Generalizing Competency Self-Assessment for Autonomous Vehicles Using Deep Reinforcement Learning

Abstract: Due to the increased role of autonomous robots in accomplishing a variety of challenging tasks alongside humans, it is essential for the human operator to establish appropriate trust towards these systems. To this end, we present a step towards generating competency-aware autonomous agents that are able to communicate their self-confidence for the given task. We develop and analyze an autonomous model-based reinforcement learning UAV ISR agent that uses a neural network based learned model of the world alongside an uncertain planner to generate a series of simulated trajectories. These trajectories, which capture uncertainties from both the planner and the model, are assessed using both reward-based Outcome Assessment (OA) metric and the more intuitive outcome-based Generalized Outcome Assessment (GOA) metric. Simulation results for the UAV ISR agent show the usefulness of leveraging learned probabilistic world models with OA and GOA self-confidence reports to assess and convey autonomous agent competencies for assigned tasks in complex uncertain environments.