University of Freiburg
Integrating Task and Motion Planning in Real-World Systems
We will illustrate how our task and motion planner is integrated as a high-level executive for mobile manipulation domains. Symbolic planners are well suited to join the multiple diverse skills of modern robotic systems into one goal-directed agent. A purely symbolic abstraction of the continuous geometric world is usually to shallow, which is why we prefer an integrated task and motion planning approach to generate executable plans. However, this only solves part of the problem as we are also acting in the real-world. The inherent partial observability means that explicitly planning for sensing is required. In addition, we need to monitor action outcomes and replan, if necessary, to be able to react on unexpected outcomes of actions. We introduce our solution to this problem embedding our planner Temporal Fast Downward/Modules (TFD/M) into a continual planning loop and demonstrate its applicability on the PR2 robot in a household scenario. We also consider, which assumptions one needs to rely on in real-world scenarios to be solvable in general and for our setting in particular.