GuidAR: Meta-Learned MoE Diffusion Costmaps from AR Trajectory Edits; War With Project Managers 110 views

My research portfolio asks what happens to human intent when it is translated through an algorithmic system. In my technical work, GuidAR, I approach this as a design problem. My system uses Mixed Reality trajectory editing, LLM/VLM semantic grounding, and adaptive diffusion-based costmaps such that operators directly shape their robot's path, and uses those edits to learn a user's preferred offset in future novel scenarios. By making intent visible and correctable, GuidAR treats the MR interface as the place where a human can demystify the black-box algorithm behind a robot's path. My STS paper, War With Project Managers, shows the opposite: what happens when black-box algorithms are used without any transparency. Drawing on Actor-Network Theory, I show how AI-enabled military targeting systems can preserve the appearance of "human-in-the-loop" oversight while black-boxing recommendations, compressing decision time, and most importantly diffusing accountability across many actors. Together, these projects argue that human-AI interfaces can either preserve human agency by keeping intent visible and revisable, or dissolve it through speed, opacity, and institutional deniability.

BS (Bachelor of Science)

mixed-reality; diffusion; robotics; motion planning; meta learning; machine learning; reinforcement learning; inverse reinforcement learning

School of Engineering and Applied Science Bachelor of Science in Computer Science Technical Advisor: Nicola Bezzo STS Advisor: Travis Elliott

English

All rights reserved by the author (no additional license for public reuse)

2026-05-08