths (100 nm) were critically studied. Both the threshold-voltage shift and the subthreshold swing (SS) were analyzed as functions of device dimensions and total dose. Our experimental results suggest that irradiated FinFET devices with narrower fins are more tolerant to TID effects. This was explained by the additional lateral gate control that attenuates the coupling effects between the front and the back gates, and decreases the fringing field effects originated from the drain terminal.

This thesis presents a multi-core architecture, the Cell processor, and an updated model-based tool suite named the Signal Processing Platform (SPP) that supports development of high performance signal processing applications, such as an automatic target recognition system, that execute on and take advantage of the computational power of the Cell. It introduces the move from single-core to multi-core architectures. Specifically, it covers the larger features of the Cell processor that allow its specialized multi-core architecture to provide a significant amount of increased performance. Specialized configurations call for specialized programming in order to harness the available performance increase. Such high computation configurations are prime targets for signal processing applications. A major goal existed for the SPP tool set to provide a system construct for executing applications on the Cell. Modifications to the tool suite were presented and monitored by porting an example automatic target recognition application to the Cell. This paper shows the steps of supporting the multi-core architecture with the design tools, yielding a significant increase in performance, and closes with future techniques for improving the automation element to the design process.

<p>A machines intelligence is related to its autonomy. Making robots autonomous is perhaps the most difficult current challenge in the state-of-the art of robotics research. Robots are said to be autonomous, if they can learn and represent percepts through their own ways and internal representations, and as a result, can undertake tasks in a real, dynamic physical world without significant human intervention. The major problems of autonomy involve acquiring, representing, and learning percepts and using them for autonomous control. Todays robots are far from complete autonomy. There is extensive human intervention in the design and construction of the perceptual and control systems. Rather than allowing the robot to learn percepts and control itself, human programmers encode their own ideas about knowledge and control algorithms in to the robots. As it is cumbersome and practically impossible to encode all knowledge about a dynamic and unstructured environment, autonomously learning the percepts may be the only way for a robot to acquire complete autonomy. Visual percepts would be among those to be learned by the robot autonomously. The large amounts of data associated with vision necessitate an efficient representation for visual percepts. The visual features to use, the data structures to represent them, and the algorithms used to learn them are critical to the success of autonomous perception.</p> <p>This work builds on existing algorithms for computer vision that learn a specific percept with minimal intervention. It combines two features, a color probability density function and a texture measure gleaned from overlapping portions of an image. It uses an approximate nearest neighbor learning algorithm implemented with a random 3-way search tree to learn the visual input features and to associate them with labeled percepts. Once learned, the search tree is used to segment the input images into classes. The time and space complexities of the algorithm are studied and compared with a pure (naïve) nearest neighbor search implementation. The accuracy of the system is then tested.</p> <p>The results of the experiments suggest that the random tree implementation is an efficient and accurate algorithm. The classification running time of the algorithm is found to be approximately logarithmic with linear space requirements, whereas the run time of the pure algorithm is approximately quadratic for a very high dimensional feature vectors like the ones used in this project. </p>

The demand for high power high frequency semiconductor devices has led to the development of microwave power devices using GaN and SiC. AlGaN/GaN HEMTs have shown power densities of 9.8 W/mm at 8 GHz. Although these results are very encouraging, significant work needs to be done to improve performance. It is generally recognized that trapping effects limit the performance of these devices. In this dissertation we study the impact of bulk traps on three distinct characteristics of these devices. These 3 mechanisms are: degradation in the IV characteristics, self-heating and gate-lag. Displacement-damage induced degradation in AlGaN/AlN/GaN HEMTs with polarization charge induced 2DEGs is examined using simulations and experiments. Carrier removal in the unintentionally doped AlGaN layer changes the space charge in the structure and this changes the band bending. The band bending decreases the 2DEG density, which in turn reduces the drain current in the device. The effect of the defect energy levels on the 2DEG density is also studied. The interplay between band bending, mobility degradation, and the charged defects is analyzed and quantified. Experiments and TCAD simulations are used to study the relationship between bulk traps, self-heating and mobility degradation in AlGaN/AlN/GaN HEMTs. Bulk traps in the GaN channel region and other regions of the device degrade the 2DEG density and the mobility in the device. This in turn degrades the performance of the device. Mobility degradation is closely coupled with the self-heating in the device. The interplay between bulk traps, mobility degradation and self-heating is analyzed and quantified. Experiments and simulations showing the impact of proton irradiation induced bulk traps on gate lag in AlGaN/AlN/GaN HEMTs are analyzed. Pre-existing donor-like surface traps in the gate-drain and source-gate access regions cause the majority of the gate-lag in the device. The simulations indicate that these traps at the AlGaN/Nitride surface are very close to the valence band. Gate lag increases with increased bulk traps. This is due to the reduction in the 2DEG density as a result of band bending and mobility degradation. The experiments and simulations did not indicate any substantial hot electron induced current collapse due to bulk traps.

