Using Behavior Activation Value Histories for Updating Symbolic Facts Frank Sch¨ onherr and Mihaela Cistelecan and Joachim Hertzberg and Thomas Christaller Fraunhofer – Institute for Autonomous Intelligent Systems (AIS) Schloss Birlinghoven, 53754 Sankt Augustin, Germany Frank.Schoenherr@ais.fraunhofer.de Abstract The paper presents a new technique for extracting symbolic ground facts out of the sensor data stream in autonomous robots for use under hybrid control architectures. The sen- sor data are used in the form of time series curves of behavior activation values, yielding an image of the environment as perceived through the eyes of useful behaviors. Similar progressions in individual behavior activation curves are aggregated to well-defined patterns, like edges and levels, called qualitative activations. Sets of qualitative activations for different behaviors occurring in the same interval of time are summed to activation gestalts. Sequences of activation gestalts are used for defining chronicles, the recognition of which establishes evidence for the validity of ground facts. The approach in general is described, and examples for a par- ticular behavior-based robot control framework in simulation are presented and discussed. 1 Background and Overview There are several good reasons to include a behavior- based component in the control of an autonomous mobile robot. There are equally good reasons to include in ad- dition a deliberative component. Having components of both types results in a hybrid control architecture, inter- twining the behavior-based and the deliberative processes that go on in parallel. Together, they allow the robot to react to the dynamics and unpredictability of its envi- ronment without forgetting the high-level goals to accom- plish. Arkin (Arkin 1998, Ch. 6) presents a detailed argu- ment and surveys hybrid control architectures; the many working autonomous robots that use hybrid architectures in- clude the Remote Agent Project (Muscettola et al. 1998; Remote Agent Project 2000) as their highest-flying exam- ple. While hybrid, layered control architectures for au- tonomous robots, such as Saphira (Konolige et al. 1997) or This work is partially supported by the German Fed. Ministry for Education and Research (BMBF) in the joint project AgenTec (01AK905B). This paper is an update of (Sch¨ onherr et al. 2001). On leave from Tech. University of Cluj-Napoca, Romania. The work was supported by a Roman-Herzog scholarship of the Alexander von Humboldt foundation. Copyright c 2001, American Association for Artificial Intelli- gence (www.aaai.org). All rights reserved. Papers from the 2001 AAAI Fall Symposium, Technical Report FS-01-01, pp. 9 – 16. 3T (Bonasso et al. 1997) are state of the art, some problems remain that make it a still complicated task to build a control system for a concrete robot to work on a concrete task. To quote Arkin (Arkin 1998, p. 207), the nature of the boundary between deliberation and reactive execution is not well understood at this time, leading to somewhat arbitrary architectural decisions. One of the problems is to keep up-to-date the symbolic world representation for the deliberative component, which is an instance of the symbol grounding problem (Harnad 1990). There are solutions to important parts of that prob- lem, such as methods and algorithms for sensor-based lo- calization to reason about future navigation actions: (Fox, Burgard, & Thrun 1999) presents one of the many examples for on-line robot pose determination based on laser scans. If the purpose of deliberation is supposed to be more general than navigation, such as action planning or reasoning about action, then the need arises to sense more generally the re- cent relevant part of the world state and update its symbolic representation based on these sensor data. We call this rep- resentation the current situation. The naive version of the update problem “Tell me all that is currently true about the world!” needs not be solved, luck- ily, if the goal is to build a concrete robot to work on a con- crete task. Only those facts need updating that, according to the symbolic domain model used for deliberation, are rele- vant for the robot to work on its task. Then, every robot has its sensor horizon, i.e., a border in space and time limiting its sensor range. The term sensor is understood in a broad sense: It includes technical sensors like laser scanners, ul- tra sound transducers, or cameras; but if, for example, the arena of a delivery robot includes access to the control of an elevator, then a status request by wireless Ethernet to de- termine the current location of the elevator cabin is a sensor action, and the elevator status is permanently within the sen- sor horizon. We assume: The world state information within the sensor horizon is sufficient to achieve satisfying robot performance. This said, the task of keeping the facts of a situation up- to-date remains to continually compute from recent sensor data and the previous situation a new version of the situation as far as it lies within the sensor horizon. The computation is based on plain, current sensor values as well as histories of situations and sensor readings or aggregates thereof. Prac-