Self-improving robot action management system with probabilistic graphical model based on task related memories

Abstract

Robots on home environment have to deal with their environment that changes every time they perform tasks. In order to reduce improving descriptions for task and action planning manually, we propose a new framework to use probabilistic graphical model, which enables robot agent know which action in the task affects the result of the task the most, and by improving parameters of the action robot can maintain high success rate of tasks. Our framework let robot infer failure action of tasks using data from early task performance which are automatically recorded and retrieved with high-level data retrieval query interface. We evaluated our approach using mobile manipulation robot PR2 on daily assistive environment and task.