Abstract
Autonomous systems are rapidly expanding into complex, real-world environments, driven by advances in artificial intelligence, perception, and computing. At the same time, the governance structures intended to regulate these technologies have struggled to keep pace with their development and deployment. This gap is especially significant because modern autonomy increasingly depends on flexible, general-purpose perception systems that are difficult to define, categorize, or constrain. As these systems become more capable and widely deployed, the boundaries of acceptable use are often through practice rather than formal deliberation in the international community. This thesis portfolio addresses this broader problem from two perspectives: first, through the development of an advanced perception system for off-road autonomous driving, and second, through an analysis of how governance failures and real-world deployment contribute to the normalization of autonomous technologies, particularly in military contexts. These projects illustrate how technical innovation and sociotechnical dynamics interact to shape both what autonomous systems can do and how they come to be accepted in practice.
The technical project focuses on improving perception for off-road autonomous driving by integrating pre-trained vision foundation models into a real-time system for semantic understanding and traversability estimation. Off-road environments present a significant challenge for autonomy because they are highly unstructured, lack standardized features such as lane markings, and require systems to generalize across diverse terrain types. To address this, the project leverages NVIDIA’s RADIO model, a distilled vision foundation model trained on multiple teacher models including SigLIP2, DINOv3, and SAM. By applying the SigLIP2 summary head to individual patch embeddings, the system produces language-aligned features at a dense level, enabling open-vocabulary semantic segmentation. This allows users to identify and segment arbitrary objects in real time simply by interacting with the system, significantly increasing flexibility compared to traditional fixed-class segmentation approaches. The model is deployed using TensorRT for accelerated inference, achieving performance well above the required 20 Hz operating rate. In addition to on-board perception, the system incorporates overhead satellite imagery combined with digital elevation models to generate terrain-aware traversability costs, accounting not only for surface type but also for slope and elevation. The result is a scalable and adaptable perception pipeline that enhances autonomous navigation in complex environments by enabling richer semantic understanding and more informed path planning.
The STS research paper examines why international efforts to regulate lethal autonomous weapons systems (LAWS) have stalled despite ongoing debate and increasing real-world deployment. The central research question asks how governance failure, definitional ambiguity, and battlefield deployment interact to normalize the use of increasingly autonomous systems. Drawing on concepts such as mutual shaping and Actor-Network Theory, the analysis argues that autonomy is a flexible and contested category that states actively manipulate to avoid binding constraints. Within international forums such as the Convention on Certain Conventional Weapons, negotiations have stalled due to disagreement over definitions and the strategic use of ambiguity by major powers. This governance failure creates a permissive environment in which deployment proceeds unchecked. Evidence from the war in Ukraine demonstrates how battlefield conditions, such as electronic jamming, incentivize the adoption of systems with greater autonomous capabilities, while the lack of strong international response allows these uses to become normalized. The paper also highlights how governance is increasingly shifting into the commercial sector, where companies and governments negotiate acceptable uses of AI through contracts and safeguards, as seen in recent disputes between AI firms and the U.S. Department of Defense. The paper concludes that norms surrounding autonomous systems are emerging through repeated use and adaptation rather than through formal agreement, a process that may constrain future regulatory efforts.
Taken together, these projects demonstrate that advances in perception systems are part of a broader transformation in how autonomous technologies are developed, deployed, and governed. The technical project shows how modern vision models enable more general, flexible, and scalable forms of autonomy in complex environments. The STS research shows that as such capabilities expand, the difficulty of defining and regulating autonomy increases, allowing real-world deployment to shape norms in advance of formal governance. While the technical work achieves its goal of improving perception and navigational decision-making, it also reflects a larger trend toward increasingly adaptable and widely applicable autonomous systems. Future work should continue to explore not only how to improve these technologies, but also how to establish clearer frameworks for their governance before patterns of use become entrenched.
