On Automation of Flight Training of AI Pilots: A Formal Language to Particularize Flight Instruction Scenarios

Considering development assurance, predictability, explainability and robustness, the lack of trustworthiness has been identified as the key challenge when it comes to the certification of adaptive flight automation systems. Such systems change their characteristics in response to changes in the mission, environment, threats, or failures, which disallows certification according to the currently applicable set of regulations that stipulate a strictly deterministic behaviour. Automation systems are supposed to maintain the status that has been certified. They are not allowed to evolve, to learn from experience, because the risks associated with a changing behaviour are unforeseeable.Rule-based, certifiable software is developed according to well-established processes and software assurance. A thorough requirements engineering results in a set of software requirements, which will be, after deployment, traced back and verified according to the specification. Learning-based systems, in contrast, learn from training data. Herewith, the challenge shifts to the production of such training data and the subsequent training. On the one hand, training data has to meet the required level of integrity. Incomplete, wrong, ambiguous, noised, or biased training data may lead to undesired behaviour. On the other hand, training data should be available in sufficient quantity for training, validation, and testing of the system in question.An anthropomorphic approach could significantly contribute to the production of training data and the training itself, given the availability — and the validity — of all the training exercises and check standards for manned aviation. Once in a machine-interpretable form, these well-established and proven exercises could be modified automatically with adaptions of the environmental conditions (weather, other traffic, …), with unexpected occurrences of system failures, or could be re-created from a base set of building blocks. The training of a future adaptive flight automation system could profit from a great availability, integrity, and diversity of data. During the training, an electronic flight instructor could demonstrate the accomplishment of such training scenarios, allowing the system to learn from example. Then, an electronic flight examiner could check whether a flight automation system satisfactorily completes the given exercise, based on check standards from manned aviation. Even after deployment, the monitoring could continue, flagging any decrease in the desired performance and disclosing adverse learning effects.Training syllabi and simulator session rundowns delineate what the assigned flight instructor is supposed to do during the training. Many syllabi are available for basic flight training, type rating training, line-oriented flight training, low-visibility training, upset prevention and recovery training, training for airports with special approach procedures and much more. Check standards are formulated, framing the allowable tolerances during a flight check.This paper copes with the challenge of compiling these syllabi and standards into an unambiguous form that can be used to automate flight training. It first introduces a classification scheme for flight training exercises, dividing them into the classes of deterministic and non-deterministic exercises, supporting the different training phases. Secondly, the general flying tasks of a pilot are categorized into prime flying tasks, indivisible tasks that are combinable in order to meet the standards in all phases of an exercise. Finally, based on these two ingredients, a formal language is proposed, interpretable by a computer system, by an electronic flight instructor / flight examiner, to particularize flight training scenarios — with a concrete example of a basic training exercise, the standard traffic pattern. This will eventually support the establishment of such an anthropomorphic approach, allowing us to train and check future adaptive flight automation systems in a traceable, proven, and extensible framework nowadays already in use for human pilots. Until then, the proposed automated flight training can be validated and refined with human pilot aspirants and human flight instructors.