Bounded Expectations for Discrepancy Detection in Goal-Driven Autonomy

Abstract

Goal-Driven Autonomy (GDA) is a model for online planning extended with dynamic goal selection. GDA has been investigated in the context of numerous abstract planning domains, and there has been recent interest in applying GDA to control unmanned vehicles. In robotic domains, certain continuous state features from sensor data must be modeled for reasoning. However, modeling these features precisely during planning and execution monitoring may be problematic, due to the inefficiency of computing exact values or sensitivity to noise. We present PHOBOS, a Hierarchical Task Network planner with bounded expectations, which we apply with a GDA agent in an underwater vehicle domain. Bounded expectations allow an agent to plan and detect discrepancies more efficiently and with fewer false discrepancies (i.e., detected but semantically meaningless differences from expectations during execution). We describe an initial simulation study that supports this claim.