Autonomous System Edge Cases: Implementing a Reinforcement Learning Pipeline for Complex Synthetic Road Environment Images; Advancing U.S. Autonomous Vehicle Regulations: Insights from Global Frameworks and Innovative Methodologies

The growing integration of artificial intelligence (AI) into transportation systems promises to redefine safety and efficiency in profound ways. This was made very clear to me after my internship this past summer which is what the contents of my capstone technical report consists of. At my internship, I was tasked with the development of a reinforcement pipeline motivated by the desire to generate synthetic image data to possibly improve vehicular AI perception systems. My internship was a great experience and showed me how more research and care needs to be put in the creation of autonomous vehicles (AV). My experience over the summer also largely influenced my decision on what my STS research topic would be, which was to address and explore the weakness of AV regulatory frameworks for these innovative systems in the United States. Together, these endeavors illuminate how a sociotechnical perspective—one that unites engineering innovation with ethical and societal considerations—can enable more responsible technological progress.
My capstone technical report is on the project that I spent this past summer developing in my internship. The main goal of my project focuses on developing a reinforcement learning pipeline to generate synthetic edge cases for AV perception systems. Edge cases, which are rare, complex scenarios such as obscured pedestrians or unfamiliar road objects, present significant challenges for AI perception. During my internship, I worked with a small team to design a pipeline capable of creating such scenarios using text-to-image generation models like Stable Diffusion, reinforced by optimization techniques such as Proximal Policy Optimization (PPO). My contributions involved constructing the pipeline's architecture, integrating a reward function to evaluate the quality of generated scenarios, and training a language model (LLM) for prompt optimization. Although I did not complete the training phase due to the end of the internship timeline, my efforts streamlined the workflow for future iterations and for other teams to continue development, promising a tool that could significantly enhance AV training and, ultimately, roadway safety.
My STS research, unlike many of my peers, was actually disjoint from my prospectus. My prospectus related heavily to the integration of AI and the internet of things into water treatment systems, but I did not feel enthusiastic about this, so it was not continued as my STS research. Despite this change, my prospectus offered me great lessons in how to gather sources and harvest information; it also showed me the true purpose behind conducting STS research through my lack of understanding in the prior semester. Thankfully, my experience with my internship allowed me to shift my focus and synthesize my current STS research topic in which I examined the regulatory challenges facing AV deployment in the U.S. Despite their potential to reduce traffic-related fatalities, AVs remain under regulated in the U.S., leaving gaps in safety oversight and public trust. Drawing from frameworks in Germany and Singapore, I explored how centralized regulations, like Germany’s comprehensive AV legislation, and experimental models, such as Singapore’s regulatory sandboxes, could inform U.S. policy. I also applied Claudia Schwarz-Plaschg’s concept of analogical reasoning to relate AV governance to historical examples of technology regulation, offering insights into proactive, adaptive policymaking. By incorporating crash-adaptive models that leverage real-time data, I proposed a dynamic regulatory approach to address AV challenges and enhance public trust.
Reflecting on these projects highlights the profound value of a sociotechnical perspective. My technical work during my internship demonstrated that AVs, which are often perceived as the epitome of reliability and technological advancement, still struggle with fundamental challenges, particularly when faced with edge cases. This realization demonstrated the importance of my STS research, which showed that these shortcomings come from the weak regulatory framework governing AVs in the United States. In addition to this, viewing these two projects together highlighted the need to bridge the technical and non-technical realms. While my internship focused on developing tools to address technical limitations, my STS research explored how laws, regulations, and public policies could complement such advancements by making sure these technologies are deployed responsibly. This dual focus demonstrated the importance of an STS perspective in identifying and addressing issues that might otherwise go unnoticed. Without this lens, the connection between the technical challenges and the societal factors contributing to them would remain obscured. By gathering insights from both projects, I realized that technological improvements are only as meaningful as the frameworks that govern them. This understanding not only increased my appreciation for STS but also highlighted the critical role of a sociotechnical approach in driving innovation.

School of Engineering and Applied Science

Bachelor of Science in Computer Science

Technical Advisor: Rosanne Vrugtman

STS Advisor: Kathryn Neeley