In recent years large advances have been made in the field of machine learning, driven by novel deep learning methods. Deep learning is a research field that focusses on creating neural networks. This field has seen a rapid advance due to an increase in computational power, availability of large amounts of data and a wide variety of novel methods that allows for more efficient training of neural networks. Deep learning has been applied in various fields to solve many different tasks. Effective training of these neural networks requires selecting the right data, network architecture and learning method. However, thorough understanding of the task for which the neural network is trained is needed to adhere to these requirements. This thesis will illustrate that deep learning methods can effectively be applied to perception tasks by thorough understanding of the task.

The subject of this thesis is to investigate whether the Cerebellar Model Articulation Controller (CMAC) can be used to anticipate controller corrections and increase performance by reducing delays in humanoid robots. This question can be divided into two subquestions. Firstly, whether the CMAC is a suitable architecture for the prediction of controller actions for a humanoid soccer robot. Using a 2D model of a robotic leg, the results of this thesis show that the CMAC can indeed learn to anticipate a corrective control signal 30 ms ahead. Secondly, whether the architecture of the aforementioned setup can increase the performance of adequately passing a ball by reducing delays. The experiments show that the use of a CMAC can increase the performance of the robotic setup.

To be able to understand the dynamic driving environment, an autonomous vehicle needs to predict the mo- tion of other traffic participants in the driving scene. Motion prediction can be done based on experience and recently observed series of past events, and entails reasoning about probable outcomes with these past ob- servations. Aspects that influence driving behavior comprise many factors, such as general driving physics, infrastructure geometry, traffic rules, weather and so on. Models that are used today are incapable of including many of these aspects. These deep learning models often rely heavily on the past driven trajectory as an information source for future prediction. Sparsely, some work has been done to include interaction between vehicles or some infrastructural features to the model. The lack of information regarding the driving scene can be seen as a missed opportunity, because it is a use- ful feature source to predict the motion of a vehicle. Especially when a map of the driving scene is available, reliable information regarding the road layout can be extracted. In this thesis, a model architecture is sought that is aware of the interaction between vehicles, and that un- derstands the geometry of the roads in the scene. Importantly, map information regarding the driving scene should be used as the primary source of information regarding the road geometry. First, a baseline deep learning model is constructed, that can generate predictions based on the past observed trajectory. To add interaction features to the model, a social pooling module is introduced. The social pooling method allows to efficiently include the driving behavior of other vehicles in the scene. To introduce reliable, map-based road information to the model, two novel methods are proposed. In the first, the predictions from a model are used to extract features in a map. These features describe the road scene around the predicted location, and are used to update these predictions. In the second method, the semantic map is only used to extract a road segment ahead of the vehicle. A road-RNN is introduced to con- struct features regarding the road segment, and an attention mechanism to determine what part of the road segment is relevant for the predictions. These modules are referred to as road-refinement and road-attention respectively. The importance of including both road geometry and interaction methods in the model is shown by con- structing 5 different models that vary in their road-geometric and interaction awareness. A baseline deep learning model is used and extended with a road-geometry module, an interaction module, a combination of the two, or none. To test the prediction capabilities of the models, they are trained on two different datasets. The first dataset, called i80, consists of trajectory recordings from a straight highway with dense traffic. The other dataset is a curved version of the i80, called i80c, where the trajectories and road are transformed to introduce road-geometric variations in the data. The prediction performances from the models clearly show the importance of both interaction-aware and road-geometry aware modeling. The road-refinement and road-attention models outperform the road-agnostic baseline model on the i80c data, showing better understanding of the road layout. Comparison between these models learns that the road-attention model is more effective for road-geometry inclusion. Additionally, the performance increase for interaction-aware models compared to the baseline clearly shows the importance of interaction in modeling on both datasets.The model that combines the interaction and road-attention modules shows outstanding prediction performance on the challenging i80c dataset compared to all other models. From the predictions it can be seen that the model understands the road layout ahead of the subject vehicle, as well as the interactive forces that are in play.

