An autonomous machine is a machine that can navigate through its environment without human interactions. These machines use sensors to sense the environment and have computing abilities for receiving and interpreting the sensory data as well as for controlling their displacement. At the University of Saskatchewan (Saskatoon, Canada), a low level autonomous machine was developed. This low level machine was the sensor system for an autonomous machine. The machine was capable of sensing the environment and carrying out actions based on commands sent to it. This machine provided a sensing and control layer, but the path planning (decision making) part of the autonomous machine was not developed.<p>This autonomous machine was developed on a Case IH DX 34H tractor with the purpose of providing a machine for testing software and sensors in a true agricultural environment. The tractor was equipped with sensors capable of sensing the speed and heading of the tractor. A control architecture was developed that received input commands from a human or computer in the form of a target heading and speed. The control architecture then adjusted controls on the tractor to make the tractor reach and maintain the target heading and speed until a new command was provided. The tractor was capable of being used in all kinds of weather, although some minor issues arose when testing in rain and snow. The sensor platform developed was found to be insufficient for proper control. The control structure appeared to work correctly, but was hindered by the poor sensor platform performance.

An autonomous machine is a machine that can navigate through its environment without human interactions. These machines use sensors to sense the environment and have computing abilities for receiving and interpreting the sensory data as well as for controlling their displacement. At the University of Saskatchewan (Saskatoon, Canada), a low level autonomous machine was developed. This low level machine was the sensor system for an autonomous machine. The machine was capable of sensing the environment and carrying out actions based on commands sent to it. This machine provided a sensing and control layer, but the path planning (decision making) part of the autonomous machine was not developed.<p>This autonomous machine was developed on a Case IH DX 34H tractor with the purpose of providing a machine for testing software and sensors in a true agricultural environment. The tractor was equipped with sensors capable of sensing the speed and heading of the tractor. A control architecture was developed that received input commands from a human or computer in the form of a target heading and speed. The control architecture then adjusted controls on the tractor to make the tractor reach and maintain the target heading and speed until a new command was provided. The tractor was capable of being used in all kinds of weather, although some minor issues arose when testing in rain and snow. The sensor platform developed was found to be insufficient for proper control. The control structure appeared to work correctly, but was hindered by the poor sensor platform performance.

We study the problem of Reinforcement Learning (RL) for Unmanned Aerial Vehicle (UAV) navigation with the smallest number of real world samples possible. This work is motivated by applications of learning autonomous navigation for aerial robots in structural inspec- tion. A naive RL implementation suffers from curse of dimensionality in large continuous state spaces. Gaussian Processes (GPs) exploit the spatial correlation to approximate state- action transition dynamics or value function in large state spaces. By incorporating GPs in naive Q-learning we achieve better performance in smaller number of samples. The evalua- tion is performed using simulations with an aerial robot. We also present a Multi-Fidelity Reinforcement Learning (MFRL) algorithm that leverages Gaussian Processes to learn the optimal policy in a real world environment leveraging samples gathered from a lower fidelity simulator. In MFRL, an agent uses multiple simulators of the real environment to perform actions. With multiple levels of fidelity in a simulator chain, the number of samples used in successively higher simulators can be reduced. / Master of Science / Increasing development in the field of infrastructure inspection using Unmanned Aerial Vehicles (UAVs) has been seen in the recent years. This thesis presents work related to UAV navigation using Reinforcement Learning (RL) with the smallest number of real world samples. A naive RL implementation suffers from the curse of dimensionality in large continuous state spaces. Gaussian Processes (GPs) exploit the spatial correlation to approximate state-action transition dynamics or value function in large state spaces. By incorporating GPs in naive Q-learning we achieve better performance in smaller number of samples. The evaluation is performed using simulations with an aerial robot. We also present a Multi-Fidelity Reinforcement Learning (MFRL) algorithm that leverages Gaussian Processes to learn the optimal policy in a real world environment leveraging samples gathered from a lower fidelity simulator. In MFRL, an agent uses multiple simulators of the real environment to perform actions. With multiple levels of fidelity in a simulator chain, the number of samples used in successively higher simulators can be reduced. By developing a bidirectional simulator chain, we try to provide a learning platform for the robots to safely learn required skills in the smallest possible number of real world samples possible.

