Being able to train machine learning models on simulated data can be of great interest in several applications, one of them being for autonomous driving of cars. The reason is that it is easier to collect large labeled datasets as well as performing reinforcement learning in simulations. However, transferring these learned models to the real-world environment can be hard due to differences between the simulation and the reality; for example, differences in material, textures, lighting and content. One approach is to use domain adaptation, by making the simulations as similar as possible to the reality. The thesis's main focus is to investigate domain adaptation as a way to meet the reality-gap, and also compare it to an alternative method, domain randomization. Two different methods of domain adaptation; one adapting the simulated data to reality, and the other adapting the test data to simulation, are compared to using domain randomization. These are evaluated with a classifier making decisions for a robot car while driving in reality. The evaluation consists of a quantitative evaluation on real-world data and a qualitative evaluation aiming to observe how well the robot is driving and avoiding obstacles. The results show that the reality-gap is very large and that the examined methods reduce it, with the two using domain adaptation resulting in the largest decrease. However, none of them led to satisfactory driving.

A robot is proficient if its performance for its task(s) satisfies a specific standard. While the design of autonomous robots often emphasizes such proficiency, another important attribute of autonomous robot systems is their ability to evaluate their own proficiency. A robot should be able to conduct proficiency self-assessment (PSA), i.e. assess how well it can perform a task before, during, and after it has attempted the task. We propose the assumption-alignment tracking (AAT) method, which provides time-indexed assessments of the veracity of robot generators' assumptions, for designing autonomous robots that can effectively evaluate their own performance. AAT can be considered as a general framework for using robot sensory data to extract useful features, which are then used to build data-driven PSA models. We develop various AAT-based data-driven approaches to PSA from different perspectives. First, we use AAT for estimating robot performance. AAT features encode how the robot's current running condition varies from the normal condition, which correlates with the deviation level between the robot's current performance and normal performance. We use the k-nearest neighbor algorithm to model that correlation. Second, AAT features are used for anomaly detection. We treat anomaly detection as a one-class classification problem where only data from the robot operating in normal conditions are used in training, decreasing the burden on acquiring data in various abnormal conditions. The cluster boundary of data points from normal conditions, which serves as the decision boundary between normal and abnormal conditions, can be identified by mainstream one-class classification algorithms. Third, we improve PSA models that predict robot success/failure by introducing meta-PSA models that assess the correctness of PSA models. The probability that a PSA model's prediction is correct is conditioned on four features: 1) the mean distance from a test sample to its nearest neighbors in the training set; 2) the predicted probability of success made by the PSA model; 3) the ratio between the robot's current performance and its performance standard; and 4) the percentage of the task the robot has already completed. Meta-PSA models trained on the four features using a Random Forest algorithm improve PSA models with respect to both discriminability and calibration. Finally, we explore how AAT can be used to generate a new type of explanation of robot behavior/policy from the perspective of a robot's proficiency. AAT provides three pieces of information for explanation generation: (1) veracity assessment of the assumptions on which the robot's generators rely; (2) proficiency assessment measured by the probability that the robot will successfully accomplish its task; and (3) counterfactual proficiency assessment computed with the veracity of some assumptions varied hypothetically. The information provided by AAT fits the situation awareness-based framework for explainable artificial intelligence. The efficacy of AAT is comprehensively evaluated using robot systems with a variety of robot types, generators, hardware, and tasks, including a simulated robot navigating in a maze-based (discrete time) Markov chain environment, a simulated robot navigating in a continuous environment, and both a simulated and a real-world robot arranging blocks of different shapes and colors in a specific order on a table.

No description available.

The industry for autonomous vehicles is growing. According to studies nine out of ten traffic accidents are due to the human factor, if the safety can get good enough in autonomous cars they have the potential to save thousands of lives every year. But obstacle detecting autonomous robots can be used in other situations as well, for example where the terrain is inaccessible for humans because of different reasons. In this project, a self navigating obstacle detecting robot was made. The robot uses ultrasonic sensors to detect obstacles and avoid them. An algorithm of the navigation of the robot was created and implemented to the Arduino. For driving the wheels, two servo motors were used. The robot consisted of three wheels, two in the back to which the servo motors were attached and one caster wheel in the front. This made it possible to implement differential drive which enabled quick and tight turns. Tests were performed which showed that the robot could successfully navigate in a room with various obstacles placed out. The placement of the sensors worked good considering the amount of sensors that was used. Improvements in detection of obstacles could have been made if more sensors had been used. The tests also confirmed that ultrasonic sensors works good for this kind of task. / Industrin för självkörande fordon växer. Enligt studier beror nio av tio trafikolyckor på den mänskliga faktorn, om säkerheten kan bli tillräckligt bra i självkörande bilar har de potential att rädda tusentals liv varje år. Men hinderupptäckande självkörande robotar kan användas i andra situationer också, till exempel i terräng som är otillgänglig för människor av olika anledningar.  I det här projektet har en självnavigerande hinderupptäckande robot byggts. Roboten använder ultraljudssensorer för att upptäcka hinder och unvika dem. En algoritm för navigationen av roboten skapades och implementerades i Arduinon. För drivningen av hjulen användes två servomotorer. Roboten hade tre hjul, två i den bakre änden till vilka servomotorerna var fästa och ett länkhjul fram. Det möjliggjorde differentialstyrning vilket också tillät snabba och snäva svängar.  Tester genomfördes som visade att roboten kunde navigera i ett rum med olika hinder utplacerade utan större problem. Placeringen av sensorerna fungerade bra med tanke på det antal sensorer som användes. Förbättringar av hinderupptäckningen hade kunnat göras om fler sensorer hade använts. Testerna bekräftade också att ultraljudssensorer fungerar bra för denna typ av uppgift.

