Learning area coverage using the co-evolution of model parameters

Abstract

The type of search where a robot’s track takes it on a path where its sensors can detect all mines located in the area is referred to as area coverage. Planning this track is an issue in robotics and is complicated when the robot is legged due to the reduced precision of its movements. Cyclic genetic algorithms have been used as a method for learning the cycle of turns and straights required for a hexapod robot to solve the area coverage problem. Although successful in a static environment, the learning system needed an anytime component to make it adaptable enough to be used in practice. This paper discusses the creation of a viable learning system by adding the anytime learning technique of co-evolving model parameters. Tests in simulation demonstrate this system's usefulness in generating search patterns despite changes in the robot's performance. cell to cell path through the area. Choset and Pignon (1997) divided the area into obstacle free sub-areas and found an exhaustive path through the adjacency graph representing these cells. Within each cell the back-andforth boustrophedic motions (Figure 1) were used to assure coverage. Ollis and Stentz (1997) used vision to control the lines in their boustrophedic motions to do automated harvesting. Hofner and Schmidt (1995) used templates appropriate for the type of robot to determine the best path within varying sized areas. In addition to dead reckoning, landmarks sensed by ultrasonic sensors were used to maintain the desired track. Hert at el. (1996) used an on-line planar algorithm and sensors for an autonomous underwater vehicle to explore areas of the sea floor with arbitrary shape. 1