Flooding can severely disrupt transportation systems. When safety measures are limited to road closures, vehicles affected by the flooding have an origin, destination, or path segment that is closed or soon-to-be flooded during the trip's duration. This thesis introduces a framework to provide routing assistance and trip cancellation recommendations to affected vehicles. The framework relies on the connected vehicle environment for real-time link performance measures and flood data and evaluates the trip of the vehicle to determine whether it is affected by the flood or not. If the vehicle is affected and can still leave its origin, the framework generates the corresponding routing assistance in the form of hyperpath(s) or set of alternative paths. On the other hand, a vehicle with a closed origin receives a warning to wait at origin, while a vehicle with an affected destination is assigned to a new safe one. This framework is tested on two transportation networks. The evaluation of the framework's scalability to different network sizes and the sensitivity of the results to various flood characteristics, policy-related variables and other dependencies are performed using simulated vehicle data and hypothetical flood scenarios. The computation times depends on the network size and flood depth but have generally an average of 1.47 seconds for the largest tested network and deepest tested flood. The framework has the potential to alleviate the impacts and inconveniences associated with flooding. / Master of Science / Flooding is a natural hazard that occurs with heavy rainfalls and high tides. In extreme situations, a flood in an area results in the evacuation of its occupant. Yet, in many cases, a flood is less severe and may only result in roads closures without necessitating evacuation. During these situations, and as transportation engineers, our ultimate goal is to maintain efficient and safe traffic operations. This thesis introduces a framework that focuses on providing routing assistance to affected vehicles and sending warnings to unaffected ones. It relies on the future connected vehicle environment which enables the communication between a traffic management center and equipped vehicles. The traffic management center collects and processes the information about the link performance measures and the weather and flood forecasts and sends them to the connected vehicles. Each vehicle has an in-vehicle navigation system in which the proposed framework is embedded. The framework, depending on the vehicle’s origin, destination, path and departure time and based on the flood’s characteristics, determines whether the vehicle is affected or not. If the vehicle is unaffected, it will receive a warning with the areas to avoid in case of any deviation and it can resume its trip as intended. If affected, the vehicle will either receive a warning to stay at its origin or routing guidance in the form of hyperpath or a set of alternative paths. The proposed framework has been evaluated on two transportation networks modeled in VISSIM based on the city of Virginia Beach, VA. Using simulated vehicle data and generated flood scenarios, several tests were executed to evaluate the scalability of the framework to different transportation networks along with the sensitivity of the results to variation in flood characteristics, policy-dependent variables and other dependencies. Concentrated, more intense and deeper floods resulted in a higher impact on the system. Yet, the analysis of the output is highly dependent on the location of the origin and destination of the vehicles with respect to the flooded roads. Thus, a lot of the output explanation are specific. Computation time increased with the increase in network size and in the flood depth. Nevertheless, it is still small and reasonable and further increase in both parameters (network size and flood depth) can be tested in future along with multiple techniques that minimize the computation time. This framework addresses the flooding hazard which road users are experiencing more and more nowadays. This hazard brings risks and inconveniences to our daily life. Thus, the development of this framework is of great interest to our society as it is a promising tool that has the potential to offer benefits, in terms of safety and mobility, to roads users exposed to a flood hazard. Its first implementation shows that it is a timely application with a potential to perform even better with future improvements.

International Telemetering Conference Proceedings / October 20-23, 2003 / Riviera Hotel and Convention Center, Las Vegas, Nevada / The intelligent vehicle navigation system is a multifunctional and complex integrated system that uses autonomous vehicle navigation, geography information, database system, computer technology, multimedia technology and wireless communication. In this paper, an autonomous navigation system based on embedded hardware and embedded operation system that is Linux is proposed. The system has advantages of low cost, small mass, multifunction and high stability, especially connecting with Internet.

International Telemetering Conference Proceedings / October 23-26, 2000 / Town & Country Hotel and Conference Center, San Diego, California / NASA Marshall Space Flight Center’s Bantam Robust Guidance Navigation & Control Project is investigating off the shelf navigation sensors that may be inexpensively combined into Kalman filters specifically tuned for launch and orbital vehicles. For this purpose, Marshall has purchased several GPS receivers and is evaluating them for these applications. The paper will discuss the receiver selection criteria and the test equipment used for evaluation. An overview of the analysis will be presented including the evaluation used to determine their success or failure. It will conclude with goals of the program and a recommendation for all GPS users.

