ROGUE is fully implemented and operational. The system completes all requested tasks, running errands between offices in our building. Execution results are presented in detail elsewhere [Simmons et al. 1997].
We have presented one aspect of ROGUE, an integrated planning and execution robot architecture. We have described how PRODIGY4.0 gives ROGUE the power
ROGUE represents a successful integration of a classical artificial intelligence planner with a real mobile robot. The complete planning and execution cycle for a single task can be summarized as follows:
Figure 9 summarizes the information exchanged between the user, PRODIGY4.0, and Xavier under ROGUE's mediation. ROGUE constrains PRODIGY4.0's decisions through calculations on task priority, task compatibility, and execution efficiency. ROGUE translates PRODIGY4.0's symbolic action descriptions into Xavier commands, and also translates Xavier's perception information into PRODIGY4.0 domain description.
Figure 9: What ROGUE does.
The contributions of our work to the Xavier project are in the high-level reasoning parts of the system, allowing the robot to efficiently handle multiple, asynchronous interacting goals, and to effectively interleave planning and execution in a real world system. Execution monitoring based on a planning model allows the systematic identification of environment monitors.
Interleaving planning with execution enhances a deliberative robot system in numerous ways. One such benefit is that the system can sense the world to acquire necessary domain knowledge in order to continue planning. For example, it can ask directions, look to see if doors are open or closed, or check whether it needs to recharge its batteries. Another benefit is reduced planning effort because the system does not need to plan for all possible failure contingencies; instead, it can execute an action to find out its actual outcome.
ROGUE advances the state of the art of the integration of planning and execution in robotic agent. In a unique novel way, ROGUE is designed as the integration of two independently developed platforms. PRODIGY4.0 is a general-purpose planner and Xavier can be viewed as a general-purpose navigational robot. ROGUE merges the functionality of these two systems in a real implementation that demonstrates the feasibility of connecting both systems in a rich task environment, namely the achievement of asynchronous user requests. (ROGUE therefore also shows how the PRODIGY4.0 planner and the TCA approach in Xavier are in fact robust architectures.)
Strictly looking at ROGUE only from the viewpoint of the integration of
planning and execution, ROGUE compares well with other special-purpose
systems such as NMRA and 
. Given the general-purpose character of the
PRODIGY4.0 planner, ROGUE could easily be applied to other executing
platforms and tasks by a flexible change of PRODIGY4.0's specification of
the domain.
The goal of our research is to build a complete planning, executing and learning autonomous robotic agent. ROGUE's contributions go beyond the integration of planning and execution. ROGUE incorporates learning from execution experience [Haigh & Veloso1998]. The learning algorithm involves extracting relevant information from real execution traces in order to detect patterns in the environment to improve the robot's behaviour. The ability to learn task-relevant knowledge conveniently matches PRODIGY4.0's search control representation, and learned situational-dependent arc costs can be incorporated into Xavier's route planner knowledge. Through its interleaved planning and execution behaviour, ROGUE provides an appropriate platform to collect the required execution data for learning.
Karen Zita Haigh