Learning high-level navigation strategies from sensor information and planner experience

Abstract

Moving a robot with shape and size in a cluttered dynamic workspace requires the capability of dealing with obstacles and local minima. The research analyzes situations where no global knowledge about the environment exists, and where the robot can only perceive the space through its local sensors. The system explores a dynamic space using a planner based on local artificial potential fields, and incrementally learns a fast way to escape from dead lock situations using a combination of sensor perceptions and field information. As main result the system learns and uses an high level description of the workspace consisting of local minimum nodes, backtracking nodes and subgoal nodes.