<p>Automotive positioning systems relying exclusively on the input from a GPS receiver, which is a line of sight sensor, tend to be sensitive to situations with limited sky visibility. Such situations include: urban environments with tall buildings; inside parking structures; underneath trees; in tunnels and under bridges. In these situations, the system has to rely on integration of relative motion sensors to estimate vehicle position. However, these sensor measurements are generally affected by errors such as offsets and scale factors, that will cause the resulting position accuracy to deteriorate rapidly once GPS input is lost. </p><p>The approach in this thesis is to use a GPS receiver in combination with low cost sensor equipment to produce a robust positioning module. The module should be capable of handling situations where GPS input is corrupted or unavailable. The working principle is to calibrate the relative motion sensors when GPS is available to improve the accuracy during GPS intermission. To fuse the GPS information with the sensor outputs, different models have been proposed and evaluated on real data sets. These models tend to be nonlinear, and have therefore been processed in an Extended Kalman Filter structure. </p><p>Experiments show that the proposed solutions can compensate for most of the errors associated with the relative motion sensors, and that the resulting positioning accuracy is improved accordingly.</p>

In modern cars, advanced driver assistance systems are used to aid the driver and increase the automobile safety. An example of such a system is the night vision system designed to detect and warn for pedestrians in danger of being hit by the car. To determine if a warning should be given when a pedestrian is detected, the system requires a prediction of the future path of the car for up to four seconds ahead in time. In this master's thesis, a new path prediction algorithm based on satellite positioning and a digital map database has been developed. The algorithm uses an extended Kalman filter to get an accurate estimate of the current position and heading direction of the car. The estimate is then matched to a position in the map database and the possible future paths of the vehicle are predicted using the road network. The performance of the path prediction algorithm has been evaluated on recorded night vision sequences corresponding to 15 hours of driving. The results show that map-based path prediction algorithms are superior to dead-reckoning methods for longer time horizons. It has also been investigated whether vision-based lane detection and tracking can be used to improve the path prediction. A prediction method using lane markings has been implemented and evaluated on recorded sequences. Based on the results, the conclusion is that lane detection can be used to support a path prediction system when lane markings are clearly visible.

Automotive positioning systems relying exclusively on the input from a GPS receiver, which is a line of sight sensor, tend to be sensitive to situations with limited sky visibility. Such situations include: urban environments with tall buildings; inside parking structures; underneath trees; in tunnels and under bridges. In these situations, the system has to rely on integration of relative motion sensors to estimate vehicle position. However, these sensor measurements are generally affected by errors such as offsets and scale factors, that will cause the resulting position accuracy to deteriorate rapidly once GPS input is lost. The approach in this thesis is to use a GPS receiver in combination with low cost sensor equipment to produce a robust positioning module. The module should be capable of handling situations where GPS input is corrupted or unavailable. The working principle is to calibrate the relative motion sensors when GPS is available to improve the accuracy during GPS intermission. To fuse the GPS information with the sensor outputs, different models have been proposed and evaluated on real data sets. These models tend to be nonlinear, and have therefore been processed in an Extended Kalman Filter structure. Experiments show that the proposed solutions can compensate for most of the errors associated with the relative motion sensors, and that the resulting positioning accuracy is improved accordingly.

The transportation routing map is increasingly used in various transportation network modeling applications such as vehicle navigation and traffic assignment modeling. A typical navigation GIS map contains all detailed road facility layers and may not be as computationally efficient as a lower-resolution map for path finding. A lower-resolution transportation routing map retains only route-finding related roadways and is efficient for path finding but may result in sub-optimal routes because of misclassification links. With the goal in balancing the traffic analysis requirement of intended application and computation requirements of transportation navigation and traffic assignment, the systematic abstraction of the lower-resolution transportation routing map from high resolution map is an important and non-trivial task. For vehicle navigation applications, the traffic analysis requirement is the shortest path quality. An innovative transportation routing map abstraction method or Connectivity Enhancement Algorithm (CEA) is proposed to deal with vehicle navigation application routing map abstraction. The algorithm starts from a low-resolution network and keeps updating the map by adding links and nodes when it processes each search set. The outcome of the algorithm is an abstract map that retains the original detailed map's hierarchical structure with quality topological connectivity at a significant computations saving. With the development of traffic assignment modeling, a detailed network is desired to describe the real world traffic network. It is the consensus that one should not directly apply a GIS map blind-sight without a systematic approach and unnecessarily overuse the network details causes excessive run time. The traffic analysis requirement of those applications is the dynamic user equilibrium (DUE) condition network performance is identical or near-identical with high resolution network. The lowest network resolution level that meets the requirements of emerging traffic analysis is not easy to determine. The proposed traffic analysis network abstraction method gives a solution for this problem. It is an iterative network abstraction approach and considers the link travel time with DUE traffic condition. The case study and numerical analysis prove that the two network abstraction methods are sound and promising. The transportation routing map abstraction method could detect most misclassification links and is robust for different network scales. The abstracted navigation map provides the identical or near-identical SP cost/travel time for any OD pair while the computation burden is much lighter than that on original map. In another hand, the case studies about the traffic analysis network abstraction tell that the method converges very quick and the rendered the abstracted network that has lowest resolution of network or least links and nodes but the DUE condition network performance or trips cost/travel time is much closer to that on the original map.