To Ask or Not To Ask: Human-in-the-loop Contextual Bandits with Applications in Robot-Assisted Feeding

Area of Science:

• Robotics
• Human-Computer Interaction
• Artificial Intelligence

Background:

• Robot-assisted bite acquisition faces challenges with diverse food items.
• Fully autonomous systems struggle to generalize across varied food properties.
• User feedback is crucial but can increase workload.

Purpose of the Study:

• To develop a human-in-the-loop strategy for robot-assisted bite acquisition.
• To minimize user querying workload while maintaining task performance.
• To adapt querying behavior based on user workload and limitations.

Main Methods:

• Formulated human-in-the-loop bite acquisition within a contextual bandit framework.
• Introduced LINUCB-QG, a method using a predictive model of querying workload.
• Trained the model on data from online studies with diverse participant groups.

Main Results:

• The proposed method effectively balances task performance and querying workload.
• LINUCB-QG outperforms autonomous and always-querying baselines.
• Querying behavior adjusts to accommodate higher workload in users with mobility limitations.

Conclusions:

• Leveraging selective user feedback improves robot-assisted eating systems.
• The LINUCB-QG method offers an adaptive and efficient approach.
• This research has implications for assistive robotics and personalized care.