Planning to learn: Integrating model learning into a trajectory planner for mobile robots

Abstract

For a mobile robot that performs online model learning, the learning rate is a function of the robot's trajectory. The tracking errors that arise when the robot executes a motion plan depend on how well the robot has learned its own model. Therefore a planner that seeks to minimize collisions with obstacles will choose plans that decrease modeling errors if it can predict the learning rate for each plan. In this paper we present an integrated planning and learning algorithm for a simple mobile robot that finds safe, efficient plans through a grid world to a goal point using a standard optimal planner, A*. Simulation results show that with this algorithm the robot practices maneuvers in the open regions of the configuration space, if necessary, before entering the constrained regions of the space. The robot performs mission-specific learning, acquiring only the information it needs to complete the task safely.