Learning User Models During Shared Autonomy

In shared autonomy, user input and robot autonomy are combined to control a robot to achieve a goal. One often used strategy considers the user and autonomous system as independent decision makers, combining the output of these two entities. However, recent work has shown that this can lead to poor performance, suggesting the need to incorporate user models into the autonomous decision maker. A key challenge for doing so is learning a good user model. Existing methods for learning user models assume autonomous assistance is not present, while prior work suggests that users behave differently in the presence of assistance. In this work, we propose a method for learning user behavior during shared autonomy. We present a cost minimization framework that utilizes this learned model to select assistance actions that minimize the cost incurred by the user. Finally, as altering the assistance strategy can change how the user behaves, we present a method for evolving the model and assistance strategy together.