The pair are part of a multi-university research team commissioned by the Air Force Office of Scientific Research to study trust in autonomous systems. It is an important and complex problem.
“Trust is a dynamic human state with multiple dimensions - it’s different for each individual and the specific system you’re using. Trusting a self-driving car if you want to go to sleep in the backseat is different than trusting Alexa to tell you the weather,” said Anderson, an assistant professor in the Ann and H.J. Smead Department of Aerospace Engineering Sciences at the University of Colorado Boulder.
The work has broad applications across the technological spectrum, but the Air Force is particularly interested due to increasing integration of autonomy in military systems and the uncertainties faced by soldiers using them said Anderson.
“There are many applications where autonomous systems may be used and – particularly with space-based applications – the human isn’t onsite with a satellite to have additional context, and you can’t always get all the data in real-time. We need to understand how users trust and view that type interaction with autonomy across a variety of situations,” Anderson said.
The initiative aims to build metrics and models for real time predictions of trust, with the goal of helping developers create better AI systems in the future said Clark, an associate professor in Smead Aerospace.
“Space and military autonomy represent critically challenging environments and being able to estimate and predict human-operator trust will enable systems to intelligently alter their behaviors to complement their human teammates,” Clark said.
During the research, test subjects will be fitted with wearable sensors while they conduct tasks with AI systems. The sensors will collect physiological data on the body’s responses – things like heart rate and respiration – as well as how users physically interact with the systems. That includes where they are looking on a computer screen, the buttons they click, and how long they take to do an activity requested by the AI powered system.
“It’s exciting to work in this emerging field where there are important questions that need to be answered to move out of the laboratory and into operations,” Anderson said.
The three-year, $900,000 grant is being led overall by the University of California, Davis. CU Boulder’s work represents nearly $500,000 of the total award.