An Architecture for the Autonomous Generation of Preference-Based Trajectories
Atkins, Ella M.
MetadataShow full item record
Numerous techniques exist to optimize aircraft and spacecraft trajectories over cost functions that include terms such as fuel, time, and separation from obstacles. Relative weighting factors can dramatically alter solution characteristics, and engineers often must manually adjust either cost weights or the trajectory itself to obtain desirable solutions. Further, when humans and robots work together, or when humans task robots, they may express their performance expectations in a "fuzzy" natural language fashion, or else as an uncertain range of more or less acceptable values. This work describes a software architecture which accepts both fuzzy linguistic and hard numeric constraints on trajectory performance and, using a trajectory generator provided by the user, automatically constructs trajectories to meet these specifications as closely as possible. The system respects hard constraints imposed by system dynamics or by the user, and will not let the user's preferences interfere with the system and user needs. The architecture's evaluation agent translates these requirements into cost functional weights expected to produce the desired motion characteristics. The quality of the resulting full-state trajectory is then evaluated based on a set of computed trajectory features compared to the specified constraints. If constraints are not met, the cost functional weights are adjusted according to precomputed heuristic equations. Heuristics are not generated in an ad hoc fashion, but are instead the result of a systematic testing of the simulated system under a range of simple conditions. The system is tested in a 2DOF linear and a 6DOF nonlinear domain with a variety of constraints and in the presence of obstacles. Results show that the system consistently meets all hard numeric constraints placed on the trajectory. Desired characteristics are often attainable or else, in those cases where they are discounted in favor of the hard constraints, failed by small margins. Results are discussed as a function of obstacles and of constraints.