Learning Topological Maps from Sequential Observation and Action Data under Partially Observable Environment Takehisa Yairi, Masahito Togami, and Koichi Hori ResearchCenterforAdvancedScienceandTechnology,UniversityofTokyo 4-6-1 Komaba, Meguro-ku, Tokyo 153-8904, Japan {yairi, togami, hori}@ai.rcast.u-tokyo.ac.jp Abstract. A map is an abstract internal representation of an environ- ment for a mobile robot, and how to learn it autonomously is one of themostfundamentalissuesintheresearchfieldsofintelligentrobotics and artificial intelligence. In this paper, we propose a topological map learning method for mobile robots which constructs a POMDP-based discretestatetransitionmodelfromtime-seriesdataofobservationsand actions.Themainpointofthismethodistofindasetofstatesornodes ofthemapgraduallysothatitminimizesthethreetypesofentropiesor uncertaintiesofthemapabout“whatobservationsareobtained”,“what actions are available” and “what state transitions are expected”. It is shown that the topological structure of the state transition model is ef- fectivelyobtainedbythismethod. 1 1 Introduction Map learning problem of autonomous mobile robots has attracted a number of researchers in the two fields of robotics and artificial intelligence for many years. From the former viewpoint, metric map construction methods such as occu- pancy grid map[6] and object location map[9] have been mainly studied. The main purpose of the metric maps is to capture the quantitatively accurate fea- tures of the environment geometry. Therefore, it requires a lot of a priori knowl- edge such as a quantitative computation model to estimate the geometric fea- tures from the robots’ sensor inputs. On the other hand, from the viewpoint of artificial intelligence, topological map construction methods have been actively studied [5, 4, 3, 8, 11]. A topologi- cal map is represented as a graph structure, where the nodes correspond to some characteristic or distinctive places the robot visited, and the arcs correspond to the travel paths or motor behaviors connecting the places. Topological map learning is important for artificial intelligence research because it is closely re- lated to the issue of abstraction or internal representation acquisition based on the interaction between the robots’ sensorimotor system and the environment. 1 Proceedings of 7th Pacific Rim International Conference on Artificial Intelligence (PRICAI2002),Tokyo,August,2002,pp.305–314