UC Berkeley

Small UAVs have shown great promise as tools for collecting aerial imagery both quickly and cheaply. Furthermore, using a team of small UAVs, as opposed to one large UAV, has shown promise as being a cheaper, faster and more robust method for collecting image data over a large area. Unfortunately, the autonomy of small UAVs has not yet reached the point where they can be relied upon to collect good aerial imagery without human intervention, or supervision. The work presented here intends to increase the level of autonomy of small UAVs so that they can independently, and reliably collect quality aerial imagery.

The main contribution of this paper is a novel approach to controlling small fixed wing UAVs that optimizes the quality of the images captured by cameras on board the aircraft. This main contribution is built on three minor contributions: a kinodynamic motion model for small fixed wing UAVs, an iterative Gaussian sampling strategy for rapidly exploring random trees, and a receding horizon, nonlinear model predictive controller for controlling a UAV's sensor footprint.

The kinodynamic motion model is built on the traditional unicycle model of an aircraft. In order to create dynamically feasible paths, the kinodynamic motion model augments the kinetic unicycle model by adding a first order estimate of the aircraft's roll dynamics. Experimental data is presented that not only validates this novel kinodynamic motion model, but also shows a 25% improvement over the traditional unicycle model.

A novel Gaussian biased sampling strategy is presented for building a rapidly exploring random tree that quickly iterates to a near optimal path. This novel sampling strategy does not require a method for calculating the nearest node to a point, which means that it runs much faster than the traditional RRT algorithm, but it still results in a Gaussian distribution of nodes. Furthermore, because it uses the kinodynamic motion model, the near optimal path it generates is, by definition, dynamically feasible.

A nonlinear model predictive controller is presented to control the non-minimum phase problem of tracking a target on the ground from a UAV with a fixed camera. It is shown that this novel controller is probabilistically guaranteed to asymptotically converge to the path that minimizes the cross-track error of the UAV's sensor footprint. In addition, for a minimum phased problem, it is shown that its tracking performance is on par with a sliding mode controller, which at least theoretically, is capable of achieving perfect tracking.

Finally, all three of these contributions are experimental validated by performing a variety of tracking tasks using the Berkeley Sig Rascal UAV.