Enabling Multiple Execution Agents

1. Introduction This extended abstract describes work at the University of Salford Mobile Robotics Lab with a hybrid architecture which combines a symbolic predictive planner-embodied in a Reflective Agent with no sensing capabilities-with a number of non-symbolic reactive (in fact Behav-ioural) agents-embodied in multiple cooperating mobile robots. As described below, the behavioural agents (BAs) are constructed using the Behavioural Synthesis Architecture (BSA) created in Salford (Barnes&Gray 91, Barnes 96), while the Reflective Agent uses a standard non-linear planning system, currently ZENO from the University of Washington (Penberthy&Weld 94). These components are described below in section 3. The motivation for MACTA was to try to combine the strengths of predictive planning-efficiency , coherence and high-level user interface-with the strengths of a behavioural system-resilience and responsiveness. From a planning perspective this involves integrating planning and execution, and the project team has drawn a number of lessons about how some of the problems inherent in this relationship may be tackled which are discussed in section 5. A number of the domain characteristics outlined in the symposium call are relevant to this work. Real-world (that is, not simulated) robotic domains are inherently difficult for well-known reasons. Firstly, complete information is out of the question, either for the planning or executing systems. Accurate knowledge of the current state of the domain can only be gathered via sensors and these are practically and theoretically limited in scope. Only the part of the domain within range can be perceived and even here sensor data is noisy and difficult to interpret. The construction of higher level models from raw data is fraught with problems so that the prerequisite for correct symbolic planning-the existence of an 8 accurate and complete symbolic model-is usually far from being met. This might not be so bad if the domains were completely static, for then it might be feasible to build an accurate model off-line with static, calibrated sensors. However no useful mobile robot domain is that static in practice-changes outside the control of the robot itself are inevitable. For this reason reactive capabilities are nearly always required to allow the robot to cope with the world as it is not as it was predicted to be. Mobile robot domains thus exhibit both complexity and uncertainty. In the MACTA project, domains also include a further complication: multiple cooperating robots. This means that concurrency and coordination are also important issues. A further issue-not discussed …