Reactive Model-based Programming of Mobile Scouts

There is a growing desire in many scientific and exploration communities to use cheap, small, autonomous vehicles to perform information gathering tasks. As the number of vehicles used for information gathering tasks increases from a single large, capable, often remotely operated vehicle to a collection of smaller, more specialized, autonomous vehicles, so too does the complexity of the planning problem and burden on the scientists and operators increase. When multiple vehicles are involved, there are often coordination and safety constraints that need to be accounted for and the resulting plan is frequently more complex as the vehicles need to work together. We envision a system that would allow these information gathering agents, or scouts, to be goal-directed and interact with their human operators at a cognitive level. Such a system should have several properties. First, it should allow the operators to specify the desired behavior of the scouts directly in terms of goals on hidden state, e.g. “goal=have picture of building A”, and let the scout worry about making the plan to achieve that goal with a high level of certainty. Second, it should allow operators to specify constraints at a high-level, such as “do not enter area B after 1500” or “land within one hour of launch”. Third, it should allow operators to have an adjustable level of control over the system. In some circumstances the operator may know exactly what the scout should do, but in other situations, especially in situations where the operator cannot see live data from the scout, it would be better for the operator to describe a set of strategies (as macro actions) and let the scout decide when and where to apply the strategies. The ability to describe macro actions relevant to the task at hand would have the dual benefit of decreasing planning complexity and increasing operator trust in scout. Scouts using such a system would have the potential to be indispensable tools for a wide variety of human activities. For the physical sciences, mobile robots could be used to map pollution levels in the atmosphere or sea while automatically responding to interesting or unexpected readings. During time critical missions, devices with this technology could focus their efforts on gathering the information most pertinent to mission success. For example, a UAV tasked to assist fire fighters during a forest fire could initially work on mapping the entire fire. But, if the fire fighters’ escape route is cut off, the UAV’s mission would change to finding a new escape route and it would accordingly change its focus to mapping the areas of the region that show the best promise for a safe path. In this paper, we first describe a recent real-world deployment of autonomous underwater vehicles as a concrete motivating example. Then, we briefly describe how temporal plan networks can be used to model much of the problems of interest.