Task planning for social robots has many challenges because most task planners, so far, have required manually designed task knowledge depending on the environment, tasks, and robot specifications. It is difficult to generalize task knowledge for different robots and environments. Also, task knowledge should be explainable to humans to transfer human task knowledge to a robot. We are developing a framework to learn a transferable knowledge system for task planning.

Semantic map represents the world in a way that a human understands the world. Our algorithm can represent known and unknown objects in the map and estimate a meaning of environmet based on common sense knowledge. We also provide human-robot interface that non-expert can improve an imperfect semantic map.

We are developing hierarchical planning algorithms which can handle multiple level of abstractions. Hierarchical planning definitly increases efficiency of robot planning. However, hierarchical planning for language description is not trivial, because human language consists of multiple level of abstractions. We are developing Markov Decision Process (MDP) based hierarchical planning.

Language is an intuitive interface for human-robot interaction. We are intereseted in how to employing language to extract information from a human, how to explain robot's internal space to non-expert humans, and how to translate human language to mathematical expressions for robots