Distributed problem solving and real-time mechanisms in robot architectures

Abstract Autonomous robots operating in unstructured environments must react predictably to unexpected events, according to some form of specification. Mechanisms included in the design of the functional system architecture for an autonomous robot can help guarantee that such specifications are met. Examples of specifications include behaviours, meeting deadlines after events, or control performance. This paper considers the design of robot architectures using concepts from distributed problem solving (DPS) and real-time knowledge based systems (RTKBS). Qualitative performance measures are identified (coherency, co-ordination and consistency) and styles of problem-solving behaviour (functionally accurate, co-operative, nearly autonomous) within the architecture based on DPS. Characteristics the architecture and functional modules should exhibit if operating in real-time are described based on RTKBS. A real-time blackboard-based agent is described as a basis for experimenting with architectures. To allow the architecture and each functional module to guarantee responses to events within set deadlines (an example of a specification) a mechanism for trading quality of solution with processing time available is described. The technique involves a variation on the AO∗ search algorithm called Pruning AO∗ (PAO∗) which removes over-deadline routes from a directed graph of possible solutions. For illustration, an example architecture involving several such agents is applied to the task of sonar interpretation for an Autonomous Underwater Vehicle. Results are presented showing the performance of the architecture in various configurations in terms of quality and confidence of solutions, time available, problem decomposition, coherency, co-ordination and consistency.