A new prototype of a self-reconfigurable modular robot, M-TRAN III, has been developed, with an improved fast and rigid connection mechanism. Using a distributed controller, various control modes are possible: single-master, globally synchronous control or parallel asynchronous control. Self-reconfiguration experiments using up to 24 modules were undertaken by centralized or decentralized control. Experiments using decentralized control examined a modular structure moved in a given direction as a flow produced by local self-reconfigurations. In all experiments, system homogeneity and scalability were maintained: modules used identical software except for their ID numbers. Identical self-reconfiguration was realized when different modules were used in initial configurations.

M-TRAN is a self-reconfigurable modular robot: each module has an independent battery, two-degree-of-freedom motion, six-surface-connection capability, and intelligence with inter-module communication. The M-TRAN system can perform flexible and adaptive locomotion in various configurations using coordination control based on a central pattern generator (CPG). Various structures of several modules can perform metamorphosis, such as that between a four-legged robot and a snake-like one.In addition to these self-reconfigurations with synchronous control, M-TRAN structures having regularity can move using parallel distributed control and message exchange via the network bus. Self-reconfiguration using infrared local communication has been attempted to improve the system's scalability.

By manipulating objects in their environment, in fants learn about the surrounding environment and continuously improve their internal model of their own body. Moreover, infants learn to distinguish parts of their own body from other objects in the environment. In the field of neuroscience, studies have revealed that the posterior parietal cortex of the primate brain is involved in the awareness of self-generated movements. In the field of robotics, however, little has been done to propose computationally reasonable models to explain these biological findings. In the present paper, we propose a generative model by which an agent can estimate appearance as well as motion models from its visuomotor experience through Bayesian inference. By introducing a factorial representation, we show that multiple objects can be segmented from an unsupervised sensory-motor sequence, single frames of which appear as a random patterns of dots. Moreover, we propose a novel approach by which to identify an object associated with self-generating action.

Abstract. Dynamics of traditional neural network models are generally time-invariant. For that reason, they have limitations in contextdependent processing. We present a new method, dynamic desensitization, of varying neurodynamics continuously and construct a basic model of interaction between neurodynamical systems. This model comprises two nonmonotone neural networks storing sequential patterns as trajectory attractors. The dynamics of respective networks are modified according to the states of other networks. Using numerical experiments, we also show that the model can recognize and recall complex sequences with identical patterns in different positions.