Current Projects

​

Research Question 1: Explain tasks that traditionally cannot be modeled as Markovian models or neural networks. 

 

• Research Gap: In existing XAI literature, tasks are most commonly represented using Markovian models (e.g., reinforcement learning) or neural networks (e.g., image classification). However, many human tasks, like those in healthcare, may not be best represented using these models. How to extend existing XAI methods to explain diverse task models is an open and interesting research question. 

​​

• Proposal: First, we need to understand how humans represent tasks. In healthcare, step-by-step checklists are frequently used. However, converting checklists into computational models is not trivial because steps in checklists include complex environment observations (e.g., move the trashcan next to the patient bed), states of medical tools (e.g., keep the sterile field dry), and human-object interaction (e.g., scrub the insertion site for 30 seconds using CHG). Some steps are composed of multiple actions (e.g., use sterile techniques to open the dressing kit) – sterile techniques are not specified and may vary based on the tasks and environment. To explain tasks that involve complicated environmental observations and human-object interactions, I will first extend existing task models in robotics, such as Planning Domain Definition Language (PDDL) and Hierarchical Task Networks (HTN), as useful starting points. In the long run, I will investigate novel techniques that capture the diversity in healthcare tasks and hospital-specific practices. I will explore the use of Large Language Models (LLM) to enable healthcare experts to translate their domain knowledge into robot-interpretable computational models, and then generate high-quality explanations. Leveraging the benefits of personalized learning, these explanations will be tailored to individual learners. 

​

​

​

​

Research Question 2: Automate skill evaluation in challenging real-world domains.

 

• Research Gap: Existing Knowledge Tracing methods, traditionally used for discrete observations such as multiple choices, struggle with continuous observations. Unlike discrete domains where the observation-to-skill mapping is relatively straightforward, evidence of skill mastery in continuous, real-world domains is often implicit and largely depends on the context.  

​​

• Proposal: To tackle this fundamental question, it is important to look at how human experts evaluate skills in such domains. For rule-based skills, such as sterile techniques, we can track the number of times a student breaks each rule using multimodal data – vision, language, and interaction. For knowledge-based skills that depend on context, such as treatment for dropping blood sugar levels, I propose to use physical robot intervention to create realistic scenarios for students to practice. Simulating realistic scenarios is a common practice that experienced nurses use to train new nurses but it varies from nurse to nurse and from hospital to hospital. A standardized simulation can be realized through robot teachers. I also propose to extend existing methods to create new Knowledge Tracing methods for broad real-world domains.