Our world is rapidly changing, leading to various technological achievements in diverse areas of expertise. Among the most evolving are autonomous robots, machines and vehicles, enormously extending human capabilities. While this trend became quite familiar for land-based vehicles, innovations in the underwater domain lag behind, mainly due to the harsh environmental conditions encountered. This in large contrast with worldwide growing needs for autonomous underwater solutions, like dredging. Nevertheless, for underwater vehicles becoming autonomously, various technological challenges must be overcome. The most essential one is finding a proper answer to the question; where in the underwater world am I? Autonomous robots are usually be able to operate in complex environments using external reference systems such as Global Positioning System (GPS) to locate themselves inside their environment. However, these are not accessible in underwater applications, since water strongly attenuates electromagnetic signals, notably GPS. Hence, various alternative methods for Simultaneous Localization And Mapping (SLAM) for underwater applications are employed, among which the method of anchoring inertial measurements on environmental landmarks, perceived with an exteroceptive Sonar. The ultimate goal of this procedure is to correct for errors in the inertial measurements, using the environmental perception available in the Sonar scans. Nevertheless, due to practical underwater issues, it still is not readily available and even barely investigated in open scientific research. This thesis aims at the development of an underwater localization algorithm, using onboard inertial sensors and Sonar. The overall system design consists of several individual software procedures, executing certain assignments, together ultimately solving for the localization task. The developed system was tested using ground truth in form of a real-world dataset obtained several years ago in an abandoned marina. This showed that most individual procedures were able to provide good results, while having a low computational complexity in relation to online operations. However, it is quite a challenge to obtain results accurate enough to properly correct for positional errors, by using the designed system. Although showing potential, it still is not entirely decent in terms of accuracy and computational burden. This is likely due to the sparsity of the extracted Sonar measurements. Data is lacking to further verify the system, hence several questions remain open regarding its universal applicability in case of different scenarios, dynamic environments and long trajectories. In this context, more in-depth research is needed into the algorithms that are influenced by the Sonar measurements, to increase the accuracy of the overall system, while gaining more insight in its computational behaviour. Additionally, also more real-world experiments are essential, to extensively verify the designed system. In conclusion, the results showed that it is possible to correct for inertial errors, using the system developed. Furthermore, it demonstrates that most procedures individually produce good results in real-time. Hence, this study can be seen as a positive step in the right direction, forming a basis for future research in solutions that are generically applicable in online real underwater world operations.

Facial expression recognition today is a widely-researched topic with some pertinent applications in human-computer interaction (HCI), surveillance, etc. The diverse range of expressions makes data collection expensive in terms of time and money, making task-specific data collection inconvenient. This thesis project investigates a possible method for quickly and cheaply collecting this data, especially for a Human Computer Interaction (HCI) application. In particular, the focus is on \textit{Pepper}, a robot meant for interacting with humans through conversation. For collecting such data, emotions should be triggered in participants and their faces should be video-recorded. The triggering method used was to show videos selected for triggering specific emotions, letting participants watch them in pairs, thus enabling mutual interactions to better simulate the social environment in which the robot will operate. After watching each video, participants were asked to rate their feelings through the \textit{AffectButton}, a tool for intuitively describing emotions in a dimensional way. Video selection was based on a questionnaire in which people were asked to rate emotions a video was triggering in them. Recordings obtained from people watching the triggering video were then compared to a model of a neutral expression performed by the same participant, in order to select the frames in which expressions were shown, ignoring neutral and transition frames. The pictures obtained were included in a dataset, which is used for training a linear regressor and a Convolutional Neural Network (CNN). These were then tested on naturalistic data taken during conversations, in order to investigate whether the proposed data collection method could build a useful dataset and the results showed this method to be promising.

Recent advancements in computation power and artificial intelligence have allowed the creation of advanced reinforcement learning models which could revolutionize, between others, the field of robotics. As model and environment complexity increase, however, training solely through the feedback of environment reward becomes more difficult. From the work on robotic priors by
R.Jonschkowski et al. we present robotic auxiliary losses for continuous reinforcement learning models. These function as additional feedback based on physics principles such as Newton’s laws of motion, to be utilized by the reinforcement learning model during training in robotic environments. We furthermore explore the issues of concurrent optimization on several losses and present a continuous loss normalization method for the balancing of training effort between main and auxiliary losses. In all continuous robotic environments tested, individual robotic auxiliary losses show consistent improvement over the base reinforcement learning model. The joint application of all losses during training however did not always guarantee performance improvements, as the concurrent optimization of several losses of different nature proved to be difficult.

In computer vision pose estimation of objects in everyday scenes is a basic need for a clearunderstanding of the surrounding environment, fields of interests include; augmented reality,surveillance, navigation, manipulation, and robotics in general. Pose estimation is a wellstudied topic, however fast and robust solutions are still hard to obtain. The goal of thisresearch is to robustly and efficiently perform 3D pose estimation of multiple objects withina single RGB image in real-time (>24framespersecond (fps)). To achieve this goal an existing CNN is utilized, more specifically the YOLO network is used.This provides a stable platform for object detection and classification, the network is onlyslightly modified to include pose regression. The YOLOv2 network was originally designedto generate bounding boxes around objects and classify the objects within the boundingboxes. This research shows that using a single confidence value rather than 4 boundingbox parameters is sufficient to determine the relative location of objects within the image,limiting the number of parameters that need to be trained. This has allowed to make thenetwork more efficient and to make the network focus more on training the pose parameters(azimuth,elevation and distance) rather than the bounding box parameters. Using several techniques like data augmentation, data clustering and data selection the state-of-the-art AVP of 50.1% was achieved on the azimuth estimation problem. For the full 3Dpose (azimuth, elevation and distance) problem the AVP is limited to 30.1%, with no directcomparison this is still considered to be state-of-the-art. However, normalizing the confidenceoutput of each image in the post-processing step has increased this accuracy further, improvingbeyond the state-of-the-art results. With a normalization step the AVP score has reached 63.0% and 40.4% respectively. These results show that the pose estimation problem hasimproved significantly, getting closer to a viable solution for real-world applications. However, further improvements can be made in the post-processing step, the data augmenta-tion step and the data selection step. The research conducted has shown that accuracy gainsare not only achieved through better network architecture but are also highly dependent onthe training and processing techniques used. This is evident from the accuracy increase of 38% from an AVP of 25% to an AVP of 63%. Optimizing these techniques specifically for theYOLOposearchitecture, might result in a solution that can be used in real-world applications.

