Experience- and Model-based Transformational Learning of Symbolic Behavior Specifications

This paper describes XFRMLEARN, a system that learns symbolic behavior specifications to control and improve the continuous sensor-driven navigation behavior of an autonomous mobile robot. The robot is to navigate between a set of predefined locations in an office environment and employs a navigation system consisting of a path planner and a reactive collision avoidance system. XFRMLEARN rationally reconstructs the continuous sensor-driven navigation behavior in terms of task hierarchies by identifying significant structures and commonalities in behaviors. It also constructs a statistical behavior model for typical navigation tasks. The behavior model together with a model of how the collision avoidance module should “perceive” the environment is used to detect behavior “flaws,” diagnose them, and revise the plans to improve their performance. The learning method is implemented on an autonomous mobile robot.