This paper investigates issues in the plan design of cognition-enabled robotic agents performing everyday manipulation tasks. We believe that plan languages employed by most cognitive architectures are syntactically too restricted to specify the flexibility, generality, and robustness needed
to perform physical manipulation tasks. As a consequence, the robotic agents often have to employ flexible plan execution systems. This causes two problems. First, robots cannot understand how their plans generate the flexible behavior and second, they cannot use the mechanisms of flexible plan execution for plan improvement. We report on our research on plan design for robotic agents performing human-scale everyday manipulation tasks, such as making pancakes and popcorn. We will stress three key factors in plan design. First, the use of vague descriptions of objects, locations, and actions that the robot can reason about and revise at execution time. Second, the plan language constructs needed for failure detection, propagation, and handling. Third, plan language constructs that represent and annotate the intentions of the robot explicitly even in concurrent percept driven plans. We clarify our concepts in terms of a generalized pick and place task. In this context, failure handling in uncertain environments is still an open topic for which we demonstrate a solution using our high level plan representation. We illustrate the advantages of our approach in a simulated environment.