||Eglin Air Force Base, FL 325426810
The guidance, navigation, and control of teams of unmanned aerial vehicles (UAVs) presents many new challenges. Research opportunities are available in the area of integrated estimation-control algorithms that enable teams of autonomous vehicles to collaboratively estimate their location and the state of their environment (including estimating the positions, velocities, and orientations of potential targets and obstacles). Particle filters offer a very general and flexible method of incorporating soft information sources (e.g., human verbal cues, non-standard computer vision outputs) and they permit nonlinear and non-gaussian process and noise models. For this reason, particle filters might enable complex map-based information-fusion to improve target and ego state estimation; methods for sharing particle filter representations of uncertainty across a team of UAVs is a current research thrust.
Given a particle filter-based estimate, there is an open problem in how to design a control law or path planner that takes a set of particles as an input and returns a command signal as an output: this is a radical departure from the usual seperation principle or certainty-equivalence paradigms that are invoked to decouple estimation from control development. Methods and algorithms for exploiting the rich, non-Gaussian representation afforded by particle filters in the construction of control laws is a current research thrust. One possible technique is to grow a random dense tree over the obstacle-free state space and use a heuristic criterion to choose a good control signal which may be implemented over a small time interval; then a new tree might be grown and a new branch chosen in a receding-horizon framework. One possible method of incorporating a particle filter is to bias the growth the random tree over the particle cloud in order to generate a large number of obstacle-free trajectories that tend to fly over regions of large uncertainty. Other Monte-Carlo based control strategies might be desirable given the sample-based nature of a particle representation.
Finally, a central issue and constraint of future collaborative systems is the limited storage and bandwidth available for information flow control. There is a dearth of theory for understanding informational value and for discerning which pieces of data are most crucial for informed decision making by agents connected in a dynamic communication topology. We are interested in formulating ways of tagging data based on its utility to decision makers and we are investigating optimal encode-decode policies that are "aware" of decision-controller objectives. The ability to communicate objectives and preferences and to exploit high-level cognitive goals at the communication or sensor level might enable much higher total system performance by enabling context aware information flow management. Understanding informational value and the intersection of communication protocols, network topology, estimation uncertainty, and decision-maker objectives is a current research topic.