Autonomy Artificial Intelligence Robotics (AAIR)

Topography Dependent Path Planning using Deep Q-Learning

by Eric Martinson; Ben Purman; Andy Dallas


Future autonomous combat vehicles will need to travel off-road through poorly mapped environments. Three-dimensional topography may be known only to a limited extent (e.g. coarse height), but this will likely be noisy and of limited resolution. For ground vehicles, 3D topography will impact how far ahead the vehicle can “see”. Higher vantage points and clear views provide much more useful path planning data than lower vantage points and occluded views from trees and structures. The challenge is incorporating this knowledge into a path planning solution. When should the robot climb higher to get a better view or else continue moving along the shortest path predicted by current information? We investigated the use of Deep Q-Networks (DQN) to reason over this decision space, comparing performance to conventional methods. In the presence of significant sensor noise, the DQN was more successful in finding a path to the target than A* for all but one type of terrain.