Figure 1 shows the general architecture of the system. ROGUE accepts tasks posted by users, calls the task planner, PRODIGY4.0, which generates appropriate plans, and then posts actions to the robot, Xavier, for execution. ROGUE provides appropriate search control knowledge to the planner and monitors the outcome of execution.
Xavier is a mobile robot being developed at Carnegie Mellon University [O'Sullivan, Haigh, & Armstrong1997, Simmons et al. 1997] (see Figure 2).
It is built on an RWI B24 base and includes bump sensors, a laser range finder, sonars, a color camera and a speech board. The software controlling Xavier includes both reactive and deliberative behaviours, integrated using the Task Control Architecture (TCA) [Simmons1994]. TCA provides facilities for scheduling and synchronizing tasks, resource allocation, environment monitoring and exception handling. The reactive behaviours enable the robot to handle real-time local navigation, obstacle avoidance, and emergency situations (such as detecting a bump). The deliberative behaviours include vision interpretation, maintenance of occupancy grids & topological maps, and path planning & global navigation. The underlying architecture is described in more detail by Simmons et al. [1997].
PRODIGY is a domain-independent planner that serves as a testbed for machine learning research [Carbonell, Knoblock, & Minton1990, Veloso et al. 1995]. The current version, PRODIGY4.0 is a nonlinear planner that follows a state-space search guided by means-ends analysis and backward chaining. It reasons about multiple goals and multiple alternative operators to achieve the goals. It reasons about interacting goals, exploiting common subgoals and addressing issues of resource contention.
Xavier reliably performs actions requested of it, but has no task planning abilities. PRODIGY4.0, meanwhile, is a complex task planner that had never been used in a real execution domain; as such, it had never been used for asynchronous goals or in an environment where the state spontaneously changes. In combining the two systems, the challenges for ROGUE include developing a communication mechanism for control and feedback, as well as extending the planner to handle the dynamics of a real-world task.
Karen Zita Haigh