<p> As technology nodes scale smaller, digital circuits are able to run at higher clock frequencies, but they can become more susceptible to single-event induced errors. These types of errors can be generated in combinational logic and in storage cells. Traditional methods of characterizing digital circuits for single-event effects have difficulty distinguishing combinational logic errors from storage cell errors at high speeds. <p> In this thesis, a new approach of characterizing single-event effects in combinational logic is described called the Combinational Circuit for Radiation Effects Self-Test (C-CREST). This approach allows the SET cross-section of combinational logic to be increased while minimizing propagation delay. Various types of digital circuits can then be tested at speeds determined by their technology node along with allowing combinational logic errors to be distinguished from storage cell errors.

Wireless sensor networks offer a promising way to monitor the structural health of national infrastructure such as bridges. This has become critical because of recent catastrophes that have demonstrated the need for improved assessment and monitoring of these structures. This work examines potential computational constraints that are involved with implementing a real-time system for structural health monitoring. The target system should provide an adequate programming platform that can be used for built-in data security and on-chip data processing. It should also utilize existing technology in a way that will provide a thorough and cost-effective means of monitoring bridges in a real-time environment. The performance of the sensor is calculated from a comparative analysis of the various execution times at multiple data sampling frequencies. This constrains the processing time available to perform the computationally intensive operations required for real-time data monitoring. Results show that besides the memory limitations, the sensor can accommodate a real-time system that efficiently samples data within a relevant range.

The small band-gap and the non-ideal high-k germanium interface makes Ge p-MOSFETs susceptible to various reliability issues. Similarly, the non-ideal SiO2-SiC interface in SiC MOS capacitors makes them susceptible to radiation damage. In this work radiation and bias temperature stress response of MOS capacitors fabricated on both these materials are studied. Ge p-MOSFETs are shown to be susceptible to enhanced junction leakage in total dose environments. The mechanisms behind this increase in junction leakage are researched in this work. It is shown that the increase in surface generation current component of the Ge p+-n junction is responsible for increase in off-state leakage current in p-MOSFETs. Further modifications in Ge p-MOSFET processing, such as variation in Si monolayer thickness and variation in halo doping, are researched to find an optimum process that provides minimum junction leakage and maximum on/off current ratio. It is shown that a process with 8 Si monolayers provides much better pre-irradiation interface trap properties and maintains a better on/off current ratio than a device with 5 Si mono-layers. An optimum value of halo doping is found which provides the minimum junction leakage. Bias temperature stress (BTS) studies on Ge MOS capacitors showed that the devices without any interlayer (with high-k directly deposited on Ge) are particularly susceptible to temperature stress. Accumulation capacitance and interface trap density was found to decrease temperature stress. This indicates growth of a thin interlayer and diffusion of Ge into the high-k layer with temperature stress. The radiation response of SiC MOS capacitors with SiO2 gate dielectric is also studied in this work. MOS capacitors fabricated on 3C- and 4H-SiC polytypes are studied. These MOS capacitors are nitrided with either NO or N2O as nitridation agent. It is shown that MOS capacitors with N2O nitridation have higher starting interface trap density on both 3C- and 4H-SiC. N2O nitrided MOS capacitors trap more radiation-induced charge than the NO treated MOS capacitors on both 3C- and 4H-SiC. This is due to greater content of nitrogen deposited at the SiO2-SiC interface for NO treated MOS. Secondary ion mass spectroscopy (SIMS) measurements show that NO treated MOS devices indeed deposit a greater content of nitrogen at interface than N2O. 3C-SiC traps more charge than 4H-SiC MOS capacitors. This may be attributed to better quality of 4H-SiC substrates.