Modular Computation Models for Robotic Perception and Monte Carlo Optimization

In many situations of our everyday lives, robots increasingly gain in importance. The application scenario, the resulting design and construction decisions and finally the system size determine how many computational resources and sensorial equipment are available in such a robotic system. These can be constrained in many ways, for instance either by low power requirements or the necessity to execute several tasks in parallel. For the purpose of modeling robotic behavior based on perception, we shortly introduce a modular framework which has been extensively used in our research. Based on this framework and in the context of a research project, we develop a lightweight 3D aided person detection approach, which is then extended for usage with different sensors and on different robots. While nowadays detection and classification tasks are carried out using deep neural network based classifiers requiring a huge amount of computational power, we present a modular real-time capable pipeline which can be executed on robotic systems with limited resources. Perception, detection and classification are only a few basic components required to implement a robotic application scenario. Hence, we further elaborate on robot localization beginning with an approach which can be used within buildings using CAD floor plans, a rather problem specific solution. Inspired by our modular data processing framework, with the goal in mind to improve on the research process itself, we further developed a more generic lightweight solution for Monte Carlo based optimization. We use this approach, alongside different map representations and according model functions, to further improve on particle filter based robot localization. Finally, we employ the generic framework in a completely different problem, namely contact point detection on robot manipulators.