Task-parameterized movement representation, as an approach for the generalization of demonstrations, is used to represent data from multiple local perspectives within the global reference frame, through which more accurate information about multiple aspects of the movement is given. The estimated transformation between the different perspectives and the global reference frame in task parameter inference can be used for gesture recognition.
In this thesis, task parameter inference in the application of human-robot interaction, a method called TP-inference approach, is investigated. It consists of a combination of task parameter inference and task parameter movement retrieval. A task-driven model is used to generalize the demonstration data and the task parameter inference is achieved by using the orthogonal Procrustes analysis. The TP-inference approach is tested for various static tasks and is compared to the Probabilistic Movement Primitive (ProMP) approach [1]. The test results indicate that for simple and or similar movement of the human and robot, the TP-inference approach performs less accurate than the ProMP. For complex movements the TP-inference preforms more accurate than the ProMP.

[1] M. Ewerton, G. Neumann, R. Lioutikov, H. B. Amor, J. Peters, and G. Maeda, Learning multiple collaborative tasks with a mixture of interaction primitives, in 2015 IEEE International Conference on Robotics and Automation (ICRA) (IEEE, 2015) pp. 1535–1542

Parts of the Southwest Palace of Sennacherib (700 BC), which were located in Nineveh (Iraq), are now destroyed by the ongoing conflict in Iraq and Syria. The palace rooms used to be decorated with numerous Assyrian reliefs. Luckily, digital pictures of the site are available thanks to an Italian expedition in 2002. The goal of this research is to physically reproduce the destroyed reliefs using this photo database. First photogrammetry was used to rebuild
the global dimensions of the reliefs. The details are reconstructed by fusing a highly detailed depth map retrieved with Shape from Luminance (SFL), together with a coarser Artificial Reconstruction. The fused reconstruction was evaluated with the eigenvalues of the Hessian. The eigenvalues give an intuitive measure about the shapes of the reliefs. The RMSE of the results improve up to 44.4% compared to the usage of an intensity (grey) image. The final 3D models are used to produce full size reproductions with 3D printing and CNC milling.

An autonomous vehicle should be able to operate amidst numerous other human-driven vehicles, each driving on its own trajectory. To safely navigate in such a dynamic environment, the autonomous vehicle should be able to predict trajectories of the vehicles operating in its vicinity and use these to plan its own path. Most related work uses a vehicle's past trajectory to model its behavior, based on which the future trajectory is predicted. However, they do not focus on the influence of contextual features such as road structure from the scene that may affect the vehicle's future trajectory. This work proposes an approach to predict a long-term vehicle trajectory using not only the past trajectory of a vehicle but also contextual features from the driving scene. We model the road structure to help prediction on curved road sections. A Recurrent Neural Network is used to learn vehicle behavior from past vehicle trajectories and predict future trajectories while incorporating road structure. Using a trajectory dataset collected from a test vehicle, we compare our model's performance with the conventional prediction approach based on only past vehicle trajectory.

Many robotic systems rely on laser ranging sensors to navigate and map their environment. As robots are getting more advanced, smaller and cheaper, their sensors need to decrease in size and cost, while increasing their reliability and accuracy. Current laser-based sensors have difficulty to meet these properties. They are either large, expensive, contain moving parts or provide insufficient amounts of information.

This thesis introduces an integrated approach for the development of an eye-safe, low-cost, solid-state, wide-angle line-laser distance sensor, tackling these challenges. A prototype is derived by constructing a generic, camera independent, triangulation model. The model, combined with a set of pre-defined requirements, is used to select hardware components and predict the sensor limits. A unified calibration step is proposed to estimate both camera intrinsics as well as misalignment errors with the same set of data. Additionally, a microsecond-accurate open loop synchronization system for laser activation and imager exposure is presented.

Tests show that the prototyped sensor is able to measure distances up to 3 meters with an error of 4%. At 2 meter, the error is just under 2%. Additionally, the solid-state prototype has a field-of-view of 105 degrees, an angular resolution of 0.8 degrees, an update rate of 10Hz and an estimated cost of just below $35.

This thesis presents a visual route following method that minimizes memory consumption to the point that even Micro Aerial Vehicles (MAV) equipped with only a simple microcontroller can traverse distances of a few hundred meters. Existing Simultaneous Localization and Mapping (SLAM) algorithms are too complex for use on a microcontroller. Instead, the route is modeled by a sequence of snapshots that can be followed back using a combination of visual homing and odometry. Three visual homing methods are evaluated to find and compare their memory efficiency. Of these methods, Fourier-based homing performed best: it still succeeds when snapshots are compressed to less than twenty bytes. Visual homing only works from a small region surrounding the snapshot, therefore odometry is used to travel longer distances between snapshots. The proposed route following technique is tested in simulation and on a Parrot AR.Drone 2.0. The drone can successfully follow long routes with a map that consumes only 17.5 bytes per meter.