Date of this Version
Ore, John-Paul "Dimensional Analysis of Robot Software without Developer Annotations." Doctoral dissertation, University of Nebraska--Lincoln, 2019.
Robot software risks the hazard of dimensional inconsistencies. These inconsistencies occur when a program incorrectly manipulates values representing real-world quantities. Incorrect manipulation has real-world consequences that range in severity from benign to catastrophic. Previous approaches detect dimensional inconsistencies in programs but require extra developer effort and technical complications. The extra effort involves developers creating type annotations for every variable representing a real-world quantity that has physical units, and the technical complications include toolchain burdens like specialized compilers or type libraries.
To overcome the limitations of previous approaches, this thesis presents novel methods to detect dimensional inconsistencies without developer annotations. We start by empirically assessing the difficulty developers have in making type annotations. In a human study of 83 subjects, we find that developers are only 51% accurate and require more than 2 minutes per annotation. We further find that type suggestions have a significant impact on annotation accuracy. We find that when showing developers annotation suggestions, three suggestions are better than a single suggestion because they are as helpful when correct and less harmful when incorrect. Since developers struggle to make type annotations accurately, we present a novel method to infer physical unit types without developer annotations. This is novel because it is the first method to detect dimensional inconsistencies in ROS C++ without developer annotations, and this is important because robot software and ROS are increasingly used in real-world applications. Our method leverages a property of robotic middleware architecture that reuses standardized data structures, and we implement our method in an open-source tool, Phriky. We evaluate our method empirically on a corpus of 5.9 M lines of code and find that it detects real inconsistencies with an 87% TP rate. However, our method only assigns physical unit types to 25% of variables, leaving much of the annotation space unaddressed. To overcome these limitations, we extend our method to utilize uncertain evidence in identifiers using probabilistic reasoning. We implement our new probabilistic method in a tool Phys and find that it assigns units to 75% of variables while retaining a TP rate of 82%. We present the first open dataset of dimensional inconsistencies in open-source robotics code, to our knowledge. Lastly, we identify extensions to our work and next steps for software tool developers to build more powerful robot software development tools.
Advisers: Sebastian Elbaum and Carrick Detweiler