Crossing Boundaries: Multi-Level Introspection in a Complex Robotic Architecture for Automatic Performance Improvements

Abstract

Introspection mechanisms are employed in agent architectures to improve agent performance. However, there is currently no approach to introspection that makes automatic adjustments at multiple levels in the implemented agent system. We introduce our novel multi-level introspection framework that can be used to automatically adjust architectural configurations based on the introspection results at the agent, infrastructure and component level. We demonstrate the utility of such adjustments in a concrete implementation on a robot where the high-level goal of the robot is used to automatically configure the vision system in a way that minimizes resource consumption while improving overall task performance.