Automated design synthesis of perception systems for industrial applications

The capability to perceive the working environment is crucial for autonomous and automated systems. For instance, an assembly robot needs to know accurately the position of the assembly parts in order to grasp them successfully. However, the design synthesis of perception systems is associated with high engineering efforts due to a large number of interdependent design choices and high reliability and accuracy requirements. This is especially problematic for flexible assembly system, where an effortless adaptation to new products is necessary. The space of possible system configurations is large and the solution space comparatively small. Noise and uncertainty are immanent in the data and have a strong effect on the system performance.

In this thesis, models and methods are introduced and investigated in order to automate the perception system design process for flexible assembly systems. First, a common representation grounded in set-theory is introduced, which allows to represent procedural and declarative models for perception systems hierarchically, i.e. on different levels of abstraction. The hierarchical structure allows to represent different levels of computationally less demanding approximations of the models, which facilitates an efficient exploration of the space of system configurations. Based on the common representation a design synthesis method is introduced, which allows to jointly optimize the structure and parameterization of perception pipelines. Finally, a synthesis method based on hierarchical planning is introduced. This hierarchical approach increases the synthesis efficiency and enables the runtime adaption and failure handling of perception system. The contributions are validated on the example of industrial assembly applications in simulated and real-world environments.