Institut für
Informatik, Universität Freiburg
Autonomous operation of an agent in the real world can be made much more efficient and robust when taking into account that failures can occur. One important aspect of failure recovery is replanning, and we will discuss an approach for reducing replanning effort. First, we introduce a general framework in which one can formally deal with goals, actions, failures, and plans. Within this framework, we formulate our approach which is also well-suited for applications where one can be content with the satisfaction of only some parts of a goal specification. Furthermore, we will show how this approach can be implemented efficiently in a situation calculus framework. The central notion is that of a abeled dependency graph, which can be used to reveal the connection between actions and parts of a goal specifications. Given an action sequence generated by a GOLOG program, we finally discuss a heuristic that can be used to obtain so-called partial goal specifications, which are vital for the construction of labeled dependency graphs.