Collaborative Research: SHF: Medium: Models and Analyses for Robotic and Autonomous Systems Operating in Complex and Unpredictable Real-world

University Of Virginia Main Campus, Charlottesville VA

Investigators

Abstract

Robotic and autonomous systems (RAS) are becoming more common in our daily lives. Making sure they operate safely, however, remains challenging. RAS code is often messy, and the real world in which these systems operate is complex and unpredictable. This research project focuses on creating a family of methods to find faults in these sophisticated systems by analyzing: 1) the program artifacts they are built from to extract models that can be more easily checked for faults, and 2) the real-world data they encounter (like images taken by cameras) to determine what portions of the environment are worth simulating more accurately. If successful, the research findings will inform solutions to challenges faced in the development of RAS, affording a path to improving public safety. The research will be integrated into educational curriculum and training. The key insights enabling the project are two-fold. First, although robotics code is often messy and complex, model-relevant behavior is typically implemented via a subset of application programming interfaces and configuration files with clearly-specifiable semantics. Second, field data encodes key spatial-temporal physical constraints imposed by the real world, which provides hints on how to steer simulation to reduce the gap with reality. The investigators will leverage these insights to effectively lift useful models from real code to detect compositional faults, identify and construct simulation scenarios that capture constraints imposed by the real world, and inform and validate the evolution of robotic systems. The proposed work will produce: 1) Techniques to infer rich behavior models with physical attributes from RAS code and artifacts, and to check those models against component and system properties; 2) Techniques to infer specifications from messy real-world data, to contrast those specifications against simulation states, and to synthesize missing simulation environments; 3) Techniques to inform RAS' evolution, to understand the impact of code, physical, and world changes, and to cost-effectively correct and test changing RAS; and 4) Studies with RAS to assess the techniques, quantifying their effect on the gaps between code and models, between simulation and reality, and during evolution. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.