Path integration is an important navigation strategy in many animal species. We use a genetic algorithm to evolve a novel neural model of path integration, based on input from cells that encode the heading of the agent in a manner comparable to the polarization-sensitive interneurons found in insects. The home vector is encoded as a population code across a circular array of cells that integrate this input. This code can be used to control return to the home position. We demonstrate the capabilities of the network under noisy conditions in simulation and on a robot.