We propose to use anthropomorphic characters with limited cognitive ability as a milestone of Artificial General Intelligence (AGI) research. We think " Minidora ", a character who appears in the anime " Doraemon " is suitable as our current challenge. Though Minidora cannot speak natural language, we assume Minidora share our mental model. Since its character requires human assistance, researches based on its characteristics are desirable from both technical possibility and social value. In this paper, we will illustrate several examples of cognitive ability Minidora holds, then show the relationship between existing researches, respectively. Especially, we focus on Minidora as a platform of new direction in Human-Agent Interaction and Pattern Recognition researches.

We propose to use anthropomorphic characters with limited cognitive ability as a milestone of Artificial General Intelligence (AGI) research. We believe, small robot capable of nonverbal expression using movement of arm, facial expression and move around are effective to achieve this milestones. In this paper, we discuss the design of small communication robot can perform nonverbal expression, and report the hardware design and development status.

We communicate naturally by predictively recognizing utterance based on situation, context, and knowledges. We think that predictive recognition makes robots without ability to speak natural language possible to do linguistic communication. In this paper, we focus on Shiritori, a word game, as a simple linguistic communication to verify the hypothesis. Shiritori restricts players' utterance because of the rules, and also has several typical reply patterns. Therefore, we can easily predict partner's utterances compared with normal conversations. In order to reply predictable words, we construct kindergartener-level vocabulary. Furthermore, we make audio data with the cooperation of voice actors for the rich expression of the words. Experimental result implied that we can play Shiritori even if a player does not speak natural language.

Recent developments in deep learning have been remarkable, from the field of image processing to the field of speech recognition and natural language processing has been penetrated and developed. In this study, we first picked up the following three main approaches to implement sentence generation. 1) Markov chain, 2) automatic summary, 3) sentence generation by deep learning (RNN / LSTM / GAN). As a subject, it was commonly seen that the connection between sentence and sentence was unnatural. We tried the connection between natural sentences and sentences which are also applicable in practice by the above three methods and considered countermeasures.

With the recent advancements and developments of deep learning techniques, 'reinforcement learning,' a framework based on the interaction between an agent and the environment, attracts a great deal of attention. This presentation introduces a universal agent model called AIXI (AIξ) proposed by Marcus Hutter (references 1,2). The AIXI model is based on the algorithmic information theory founded by Ray Solomonoff and uses a universal prior distribution in the agent optimization strategy. Based on this formulation, the AIXI can be interpreted as an agent model that can take an optimal strategy under any circumstances. The formulation of such universal agent is an attempt to answer the fundamental question of universal intelligence and may give some hints on how to deepening the reinforcement learning.

We give an introduction of Solomonoff's universal induction, or algorithmic probability. The existence of universal prior (or computability) is the key of his result, which explains many aspects in artificial intelligence and philosophy of science. This introduction especially focuses on Solomonoff's view of probability.

Under large state and action spaces, it is difficult for a reinforcement learning agent to learn the agent's policy within a practical time. Previous studies have proposed methods in which a trainer gives better actions to a trainee to promote the learning. However, when action spaces of a trainer and a trainee is not the same, the instruction does not work without mapping from the instruction to the trainee's variable space. In this paper, we deal with three types of instruction: action-based expression, abstract expression from a human trainer, and expression output by Instruction-based Behavior Explanation, which is a framework to announce a reinforcement learning agent's future behavior. The three instructions were mapped to agents' action spaces with deep reinforcement learning, and we compared the mappings to consider the form of information towards heterogeneous agents' instruction.

Semi-autonomous telepresence robots are proposed by some researches. However, it is insufficient that guidelines defining which behavior should be autonomized on telepresence robots. In our previous research, we proposed a guideline of autonomization focusing on voluntary behavior and involuntary behavior. We extended the architecture focusing on those two behaviors for a telepresence robot and evaluated the impression. However, an evaluation from the viewpoint of remote operator is insufficient. In this paper, we investigated the influence through experiments that participants remotely attend meetings with the telepresence robot. From the results, it was suggested that autonomization of voluntary behaviors could be a factor of unpleasantness for remote operator.

In recent years, reinforcement learning motivated by curiosity has attracted attention. By learning features related to agents' behavior, they can focus on changes in things that are interesting from images, without being distracted by screen noise or meaningless changes. In addition, it has been reported that even with the learning result based only on the internal compensation, it has the generalization ability to be effective even at the stage other than the stage used for learning. However, in the reference paper, it was only applied to 2 stages in the "Super Mario Bros." Therefore, to further investigate the generalization ability of reinforcement learning motivated by curiosity, I experimented on more stages. As a result, I confirmed the generalization ability of reinforcement learning motivated by curiosity.

Currently in research and development, the most important key technology is general artificial intelligence (AGI). If Japan would succeed in development, we will accomplish the Fourth Industrial Revolution and realize a major breakthrough in productivity. However, even in the private sector, the scientists and academic societies, this most important factual relationship is still hardly understood among themselves. It is necessary promptly raise awareness and consciousness. For reforming in the private sector is necessary. To solve this problem, we believe our private volunteers that offers the research and development platform "AGI R&D DAO(Decentralized Autonomous Organization)"applying block chains of completely new ideas would be the most effective. In addition, we will promote the project to make manga characters "Doraemon" who everyone knows, and the educational business to educate AGI. Through these three projects, we will select and support the most effective research activities so that AGI can succeed in development at the shortest and efficient manner.