Simplifying the Control of Mobile Robots through Image-based Behaviors

The development and commissioning of mobile robots is usually a time-consuming and cost-intensive undertaking. Today, systems are realized primarily through the use of complex software and hardware architectures, which only partially meet modern requirements for mobile robots. The core discipline that a mobile robot must fulfill is the correct perception of its environment. To achieve this, various technologies are used that are useful for detecting obstacles or navigating targets. An increasingly attractive technology in this domain are artificial neural networks. Existing literature focuses on either improving the performance of existing systems or on training for very specialized applications. In the field of mobile robotics, the focus is usually on realizing a specific task, while little attention is paid to generalizability, cost and energy constraints. To fill this gap, this paper investigates the possibility of reducing the setup for a mobile robot application to a minimum while still enabling complex behaviors. For the implementation, we take a biological inspired approach and investigate the usability of artificial neural networks, in this case YOLOv4, in a mobile robot application. In particular, we examine whether the currently available technologies meet todays requirements.