Towards more Realistic Logic-based Robot Controllers in the GOLOG Framework

High-level robot control languages should not only be expressive but should also support reasoning about actions, in particular, the projection of robot plans. Projection is useful for the robot when choosing among different courses of action as well as for the designer of robot controllers, since projections allow for qualitative simulations. The high-level programming language GOLOG was specifically proposed for this purpose. The semantics of GOLOG, which offers constructs such as sequences, iterations and recursive procedures, is based on the situation calculus, a logical language for reasoning about action and change. In particular, every primitive GOLOG action is an action of the underlying situation calculus theory, which allows reasoning about the effects of primitive actions or complex GOLOG programs. While GOLOG comes equipped with a powerful projection mechanism, however, it lacks the expressiveness provided by non-logic-based robot programming languages like RPL, RAP or Colbert. In particular, it does not provide facilities for dealing with continuous change, event-driven behavior, and communication with lower-level routines for navigation or localization, to which we refer to as lowlevel processes. Another limitation of GOLOG is that it assumes that actions have deterministic effects and cannot represent probabilistic uncertainty. In realistic domains, however, uncertainty seems to be ubiquitous: a robot has often only probabilistic beliefs about the state of the world, and low-level processes have probabilistic outcomes. In this thesis, we show how the GOLOG framework can be extended to deal with issues like continuous change, event-driven actions and low-level processes in a natural way, thus shortening the gap in expressiveness between non-logic-based and logic-based robot control languages. In particular, we integrate continuous time and change directly in the ontology of GOLOG. To facilitate the actual execution of high-level plans on a real robot, we employ a layered robot control architecture where a high-level controller communicates via message with the low-level processes provided by the basic-task execution system. Our framework allows not only the projection and the actual (on-line) execution of the same plans, but also supports the specification of plans with interleave projection and on-line execution. Furthermore, we provide means to represent and deal with probabilistic beliefs and noisy sensors and effectors. Finally, the extended GOLOG formalism is implemented in PROLOG and evaluated in several experiments, including delivery tasks where the mobile robot CARL operates in the Computer Science Department V at Aachen University of Technology.