Learning and Control of UAV maneuvers Based on Demonstrations

Many maneuvers of Unmanned Aerial Vehicles (UAV) can be considered within a framework of trajectory following. Though this issue can differ from one application to another, they all share the same problem of finding an optimal path (or signal) to perform the specified task. Finding this optimal trajectory is a challenging issue since it depends on both having an accurate mathematical model of the UAV, and designing the desired trajectory based on this dynamical model. %The former is usually tackled by estimating the nonlinear model with a locally linear model around the desired states, while the latter is mostly relaxed by defining the desired path with a polynomial or spline curves, and then designing a controller to follow it. However, there still remains some ambiguities about the accuracy and performance of the result. In response to these concerns, statistical modeling approaches have proved to be interesting alternatives to classical control and planning approaches for modeling of the intrinsic dynamics of the robot's body when it cannot be well estimated. Furthermore, within the framework of programming by demonstration (PbD), statistical methods have been proposed as means to learn a generic trajectory across sets of demonstration. In this work, we implemented an algorithm based on PbD to estimate both dynamics of the UAV and to infer the underlying maneuver. The main advantage of the proposed algorithm over our previous works lies in the fact that with this modeling approach, the effect of robot's dynamics is taken into account.

Related material