Integrating Deliberation in an Intelligent Agent Architecture

We have been pursuing the integration of a state-based planner as the deliberative top layer of a three-tiered robot control architecture. The middle layer is implemented in the RAP system, while the bottom layer consists of a suite of reactive skills which can be configured by the RAPs system into Brooksian machines. Our target environment has been the ground control center of a space station command center. Our current system has been shown to provide a higher level of human supervision that preserves safety while allowing for task level direction, reaction to out-ofnorm parameters, and human intervention at all levels of control. For this workshop, we discuss the issues we have been addressing with regard to where the boundary lies between local reactive control and longrange deliberation. Such issues include whether the seat of control should be at the RAPs ystem or at the planner, interruption of the RAPs system by the planner due to priority changes, how much of the planner’s instruction can be taken as guidance versus direction, and levels of human interaction between the planner and the RAPs ystem. Background and Motivation Since the late eighties we have investigated ways to combine deliberation and reactivity in robot control architectures [Sanborn et ai 1989, Bonasso 91, & Bonasso et al 92], in order to program robots to carry out tasks robustly in field environments. The robot control software architecture we are using is an outgrowth of several lines of situated reasoning research in robot intelligence [Firby 89, Gat 91, Connell 91, Slack 92, Yu et al 94, Elsaesser & Slack 94], and has proven useful for enabling mobile robots to accomplish tasks in field environments. This architecture, which is similar to ATLANTIS [Gat 91] and Cypress [Wilkins et al 94] in its concept if not in its implementation, separates the general robot intelligence problem into three interacting pieces (see Figure 1): o A set of robotic specific reactive skills. For example, grasping, object tracking, and local navigation. These are tightly bound to the specific hardware of the robot and must interact with the world in real-time. o A sequencing capability which can differentially activate the reactive skills in order to direct changes in the state of the world and accomplish specific tasks. For example, exiting a room might be orchestrated through the use of reactive skills for door tracking, local navigation, grasping, and pulling. In each of these phases of operation, the skills of the lower-level are connected to function as a what might be called a "Brooksian" robot a collection of networked state machines. We are using the Reactive Action Packages (RAPs) system [Firby 89] for this portion of the architecture. o A deliberative planning capability which reasons in depth about goals, resources and timing constraints. We are using a state-based non-linear hierarchical planner known as AP [Elsaesser & MacMillan 91], the adversarial planner, since it grew out of research in counter planning. AP is a multi-agent planner which can reason about metric time for scheduling, monitor the execution of its plans, and replan accordingly. These capabilities allow a robot, for example, to accept guidance from a human supervisor, plan a series of activities at various locations, move among the locations carrying out the activities, and simuitaneously avoid danger and maintain nominal resource levels. We have been successful in applying the reactive portion of this architecture to mobile land [e.g., Bonasso et al 92] and undersea robots [Bonasso & Barrett 93]. Recently, we have integrated the planner to control a free-flying, twoarmed manipulator system in support of the maintenance of NASA’s planned space station. This paper discusses issues associated with integrating planning with reactive control. Such issues include whether the seat of control should be at the RAPs ystem or at the planner, interruption of the RAPs ystem by the planner due to priority changes, how much of the planner’s instruction can be taken as guidance versus direction, and levels of human interaction between the planner and the RAPs ystem. Automated Robotic Maintenance of Space