Integrating Golog and Planning : An Empirical Evaluation

The Golog family of action languages has proven to be a useful means for the high-level control of autonomous agents, such as mobile robots. In particular, the IndiGolog variant, where programs are executed in an online manner, is applicable in realistic scenarios where agents possess only incomplete knowledge about the state of the world, have to use sensors to gather necessary information at runtime and need to react to spontaneous, exogenous events that happen unpredictably due to a dynamic environment. Often, the specification of such an agent’s program also involves that certain subgoals have to be solved by means of planning. IndiGolog supports this in principle by providing a variety of lookahead mechanisms, but when it comes to pure, sequential planning, these usually cannot compete with modern state-of-the-art planning systems, most of which being based on the Planning Domain Definition Language PDDL. Previous theoretical results provide insights on the semantical compatibility between Golog and PDDL and how they compare in terms of expressiveness. In this paper, we complement these results with an empirical evaluation that shows that equipping IndiGolog with a PDDL planner (FF in our case) pays off in terms of the runtime performance of the overall system. For that matter, we study a number of example application domains and compare the needed computation times for varying problem sizes and difficulties.