Abstract: Unmanned aerial systems (UAS) are already actively used in the military domain, predominantly in Information, Reconnaissance and Surveillance missions. Future projections anticipate a growing demand for civil applications as well. A prerequisite for a broader development of these systems is to increase their level of autonomy. However, developing UAS controllers that require only a minimal degree of human supervision and work in a realistic environment is an extremely challenging task. This talk will discuss our current work on some aspects of this problem.
Consider M UAVs monitoring N>M sites, collecting ”rewards” as they visit the sites. The states of the sites evolve as independent Markov chains, and the goal of the controller is to decide where to send the UAVs at each period. In certain restricted situations, with only one vehicle, this problem can be solved optimally using Gittins’ dynamic allocation indices for the multi-armed bandit problem. For more interesting scenarios, the problem quickly becomes provably intractable. We present relaxation techniques leading to performance bounds and efficiently computable policies approaching these bounds for problems with multiple agents, partial information, and additional costs for traveling between the sites.
The last part of the talk will be concerned with the traditional hierarchical architectures used for tractable designs of autonomous robot controllers. Typically, the decision-making module, such as the task scheduler considered above, generates a sequence of waypoints for the robot, without taking into account the differential constraints. However, there seems to be much performance gain to be expected if we can incorporate even an approximate model of the dynamics of the vehicle at the planning level. In this context, we present some recent work on the traveling salesman problem for the Dubins’ car, which is a useful approximation of the dynamics of fixed wing UAVs.