The focus of this study was deep learning. A small, autonomous robot car was used for obstacle avoidance experiments. The robot car used a camera for taking images of its surroundings. A convolutional neural network used the images for obstacle detection. The available dataset of 31 022 images was trained with the Xception model. We compared two different implementations for making the robot car avoid obstacles. Mapping image classes to steering commands was used as a reference implementation. The main implementation of this study was to separate obstacle detection and steering logic in different modules. The former reached an obstacle avoidance ratio of 80 %, the latter reached 88 %. Different hyperparameters were looked at during training. We found that frozen layers and number of epochs were important to optimize. Weights were loaded from ImageNet before training. Frozen layers decided how many layers that were trainable after that. Training all layers (no frozen layers) was proven to work best. Number of epochs decided how many epochs a model trained. We found that it was important to train between 10-25 epochs. The best model used no frozen layers and trained for 21 epochs. It reached a test accuracy of 85.2 %.

No description available.

Off-policy reinforcement learning is useful in many contexts. Maei, Sutton, Szepesvari, and others, have recently introduced a new class of algorithms, the most advanced of which is GQ(lambda), for off-policy reinforcement learning. These algorithms are the first stable methods for general off-policy learning whose computational complexity scales linearly with the number of parameters, thereby making them potentially applicable to large applications involving function approximation. Despite these promising theoretical properties, these algorithms have received no significant empirical test of their effectiveness in off-policy settings prior to the current work. Here, GQ(lambda) is applied to a variety of prediction and control domains, including on a mobile robot, where it is able to learn multiple optimal policies in parallel from random actions. Overall, we find GQ(lambda) to be a promising algorithm for use with large real-world continuous learning tasks. We believe it could be the base algorithm of an autonomous sensorimotor robot.

