Integrating Task Level Planning with Stochastic Control

Real-world tasks require planners to search for solutions in a partially observable state space. Planning in partially observable domains is computationally expensive—it is impossible to plan for all possible contingencies in advance. Furthermore, real world scenarios introduce a huge variety of context-dependent variation in the dynamics of belief in response to new information. Our approach investigates the use of models of belief dynamics designed to exploit latent structure in the environment in the form of “objects” and we evaluate the possibility of accumulating thousands of such models in a practical, high performance planning system. We evaluate the efficacy of our approach in various tasks using a real robot.