Evolved Navigation Control for Unmanned Aerial Vehicles

摘要

Using multi-objective GP, we were able to evolve navigation controllers for UAVs capable of flying to a target radar, circling the radar site, and maintaining an efficient flight path, all while using inaccurate sensors in a noisy environment. Controllers were evolved for five radar types using both direct evolution and incremental evolution: continuously emitting, stationary radars; continuously emitting, mobile radars; intermittently emitting, stationary radars with regular periods; intermittently emitting, stationary radars with irregular periods; and intermittently emitting, mobile radars with regular periods. The use of incremental evolution dramatically increased the chances of producing successful controllers compared to direct evolution. Incremental evolution also produced controllers able to handle all five radar types. Controllers were evolved to use inaccurate sensors in a noisy environment. We tested the transferability of the evolved controllers by using them to control a wheeled mobile robot. Evolved UAV controllers were successfully transferred to a wheeled mobile robot equipped with a passive sonar system which provided the angle and amplitude of sound signals from a stationary speaker. Using evolved navigation controllers, the mobile robot moved to the speaker and circled around it. The results from this experiment demonstrate that our evolved controllers are capable of transference to real physical vehicles. We developed a series of robustness tests for evolved navigation controllers for UAV controllers developed in simulation. Before testing evolved controllers on a real UAV, we needed some assurance that the off-design performance of these controllers would be sufficient to accomplish the desired task and that controllers would be able to avoid behaviors that could potentially damage the aircraft. Also, since the controllers were evolved using multi-objective optimization, we needed to select a single best controller. When evolving controllers for systems where tests may be dangerous to the vehicle, robustness tests might be useful in selecting a controller and seeing how well it performs. The robustness tests described here apply several sources of sensor and state noise. If the real-world noise falls within the range where tests in simulation performed well, we can expect that transference will be successful.