Decision-Theoretic Layered Robotic Control Architecture

One of the current methods for developing task control software for robots is a layering approach. This approach generally consists of a symbolic planner, a task sequencer, and a behavioral robotic controller. The task sequencer is responsible for taking a command from an abstract plan and selecting which robot level actions and behaviors to execute. This representation leads to a robust functioning software control for a robot and a single task [Bonnasso and Kortenkamp, 1996]. When the robot must be reconfigured for a new task, elements must be added to the sequencer, and behaviors to the behavioral controller. We are currently developing a decision-theoretic planner to function as the planning and sequencing layers for the architecture. It is our expectation that using a decision-theoretic planner as the sequencer will reduce the amount of work to reconfigure for a new task. We will verify the reconfigurability of our system by creating one set of behavior controllers for our robots and demonstrating the effectiveness of the controllers on multiple diverse plans. Nourbakhsh has implemented a similar system using a symbolic planner in which all planning was abstracted into three levels [Nourbakhsh, 1997]. We choose to incorporate a decision-theoretic planner instead of a symbolic one to make tradeoffs between risk and desire. In addition, this representation allows us to formally reason about the uncertainty that is inherent with robot tasks. Our planner, DT-Graphplan, adds decision theory into the Graphplan algorithm, extending the domain to handle contingent and probabilistic actions as well as utility driven search. This is an extension of the recent work conducted on extending Graphplan to handle probabilities [Weld et al. 1998]. We incorporate utility reasoning into the existing multiple world approach that represents the effect of each action in all possible worlds. Instead of specifying a goal criterion, a minimum acceptable utility threshold is set for the planner. The planner searches for a plan that meets this minimum threshold, pruning world state with low utility values. Requests made to the robot are not represented as goals, but receive a utility commensurate with their value, and instigate replanning. Certain elements in a robotic domain are best represented with decision-theoretic methods. One such