I am on the job market for tenure-track faculty & postdoc positions for fall'25! Open to relocate within the US. Check out my CV and contact me!
Projects
AI Teacher
AI Teacher is a human-in-the-loop explainable AI (XAI) framework for explaining robot behavior. It includes a policy summarization algorithm based on Bayesian user modeling, an interactive user interface allowing users to ask user-specific questions, and a detailed user study showing the effect of interactive teaching techniques for XAI.
Related Papers:
-
AI Teacher 1.0: Evaluating the Role of Interactivity on Improving Transparency in Autonomous Agents (AAMAS '22) [pdf]
-
AI Teacher 2.0: Interactively Explaining Robot Policies to Humans in Integrated Virtual and Physical Training Environments (HRI'24 LBR) [pdf] [poster]
-
AI Teacher 3.0: Personalized Policy Summarization for Explaining Sequential Behavior of Autonomous Agents (AIES'24) [pdf] [poster]
Related Video Demos (some of the videos do not have sound):
Robot-assisted Nursing / Robot Tutor
Robot Tutor is an Intelligent Robot Tutoring system developed to train and assess novice nurses on medical procedures. The system includes planning, perception, and speech capabilities.
Related Papers:
-
Robotic Tutors for Nurse Training: Opportunities for HRI Researchers (RO-MAN'23) [pdf]
In the demo above, Stretch performs simple routine tasks in a hospital room such as fetching and placing objects, and pushing the bedhead button to adjust the hospital bedhead. The geometric shapes can be replaced with hospital-related objects such as blood samples, masks, gloves, and tool kits.
I implemented Bayesian Knowledge Tracing (BKT) to evaluate a nurse's skills for maintaining sterilized (in this example, the hands need to be always above the waistline). The original BKT is not designed for continuous assessment so I modified it based on this paper, essentially by running BKT every 0.1 seconds. Two versions are implemented:
1. BKT w/o learning assumes the user is not gaining new knowledge during this assessment, thus using the initial prior for every step of the assessment;
2. BKT w/ learning assumes the user continues to learn during the assessment, thus using an updated prior.
* The vision system (left side of the screen) is built by my summer intern student Filip Bajraktari.
I-CEE
One of the biggest challenges faced by AI researchers is the lack of interpretability of deep neural networks - the machine learning model worked, but why? or the machine learning model made an error, why? To address this problem, my co-authors and I propose a framework, titled I-CEE, that utilizes human cognition modeling to help human users make sense of the machine learning model. This interdisciplinary work marks novel contributions toward generating personalized explanations for users.
Related Papers: