Here - Agents Lab - University of Nottingham

It has been argued that building robot systems for environments in whichrobots need to co-exist and cooperate with humans requires taking a cognitivestance [12] . According to [12], translating the key issues that such robot systemshave to deal with requires a cognitive robot control architecture. Takinga cognitive stance towards the design and implementation of a robot systemmeans that such a system needs to be designed to perform a range of cognitivefunctions. Various cognitive architectures, such as ACT-R [13] and SOAR [14],have been used to control robots. These architectures were not primarily aimedat addressing the robot control problem and in this sense are similar to agentprogramming languages, the technology that we advocate here for controllingrobots. SOAR has been used to control the hexapod HexCrawler and a wheeledrobot called the SuperDroid [3]. ADAPT (Adaptive Dynamics and Active Perceptionfor Thought) is a cognitive architecture based on SOAR that is specificallydesigned for robotics [15]. The SS-RICS (Symbolic and Sub-symbolic Robotic IntelligentControl System) architecture for controlling robots is based on ACT-R;SS-RICS is intended to be a theory of robotic cognition based on human cognition[16, 2]. Unlike [17], we are not mainly concerned here with the long-termgoal of developing robotic systems that have the full range of cognitive abilitiesof humans based on a cognitive architecture such as SOAR. Our work is orientedtowards providing a more pragmatic solution for the robot control problem asdiscussed above. Although our work may contribute to the larger goal that [17]sets (as there are quite a few similarities between BDI-based agents and cognitivearchitectures such as ACT-R and SOAR), we do not address this question.3 Cognitive Robot Control ArchitectureThis section introduces our cognitive robot control architecture. We discuss severalissues including the processing and mapping of sensory data to a symbolicrepresentation, the translation of high-level decisions into low-level motor controlcommands, and the interaction of the various architecture components.3.1 Overall Design of the ArchitectureFigure 1 shows the main components of our architecture, including a symbolic,cognitive layer (Goal), a middle layer for controlling robot behavior (written inC++), and a hardware control layer (using URBI 1 , a robotic programming language).The Environment Interface Standard (EIS) layer provides the technologywe have used to manage the interaction between the symbolic and sub-symboliclayers. EIS provides a tool to deal with the issue that sub-symbolic sensory datais typically noisy, incomplete, quantitative measurements, whereas the symboliclayers need the symbolic representation that supports logical reasoning.The main functions for controlling a robot are placed in the behavioural controllayer. In addition to the functions such as (object) recognition, navigation,1 http://www.urbiforge.org/57