Logic-based subsumption architecture

We describe a logic-based AI architecture based on Brooks' subsumption architecture. In this architecture, we axiomatize different layers of control in First-Order Logic (FOL) and use independent theorem provers to derive each layer's outputs given its inputs. We implement the subsumption of lower layers by higher layers using nonmonotonic reasoning principles. In particular, we use circumscription to make default assumptions in lower layers, and nonmonotonically retract those assumptions when higher layers draw new conclusions. We also give formal semantics to our approach. Finally, we describe layers designed for the task of robot control and a system that we have implemented that uses this architecture for the control of a Nomad 200 mobile robot. Our system combines the virtues of using the represent-and-reason paradigm and the behavioral-decomposition paradigm. It allows multiple goals to be serviced simultaneously and reactively. It also allows high-level tasks and is tolerant to different changes and elaborations of its knowledge in runtime. Finally, it allows us to give more commonsense knowledge to robots. We report on several experiments that empirically show the feasibility of using fully expressive FOL theorem provers for robot control with our architecture and the benefits claimed above.