HIFuzz: Human Interaction Fuzzing for small Unmanned Aerial Vehicles

Small Unmanned Aerial Systems (sUAS) must meet rigorous safety standards when deployed in high-stress emergency response scenarios. However, tests that execute perfectly in simulation can fail dramatically in real-world environments. Fuzz testing can be used to increase system robustness by providing malformed input data aimed at triggering failure cases. In this paper, we apply fuzzing to support human interaction testing. Initial tests are run in simulation to provide broad coverage of the input space in a safe environment; however, they lack the fidelity of real-world tests. Field tests provide higher fidelity but can result in costly or dangerous crashes. We, therefore, propose and demonstrate HiFuzz, which executes large numbers of fuzz tests in simulation and then down-selects tests for deployment in human-in-the-loop simulations and safety-aware physical field tests. We apply \hf to a multi-sUAS system and show that each test level serves a unique purpose in identifying known and unknown failures associated with human interactions.