NASA NTRS2007
A new approach is suggested to define and evaluate key metrics as to autonomous aerial vehicle performance. This approach entails the conceptual definition of a "Turing Test" for UAVs. Such a "UAV Turing test" would be conducted by means of mission simulations and/or tailored flight demonstrations of vehicles under the guidance of their autonomous system software. These autonomous vehicle mission simulations and flight demonstrations would also have to be benchmarked against missions "flown" with pilots/human-operators in the loop. In turn, scoring criteria for such testing could be based upon both quantitative mission success metrics (unique to each mission) and by turning to analog "handling quality" metrics similar to the well-known Cooper-Harper pilot ratings used for manned aircraft. Autonomous aerial vehicles would be considered to have successfully passed this "UAV Turing Test" if the aggregate mission success metrics and handling qualities for the autonomous aerial vehicle matched or exceeded the equivalent metrics for missions conducted with pilots/human-operators in the loop. Alternatively, an independent, knowledgeable observer could provide the "UAV Turing Test" ratings of whether a vehicle is autonomous or "piloted." This observer ideally would, in the more sophisticated mission simulations, also have the enhanced capability of being able to override the scripted mission scenario and instigate failure modes and change of flight profile/plans. If a majority of mission tasks are rated as "piloted" by the observer, when in reality the vehicle/simulation is fully- or semi- autonomously controlled, then the vehicle/simulation "passes" the "UAV Turing Test." In this regards, this second "UAV Turing Test" approach is more consistent with Turing s original "imitation game" proposal. The overall feasibility, and important considerations and limitations, of such an approach for judging/evaluating autonomous aerial vehicle "intelligence" will be discussed from a theoretical perspective.