We have been developing and researching chat robot for elderly people. This paper proposed category estimation for voice based chat robot system. Proposed method estimates a category of the user's utterance sentence. Category or topic estimation have been studied by many researchers, however voice base chat system needs a quick response. Therefore our system is simple method just using keyword match and previous dialogue history. This paper analyzed the validity of category estimation compares with manual annotation. Category estimation is a difficult task even if human, then each sentence were categorized by three annotator to 18 categories. In this research, we collected 13727 dialogue sentences from an experiment with ten elderly participants. As the result, 12783 sentences can be categorized (2 or 3 annotator categorized for the same category). We compare the categorized sentences and robot's estimated category. The robot categorized as same as human around 50%. This coincident ratio is depending on the category. “Society” category is difficult to categorize, but “go-out” and “music” can be categorized as around 60% ratio. In the future works, we have to analyze the similarity around 18 categories and how to control the number of categories.

The authors developed cloud based chat robot for exhilarating elderly people. Our robot is designed to keep users from being bored. Conventional chat robot systems use huge phrase database in advance. However, fixed phrase involves an immense amount of time and effort to make. The proposed system uses dialogue histories which are collected from users. This paper reports the effect of updating the phrase database by using histories.

Human commonsense is required to improve quality of robotic application. However, to acquire the necessary knowledge, robot needs to evaluate the appropriateness of the data it has collected. This paper presents an evaluation method, by combining the weighting mechanism in commonsense databases with a set of weighting factors. The method was verified on our Basic-level Knowledge Network. We conducted questionnaire to collect a commonsense data set and estimate weighting factors. Result showed that, the proposed method was able to build Robot Technology (RT) Ontology for a smart “Bring something” robotic service. More importantly, it allowed robot to learn new knowledge when necessary. An intuitive human-robot interface application was developed as an example base on our approach.

This paper proposes our approach to disambiguate human pointing gesture by using commonsense knowledge. Pointing gesture is a natural way to specify the object that a user needs in the “Bring something” service. However, error in gesture recognition makes robot difficult to select the correct candidate object when there are multiple objects located closely together. To solve this problem, Robot Technology (RT) Ontology was used to make assumption about human activity and necessary objects for the activity. By combining this commonsense reasoning with the spatial constrain from human's pointing direction, robot is able to find appropriate object. Experiment showed the successful rate of object selection was improved, and therefore robotic service becomes more intuitive.

Recent research on hidden-state reinforcement learning (RL) problems has been concentrated in overcoming partial observability by using memory to estimate states. Switching Q-learning (SQ-learning) is a novel memoryless approach for RL in partially observable environments. The basic idea of SQ-learning is that "non-Markovian" tasks can be automatically decomposed into subtasks solvable by memoryless policies, without any other information leading to "good" subgoals. To deal with such decomposition, SQ-learning employs ordered sequences of Q-modules in which each module discovers a local control policy. Furthermore, a hierarchical structure learning automaton is used which finds appropriate subgoal sequences. We apply SQ-learning to three partially observable maze problems. The results of extensive simulations demonstrate that SQ-learning has the ability to quickly learn optimal or near-optimal policies without huge computational burden.