Achieving Fully-Automated Aircraft Flight with Limited Resources

Fully-automating aircraft flight presents a number of challenges for researchers. Perhaps most important is the criticality of real-time responses -an aircraft cannot remain safe indefinitely once it has left the ground. Additionally, aircraft flight dynamics are complex and not fullyunderstood, so computations are often time-consuming and approximate at best. Current aircraft use low-level controllers to handle ordinary situations, but rely on human pilot commands to respond when emergencies or unexpected situations (e.g., course changes) arise. We are working to expand the role of autopilots to also handle emergency situations. Unfortunately, there are many possible emergencies, ranging from a failed sensor to collision-course traffic to unexpectedly flying into a tornado. In fact, the comprehensive set of possible emergencies is so large that it is infeasible to build a database of pre-planned responses. Instead, we approach the problem by building control plans to handle “expected”, or highly-probable, situations. Then, we detect and replan to handle unexpected situations or emergencies as they arise. Flexible computation is very important during these replanning phases. Ideally, the planner could carefully expand all reachable states, selecting actions to keep the plane safe while reaching the goal. However, a quick response is usually necessary because the aircraft cannot just “stop” and wait for a new plan. For example, if all engines fail, the new plan must be completed in time for the aircraft to fly to a desirable landing location (e.g., nearby airport or open field) before it loses too much altitude. Thus, we should restrict planning time to allow production of an approximate plan in time to execute successfully. If planning time permitted, this plan may be able to reach the desired goal. Alternatively, if allowable planning time was brief, this approximate plan may simply maintain a safe set of states, buying the time required to develop a complete plan that can reach the goal. We have investigated replanning for unexpected situations during aircraft flight in the context of CIRCA (Cooperative Intelligent Real-time Control Architecture) [Musliner], which combines a planner, scheduler, and separate real-time plan execution module such that plans are built, scheduled, then executed with real-time guarantees of system safety. Section 2 briefly describes CIRCA, including recent enhancements that admit probabilistic knowledge specification as well as the detection and handling of unexpected situations. Section 3 describes flight simulation tests, followed by a discussion of future additions we feel are required to achieve a sufficient level of robustness when replanning must occur rapidly (Section 4).