Control-oriented modeling of flight demonstrations for quadrotors using higher-order statistics and dynamic movement primitives

In this paper, we present a novel method for parsing demonstrations, and further characterizing the segments as subactions which are easy to implement by low-level motion controllers. Demonstration data's attributes of Gaussianity and linearity are linked to teacher's control manners and intentions, and Hinich's Gaussian and linear tests with higher-order statistics are adopted for segmentation. Wigner spectrum tools are applied to locate the rhythmic phases. Having parsed the demonstration into segments based on the test results, segmental features are parametrically represented in different ways, among which dynamic movement primitives (DMPs) are used to unifiedly model the nonlinear and rhythmic segments. For the multi-dimensional demonstrations, rules of selecting suitable variables for characterization are presented for both linear and nonlinear cases. The adverse effects brought about by inter-axis couplings are discussed and recognized using heuristics. For the case of multiple demonstrations, three-leveled feature consistency problem is also addressed. The proposed techniques are integrated into a whole for learning from flight demonstrations, and evaluated in simulations by modeling a quadrotor's axial roll maneuver. Results show the effectiveness of our method.