Ronald Garcia

Ronald Garcia

Biography

Ronald Garcia is a first-generation Mechanical, Computer, and Electrical engineering student at Johnson Hopkins University. He is Cuban-American, and grew up in Miami,Florida in a limited income home. His interest in engineering was possible as a result of the support system he’s built at Hopkins. This support sparked his interest in playing a vital role in education, mentorship, and teaching. As a tutor, he experienced first-hand that a key component to solving real-world problems is viewing these problems through diverse lenses. He is applying his diverse experiences to problems in the field of robotics, striving to translate academic insights into tangible ways to give back. Through his interdisciplinary work and experiences, he strives to create environments where others feel empowered to learn beyond the classroom. Above all, he believes that nurturing curiosity and sharing knowledge can create lasting and positive change in individuals, communities, and the world.

Abstract

Smart Testing: How to Adaptively Find Failures with ECI

Testing of Autonomous Driving Systems (ADS) is a vital step in ensuring the on-road safety of surrounding vehicles and Vulnerable Road Users (VRUs), such as bicyclists, pedestrians, and vehicle occupants. To determine if an ADS is safe, the Vehicle Under Test (VUT) undergoes hardware and simulation testing on standardized scenarios designed based on recommendations from regulatory testing authorities such as Euro New Car Assessment Programme. While these scenarios are hand-designed from historical accident reports, they are low-dimensional and insufficient in determining the effectiveness of a particular ADS. We propose an innovative online testing paradigm that uses the model performance of the VUT as feedback to create adaptive testing scenarios. This method uses Gaussian Processes to develop a surrogate model of the VUT testing environment. Using MATLAB’s AEB test bench, we demonstrate that this method can tackle higher-dimensional testing scenarios, explore the parameter space, and cover rare failure modes by identifying new test points based on their expected coverage improvement (ECI). This improvement can be quantified with either an objective or subjective metric, allowing for abstract failure modes. Future work will integrate our method with a higher-level optimization on the parameter space to further optimize failure mode coverage with low data.