Robustness analysis of genetic programming controllers for unmanned aerial vehicles

While evolving evolutionary robotics controllers for real vehicles is an active area of research, most research robots do not require any assurance prior to operation that an evolved controller will not damage the vehicle. For controllers evolved in simulation where testing a poorly performing controller might damage the vehicle, thorough testing in simulation - subject to multiple sources of sensor and state noise - is required. Evolved controllers must be robust to noise in the environment in order to operate the vehicle safely. We have evolved navigation controllers for unmanned aerial vehicles in simulation using multi-objective genetic programming, and in order to choose the best evolved controller and to assure that this controller will perform well under a variety of environmental conditions, we have performed a series of robustness tests. The results show that our best evolved controller outperforms two hand-designed controllers and is robust to many sources of noise.