Disputation Sebastian Buck
am Mittwoch, 9. Mai 2018, um 16:00 Uhr in Raum A 104, Sand 1
Modeling Robotic Systems with Activity Flow Graphs
Berichterstatter 1: Prof. Dr. Andreas Zell
Berichterstatter 2: Prof. Dr. Andreas Schilling
Kurze Zusammenfassung des Vortrags
Autonomous robotic systems are becoming increasingly common in our society, with research efforts towards automated goods transportation, service robots and autonomous cars. These complex systems have to solve many different problems in order to function robustly. Two especially important areas of interest are perception and high level control. Intelligent systems have to perceive their surroundings in order to facilitate autonomy. With an understanding of the environment, they then can make their own decisions based on high level control policies defined by the developers.
Robotic systems differ drastically in their sensory capabilities, their computational power, and their designated tasks. When developing algorithms, however, we need to have a common modeling framework that enables us to generalize and re-use existing solutions. A modular approach, which is coherent across different platforms, also allows faster prototyping of new systems. In this dissertation we develop a modeling framework based on data flow that achieves this goal.
We first extend the existing Synchronous Data Flow (SDF) model and combine it with reactive programming ideas and finite-state machines. Together, these existing frameworks enable us to model many aspects of complex robotic systems. We apply this model to a robot in a warehouse scenario to demonstrate the viability of the approach.
Using three disjoint formalisms to model a robotic system has many downsides. In a first unification step we merge SDF and reactive programming into Hybrid Flow Graphs (HFGs), where we explicitly model synchronous and asynchronous data flow. We then apply the HFG model to the perception system of an autonomous transportation robot.
In a last step, we eliminate the need for separate finite-state machines by introducing the concept of activity into the data flow. We therefore unify the different aspects into a single and coherent framework which we call Activity Flow Graphs (AFGs). The flow of activity enables us to model high level state directly in the data flow graph. The result is a single computation graph that can express both perception and high level control aspects of any robotic system. We then demonstrate this with multiple high level robotic system models.
Finally, we make use of the uniform AFG model to provide a single graphical user interface that allows a developer to rapidly prototype complete robotic systems. Since all aspects of a robot can be implemented using the same theoretical framework, there is no need to switch between different paradigms. The user interface is designed to give immediate feedback, which speeds up prototyping, testing and evaluation, as well as debugging when working with real robots.