On the Role of Stored Internal State in the Control of Autonomous Mobile Robots

of local sensor data (for example, Nilsson [1980]). Reactive approaches, brandishing slogans such as “the world is its own best model,” prefer local sensor data and tend to avoid the use of internal state whenever possible (for example, Connell [1989]). Reactive approaches arose as a sort of “popular revolt” backlash against the shortcomings of traditional sense-plan-act architectures. When the slow, plodding operation of traditional approaches proved ineffective at dealing with the realities of reality, Rodney Brooks turned the problem sideways: Instead of general-purpose functional modules, Brooks proposed an architecture based on special-purpose, task-oriented modules that coupled sensors directly to actuators. The dramatic result of Brooks’s work was that many tasks that had previously been thought to be difficult (most notably, the task of collision-free navigation) turned out to be achievable using simple control mechanisms (Brooks 1986). In the subsequent euphoria, internal state was somehow transformed into a greater villain than Brooks ever intended.1 No reactive architecture advocates the wholesale elimination of internal state. Internal state is necessary, reactivists readily concede, but a few odd bytes scattered here and there should suffice. It is classical planning, along with its associated centralized world models, abstractions, and large linked data structures, that ■ This article informally examines the role of stored internal state (that is, memory) in the control of autonomous mobile robots. The difficulties associated with using stored internal state are reviewed. It is argued that the underlying cause of these problems is the implicit predictions contained within the state, and, therefore, many of the problems can be solved by taking care that the internal state contains information only about predictable aspects of the environment. One way of accomplishing this is to maintain internal state only at a high level of abstraction. The resulting information can be used to guide the actions of a robot but should not be used to control these actions directly; local sensor information is still necessary for immediate control. A mechanism to detect and recover from failures is also required. A control architecture embodying these design principles is briefly described. This architecture was successfully used to control realworld and simulated real-world autonomous mobile robots performing complex navigation tasks. The architecture is able to incorporate standard AI planning and world-modeling algorithms into a real-time situated framework.