Extending Reward-based Hierarchical Task Network Planning to Partially Observable Environments

Rapid, recent developments in robotic applications demand feasible task planning algorithms capable of handling large search spaces. Hierarchical task network (HTN) planning complies with such demand by extending classical planning with task decomposition. Recent advances have extended HTN planners to include the use of reward functions, increasing their flexibility. Nonetheless, such planners assume a fully observable environment, which is often violated in realistic domains. This work contributes to this challenge by presenting POST-HTN, a tree-search based solver which accounts for partial observable environments. A qualitative comparison of POST-HTN with the PC-SHOP HTN solver is given in multiple domains, such as industrial inspection, which is executed on a mobile robot in the real world.