Learning action models to support efficient navigation planning for unmanned ground vehicles
Natasha Balac, Daniel M. Gaines, Siripun Thongchai, Doug H. Fisher, Surachai Suksakulchai · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2001
An effective navigation planner must have knowledge not only of the effects its actions will have, but also the effect that the environment will have on its actions (e.g. the UGV may travel more slowly over rough terrain). This is needed because the shortest path to the goal is not always the most efficient when you consider the rate of travel over the terrain. To address this issue, we have developed an approach called ERA which uses regression tree induction to learn action models that predict the effect terrain conditions will have on a UGV's navigation actions. The action models support a high level mission planner that finds efficient navigation plans consisting of way-points through which the UGV should travel. We will present results from our experiments in a simulated environment and on an RWI ATRV-Jr robot. The studies evaluate the performance of ERA in different mission scenarios with different amounts of sensor and actuator noise. Advantages of our approach include the ability to automatically learn action models, generate efficient high level navigation plans taking into account terrain conditions and transfer learned knowledge to other missions.