Safety validation of sense and avoid algorithms using simulation and evolutionary search

We present a safety validation approach for Sense and Avoid (SAA) algorithms aboard Unmanned Aerial Vehicles (UAVs). We build multi-agent simulations to provide a test arena for UAVs with various SAA algorithms, in order to explore potential conflict situations. The simulation is configured by a series of parameters, which define a huge input space. Evolutionary search is used to explore the input space and to guide the simulation towards challenging situations, thus accelerating the process of finding hazards and supporting the validation process. We applied our approach to the recently published Selective Velocity Obstacles (SVO) algorithm. In our first experiment, we used both ran-dom and evolutionary search to find mid-air collisions where UAVs have perfect sensing ability. We found evolutionary search can find some hazards (here, interesting problems with SVO) that random search takes a long time to find. Our second experiment added sensor noise to the model. Random search found similar problems as it did in experiment one, but the evolutionary search found some interesting new problems. The two experiments show that the proposed approach has potential for safety validation of SAA algorithms