[BK세미나] 11/20(금) 원창희 교수(Temple University) "Toward Autonomous Robotic Vibrotactile Imaging for Tumor Characterization using Reinforcement Learning-based Tracking Control"

기계공학부 구성원들의 많은 관심과 참여 부탁드립니다.

▣ 주 제: Toward Autonomous Robotic Vibrotactile Imaging for Tumor Characterization using Reinforcement Learning-based Tracking Control

▣ 연 사: 원창희 교수

▣ 소 속: Temple University

▣ 일 시: 2026. 11. 20.(금) 11:00

▣ 장 소: 제4공학관 D602호

▣ 초 청: 김종백 교수

▣ 초 록

Early and accurate characterization of malignant tumors is critical to reducing cancer mortality, yet many existing screening approaches depend heavily on operator skill and are difficult to scale. In this talk, I will present our work toward an autonomous robotic vibrotactile imaging platform that characterizes tumors non-invasively and acquires data with minimal human intervention.

First, I will introduce a differential vibrotactile imaging system that augments tactile imaging with ultrasound-induced vibration. The system captures multiple tactile and vibrotactile images and synthesizes them into representative profile diagrams. By quantifying pixel-level variations between the tactile and vibrotactile profiles, we obtain differential profile diagrams that reveal subtle mechanical properties of the tissue, which convolutional neural networks then classify as benign or malignant. I will share results from tissue-mimicking phantoms, as well as in vivo human data acquired with the tactile modality.

Second, I will address a key barrier to clinical translation: data acquisition in these systems has largely relied on manual operation, which limits repeatability, efficiency, and scalability. To enable autonomous sensing, we developed a Minimal Cost Variance (MCV) tracking control framework for stochastic robotic systems. I will describe a continuous-time, infinite-horizon, model-based MCV tracking controller derived through policy iteration, together with its convergence and robustness analysis under bounded uncertainty. Because an accurate model of the sensing platform is often unavailable in practice, we also developed a model-free reinforcement learning algorithm that learns the optimal tracking policy directly from measured data. A ridge-regression-based policy evaluation improves numerical stability for ill-conditioned robotic dynamics, and our analysis shows that the learned gain remains near-optimal under small, bounded data disturbances.

Together, these advances lay the groundwork for autonomous robotic vibrotactile imaging that delivers repeatable, operator-independent tumor characterization. I will conclude with remaining challenges and future directions toward clinical deployment.