Thesis (MScEng)--Stellenbosch University, 2012. / ENGLISH ABSTRACT: Simultaneous localization and mapping (SLAM) is vital for autonomous robot navigation. The robot must build a map of its environment while tracking its own motion through that map. Although many solutions to this intricate problem have been proposed, one of the most prominent issues that still needs to be resolved is to accurately measure and track landmarks over time. In this thesis we investigate the use of stereo vision for this purpose. In order to find landmarks in images we explore the use of two feature detectors: the scale-invariant feature transform (SIFT) and speeded-up robust features (SURF). Both these algorithms find salient points in images and calculate a descriptor for each point that is invariant to scale, rotation and illumination. By using the descriptors we match these image features between stereo images and use the geometry of the system to calculate a set of 3D landmark measurements. A Taylor approximation of this transformation is used to derive a Gaussian noise model for the measurements. The measured landmarks are matched to landmarks in a map to find correspondences. We find that this process often incorrectly matches ambiguous landmarks. To find these mismatches we develop a novel outlier detection scheme based on the random sample consensus (RANSAC) framework. We use a similarity transformation for the RANSAC model and derive a probabilistic consensus measure that takes the uncertainties of landmark locations into account. Through simulation and practical tests we find that this method is a significant improvement on the standard approach of using the fundamental matrix. With accurately identified landmarks we are able to perform SLAM. We investigate the use of three popular SLAM algorithms: EKF SLAM, FastSLAM and FastSLAM 2. EKF SLAM uses a Gaussian distribution to describe the systems states and linearizes the motion and measurement equations with Taylor approximations. The two FastSLAM algorithms are based on the Rao-Blackwellized particle filter that uses particles to describe the robot states, and EKFs to estimate the landmark states. FastSLAM 2 uses a refinement process to decrease the size of the proposal distribution and in doing so decreases the number of particles needed for accurate SLAM. We test the three SLAM algorithms extensively in a simulation environment and find that all three are capable of very accurate results under the right circumstances. EKF SLAM displays extreme sensitivity to landmark mismatches. FastSLAM, on the other hand, is considerably more robust against landmark mismatches but is unable to describe the six-dimensional state vector required for 3D SLAM. FastSLAM 2 offers a good compromise between efficiency and accuracy, and performs well overall. In order to evaluate the complete system we test it with real world data. We find that our outlier detection algorithm is very effective and greatly increases the accuracy of the SLAM systems. We compare results obtained by all three SLAM systems, with both feature detection algorithms, against DGPS ground truth data and achieve accuracies comparable to other state-of-the-art systems. From our results we conclude that stereo vision is viable as a sensor for SLAM. / AFRIKAANSE OPSOMMING: Gelyktydige lokalisering en kartering (simultaneous localization and mapping, SLAM) is ’n noodsaaklike proses in outomatiese robot-navigasie. Die robot moet ’n kaart bou van sy omgewing en tegelykertyd sy eie beweging deur die kaart bepaal. Alhoewel daar baie oplossings vir hierdie ingewikkelde probleem bestaan, moet een belangrike saak nog opgelos word, naamlik om landmerke met verloop van tyd akkuraat op te spoor en te meet. In hierdie tesis ondersoek ons die moontlikheid om stereo-visie vir hierdie doel te gebruik. Ons ondersoek die gebruik van twee beeldkenmerk-onttrekkers: scale-invariant feature transform (SIFT) en speeded-up robust features (SURF). Altwee algoritmes vind toepaslike punte in beelde en bereken ’n beskrywer vir elke punt wat onveranderlik is ten opsigte van skaal, rotasie en beligting. Deur die beskrywer te gebruik, kan ons ooreenstemmende beeldkenmerke soek en die geometrie van die stelsel gebruik om ’n stel driedimensionele landmerkmetings te bereken. Ons gebruik ’n Taylor- benadering van hierdie transformasie om ’n Gaussiese ruis-model vir die metings te herlei. Die gemete landmerke se beskrywers word dan vergelyk met dié van landmerke in ’n kaart om ooreenkomste te vind. Hierdie proses maak egter dikwels foute. Om die foutiewe ooreenkomste op te spoor het ons ’n nuwe uitskieterherkenningsalgoritme ontwikkel wat gebaseer is op die RANSAC-raamwerk. Ons gebruik ’n gelykvormigheidstransformasie vir die RANSAC-model en lei ’n konsensusmate af wat die onsekerhede van die ligging van landmerke in ag neem. Met simulasie en praktiese toetse stel ons vas dat die metode ’n beduidende verbetering op die standaardprosedure, waar die fundamentele matriks gebruik word, is. Met ons akkuraat geïdentifiseerde landmerke kan ons dan SLAM uitvoer. Ons ondersoek die gebruik van drie SLAM-algoritmes: EKF SLAM, FastSLAM en FastSLAM 2. EKF SLAM gebruik ’n Gaussiese verspreiding om die stelseltoestande te beskryf en Taylor-benaderings om die bewegings- en meetvergelykings te lineariseer. Die twee FastSLAM-algoritmes is gebaseer op die Rao-Blackwell partikelfilter wat partikels gebruik om robottoestande te beskryf en EKF’s om die landmerktoestande af te skat. FastSLAM 2 gebruik ’n verfyningsproses om die grootte van die voorstelverspreiding te verminder en dus die aantal partikels wat vir akkurate SLAM benodig word, te verminder. Ons toets die drie SLAM-algoritmes deeglik in ’n simulasie-omgewing en vind dat al drie onder die regte omstandighede akkurate resultate kan behaal. EKF SLAM is egter baie sensitief vir foutiewe landmerkooreenkomste. FastSLAM is meer bestand daarteen, maar kan nie die sesdimensionele verspreiding wat vir 3D SLAM vereis word, beskryf nie. FastSLAM 2 bied ’n goeie kompromie tussen effektiwiteit en akkuraatheid, en presteer oor die algemeen goed. Ons toets die hele stelsel met werklike data om dit te evalueer, en vind dat ons uitskieterherkenningsalgoritme baie effektief is en die akkuraatheid van die SLAM-stelsels beduidend verbeter. Ons vergelyk resultate van die drie SLAM-stelsels met onafhanklike DGPS-data, wat as korrek beskou kan word, en behaal akkuraatheid wat vergelykbaar is met ander toonaangewende stelsels. Ons resultate lei tot die gevolgtrekking dat stereo-visie ’n lewensvatbare sensor vir SLAM is.

The thesis deals with design of the device for testing the localization techniques for indoor navigation. Autonomous robot was designed as the most appropriate platform for testing. The thesis is divided into three parts. The first one describes various kinds of robots, their possible use and sensors, which could be of use for solving the problem. The second part deals with the design and construction of the robot. The robot is built on the chassis of the differential type with support spur. Two electric motors with a gearbox and output shaft speed sensor represent the drive unit. Coat of the robot was designed for good functionality and attractive overall look. The robot is also used for the presentation of robotics. Thesis provides complete design of chassis and body construction, along with control section and sensorics. The last part describes a statistical model of the robot movement, which was based on several performed experiments. The experiments were realized to find any possible deviations of sensor measurement comparing to the real situation.

The goal of this diploma thesis is a proposal of autonomous two wheeled balancing robot, differentially driven. This kind of robot is especially suitable in confined space, where it can utilize its maneuver skills. Many criteria as operational conditions, materials, size and weight of the robot, suitable hardware and sensors must to be considered, when designing the robot. Development and implementation of autonomous balancing control system is also part of the thesis.