General classification methods only involve consideration of learned classes, and do not cover undefined targets such as unintended characteristics in the learning process. This paper proposes a novel probabilistic neural network can treat unlearned class. The proposed method incorporates two types of probabilistic distribution: normal and complementary Gaussian distribution and can reach multi-class classification and unlearned class detection with a single network. In the experiments, artificial data and electromyogram (EMG) patterns were classified to demonstrate the capabilities of the proposed method. The results showed that the approach produces high performance for classification, and there were significant differences between the proposed and previous methods.

In order to achieve an autonomous vessel, it is important how to organize the map. The map is required to be fine when navigating in confined space but to be coarse when navigating in sparse space to save memory consumption and computation time. However switching between fine and coarse maps may cause discontinuities of navigation. In this paper, we propose a concept that integrates coarse and fine maps without switching between them named SEAMLESS. To resolve the ambiguity near the map boundary and optimize the number of map grids, our approach integrates the wide area maps and the narrow area maps with depth-optimized Quad Trees according to each sensor's distance error and self-positioning error. The concept has been demonstrated in real environment on our vessel with multi-modal system which generates both wide and narrow area maps based on obstacle detection.

It has been proposed to apply adaptive function which is one of the characteristics of passive dynamic walking in designing legged walking robot hardware. Although the adaptive function has been confirmed to appear only in computer simulations, it has not yet been confirmed in actual experiments. We tried to conduct the phenomenon caused by Adaptive function, by several verification experiments, and then confirmed that Adaptive function appeared. Addition to it, we confirmed that the walking motion of passive dynamic walking can be stabilized by applying adaptive function by several experiments. In this paper, we describe the verification of adaptive function and the stabilization of walking motion by experiments.

We are developing a system to treat kidney, liver and other organs, as well as stones and cancer in these organs using high-power focused ultrasound (HIFU) while tracking lesions that move by breathing and body movements. The system estimates organ movement by analyzing ultrasound images obtained from the probe and compensates for this movement with robotic control. In recent years, various medical image analysis methods using deep learning have been proposed and studied, but they have not been fully explored as methods to detect specific organs using robotic systems. In this paper, we compared and validated the performance of Faster R-CNN, which is commonly used for object detection, and the proposed methods, Regression Network (RegNet) and Segmentation In Regression Network (SegInRegNet), the proposed method based on the problems of Faster R-CNN, in a kidney detection task. And then, we show that 1) there are some problems with Faster R-CNN as a kidney detection method operating on a robotic system and that 2) proposed method performs better than Faster R-CNN in terms of detection speed and accuracy.

This study proposes a new four-fingered robot hand mimicking tree frog's sucker-attached fingers. We introduced tree frog's bone structure, movement of its fingers (addiction/abduction, flexion/extension), and its fingertip's suction function. In addition, considering the feasibility of mimicking tree frog's fingers from engineering perspective, we improved the prototype by: i) simplifying its mechanism and control by settling the thumb opposable to the middle finger, ii) simplifying control by 2 degrees of freedom handling, one motor for flexing/extending the four fingers at once by moving the wires, and one linear actuator and linking mechanism for addicting/abducting the pointed finger and the ring finger, iii) adding 4th passive joint with suction cups and a claw on the fingertip to increase contact force from the robot hand onto the object. The experiment showed the possibility that this robot hand may grasp several types of objects, whether the object is planar or 3D shape.