Conformal prediction provides distribution-free uncertainty guarantees for learned perception, but existing pose-estimation methods often construct joint guarantees from per-keypoint uncertainty estimates using conservative aggregation methods. This project develops a geometry-aware conformal framework for multi-keypoint pose uncertainty by modeling spatial dependence between keypoint localization errors directly from calibration residuals and object geometry. We will study how keypoint failure correlations depend on 3D structure, derive calibratable joint coverage bounds that account for this dependence, and design conformal score functions that better capture multi-keypoint error patterns, including outlier images where many predictions fail simultaneously. The resulting method will be empirically evaluated on benchmark pose-estimation data, with the goal of producing tighter, statistically valid uncertainty bounds for safety-critical robotic perception.

Through SuperUROP, I hope to deepen my experience working on research that connects mathematical ideas with real-world problems in robotics. I am especially excited by this project because it asks how we can make learned perception systems not only more accurate, but also more trustworthy when deployed in uncertain environments. I see SuperUROP as an opportunity to grow as both a researcher and communicator while contributing to a project with meaningful implications for safe and reliable autonomy.