Kousuke Inoue

Iterative Transportation by Cooperative Mobile Robots in Unknown Environment

This paper focuses on a planning method for an iterative transportation task by cooperative mobile robots in an unknown environment. This task requires the acquisition of environmental information, the generation of appropriate robot paths based on the acquired information, and the formation of the group of robots. In order to realize an efficient transportation, a motion planning architecture consisting of ‘environmental exploration phase,’ ‘path-generation phase’ and ‘strategy-making phase’ is proposed. In the first phase, robots explore the environment using a learned visibility graph while transporting. Next, transportation paths consisting of 1-lane paths and 2-lane paths are generated using two kinds of C-spaces. In the final phase, every robot learns a behavior strategy by reinforcement learning and efficient formation of transportation is acquired. Simulation results indicate the effectiveness of proposed architecture.

Acceleration of reinforcement learning by a mobile robot using generalized rules

We propose an architecture to accelerate reinforcement learning by a mobile robot. In order to solve the problem of explosion of learning time in former reinforcement learning methods, we introduce a mechanism to acquire and utilize a set of widely applicable generalized rules into a reinforcement learning system. The mechanism extracts these rules by a statistical analysis of experienced data from the learning process. By applying these rules, the learning process can be accelerated by reducing search space. Simulation results indicate the effectiveness of the proposed method.

Flexible transport system by cooperation of conveyer-loaded AGVs

A transport module for constructing an efficient and flexible material handling system are proposed in the paper: an AGV (automated guided vehicle) module loaded a conveyer in which it is possible to hand-over operation between two modules. The operation can solve the deadlock problem among AGVs in the process of material handling. Prototype modules are created and hand-over experiments are made to measure motion characteristics of the modules. Transport simulations are made for a model plant layout by using values from former experiments. The results show the effectiveness of the proposed module from the viewpoint of transport efficiency, especially for cellular-type or process-type plant layouts.

State and action space construction using vision information

To apply reinforcement learning to the real world, it needs pre-processed sensor data which is adequate for action learning. Since it is difficult to construct state space and learn an appropriate action simultaneously, we assume that an estimation is given to each step of action, whether it is good or bad. Under this condition, we propose a method of dividing and clustering the state space. The TRN (topology representing network) is a vector quantization algorithm, and it can preserve topology in the input space. We apply the TRN algorithm to our problem with dynamically increasing nodes and the idea of a radial basis function.

In vitro reconstruction and functional development of the superior colliculus in the retinotectal pathway.

In order to examine the formation of a neural network and the functional development of a visual pathway, we performed in vitro reconstruction of the retinotectal pathway using organotypic explants and co-culture methods. Retinas and superior colliculus (SC) slices obtained from embryonic rats were co-cultured on microelectrode array (MEA) substrates for four weeks. We observed retinal ganglion cell neurites innervating SC slices that evoked responses in retinas or SC slices after applying electrical stimulation. Functional connections between retinas and SC slices were formed in the cultures. At the same time, spontaneous electrical activities were recorded from both the retinas and SC slices over the four weeks. In the co-cultured SC slices, sporadic firings were initially observed at 3-4 days in vitro (DIV), and thereafter the frequency of spontaneous firing increased and synchronized activities occurred after two weeks in vitro (WIV). In most of the single-cultured SC slices, however, only sporadic firings were observed over four weeks. In addition, the retinas and SC slices were co-cultured to enable the exchange of soluble factors with each other via culture medium but not via direct neural connections. The activity patterns resembled ones of single-cultured SC slices. These results suggest that signal inputs from retinas through direct neural connections affect the development of SCs in the retinotectal pathway.

Autonomous State-Space Construction in Pomdp with Continuous Observation Space

Abstract We propose a method to construct a state-representation that interprets continuous and non-Markovian perception into Markovian state. In our method, short-term memory is represented as a decision tree in order to distinguish states according to the short-term memory and continuous observation space represented in the tree incrementally segmented. The simulation result shows that the method constructs appropriate state-representation for a simple embodiment, environment, and task.