UC Berkeley

This dissertation addresses how a single human operator can interactively control a network of collaborative unmanned aerial vehicles (UAVs). It provides a model of computation for network-level controllers. These network-level controllers enable a single human operator to monitor, order, and supervise the progress of an entire network of autonomous collaborative UAVs. As the human operator observes the network's progress, they may make changes by applying runtime patches to the network-level controller. The resulting network-level controller can be analyzed to verify that it meets required conditions.

UAV networks produce significantly more information than any single human can concentrate on. To ease this cognitive burden, the human operator should interact with the UAVs at an appropriately high level of abstraction. They should focus on monitoring what is being done, deciding what should be done, and determining in what order. Other lower-level decisions can be automated by the network. To this end, Petri nets are used as a theoretical basis for stating network-level controllers. Petri nets have a graphical description that is extremely intuitive. They also convey exactly what is being done, what will be done, and in what order using a network-focused perspective.

Developing a novel Petri net-based model of computation for network-level control is the main contribution of this work. This includes forming the syntax and semantics for network-level controllers. Other related contributions include identifying invariant and analyzable properties of the network-level controllers, proofs of their correctness, and interpretations of their meanings. These allow a human operator to understand and assert that a proposed controller is indeed correct. A runtime patching language to enable modifications by the human operator is another contribution.

The theoretical concepts behind network-level controllers provide a mathematical basis for the more concrete Collaborative Sensing Language (CSL). This formal XML-based language allows a precise specification of network-level controllers for UAVs. This dissertation also describes an implementation that enabled CSL for the UAV fleet at the Center for Collaborative Control of Unmanned Vehicles (C3UV). CSL and its implementation are also original contributions.

Previous related research has focused on the detailed off-line specification of individual reactive UAV behaviors. This creates a fixed preprogrammed network which produces a specific but predetermined behavior. The various existing alternatives are not well suited for interactive control of a network by a single human operator. They either suffer from no graphical description (process algebras), excessive information (hybrid systems), an individual component and not network focus (networks of finite state machines), a fixed dimension network, or a static specification that cannot be affected on-line. The novelty and usefulness of this dissertation lies in its ability to address these issues.