An On-Line Planner for MARIE

In this paper we introduce an on-line planner implemented on MARIE, a real autonomous robot. Planning occurs at a high level, the planner controls other modules such as the path planner, which contain planners themselves. Plans are represented in a tree structure. The on-line planner has a library of partial trees to choose from when replanning is needed. We think that our research is unique in applying on-line planning at such a high level on a real robot. The on-line planner uses a simple algorithm based on elements of POP [10] in combination with random choice. The random choice is essential for real world applications with information errors, because it prevents the robot from taking the ’best’ choice repetitively. Despite being simple the algorithm performs very well in practice. Tests show that the on-line planner gives a major increase in robustness and reliability while not being significantly slower than perfect algorithms.