Stereo Visual Odometry (VO) plays a critical role in enabling precise localization in GPS-denied environments, such as indoor, urban, and subterranean settings. Despite its importance, one of the major challenges of stereo VO is the high computational cost involved. To address this, this study focuses on reducing the computational load by increasing the frame interval at which VO is performed. However, expanding the frame interval may result in accuracy degradation, making it necessary to adjust the interval to prevent significant loss of accuracy. Understanding the factors that influence VO accuracy is essential for this adjustment. Causal analysis is employed to identify these factors, and a method is proposed to determine the optimal frame interval based on these findings. Experimental results demonstrate that the proposed approach successfully extends the frame interval while maintaining accuracy, achieving a balance between computational efficiency and reliable localization.

This study presents the design and evaluation of a robotic system for long-term surveys on remote islands. Nishinoshima Island, the focus of this research, is undergoing rapid geological changes due to active volcanic eruptions, making conventional manned surveys difficult. To address these challenges, we developed a robotic system equipped with a wireless power transfer mechanism, enabling long-term operations with limited energy resources. The rover utilizes solar power and works in conjunction with a charging station, ensuring essential functions remain operational during nighttime or inclement weather. During a 2024 survey on Nishinoshima, the rover successfully conducted night-time observations of bird nesting sites, demonstrating the effectiveness of the power management system, though challenges such as vehicle overturns and transportation limitations were identified. This study highlights the potential of robotic operations in isolated natural environments and provides insights for improving sustainable survey systems in the future.

Self-organizing robots has been proposed that connects multiple small robots and behaves as one large robot. Conventional self-organizing robots perform the robot coupling operation from a predetermined position and direction. However, since precise position control is required, it takes a considerable amount of time to form an aggregate. Therefore, we have developed self-organizing robots with ``omni-directional connector'' that can be connected from all directions. In this paper, we will implement a coupling operation that does not clearly define the starting position of the connection and does not require precise position control by simulation using multi-agent deep reinforcement learning, and conduct experiments on actual machines.

Focusing on obstacle avoidance, a mobile robot is typically required to move straight for a certain duration after avoiding an obstacle before reorienting itself toward the destination. In this paper, we propose a novel motion planner based on a Time-Delay CNN that uses time-series image inputs. Through experiments, we demonstrate that the robot is able to plan appropriate avoidance motions as described above and navigate toward the destination in both simulation and real-world environments with multiple dynamic obstacles.

Silent speech recognition, which estimates spoken content from mouth movements without vocalization, has gained attention as an alternative to conventional voice-based systems. However, most existing studies have focused on English, and applications to Japanese remain limited. In particular, recognition at finer linguistic units such as morae has not been sufficiently explored. This study proposes a classification method for silent speech using IMUs, targeting Japanese morae. Experimental results show that Japanese seion can be accurately classified by leveraging phonological features. These findings confirm the feasibility of mora-level silent speech recognition in Japanese.

Autonomous navigation in crowded environments with frequent occlusions is challenging. This paper presents a mobile robot equipped with an embedded quantum-inspired optimization computer for robust multi-object tracking under occlusion and predictive path planning. An integrated simulation platform has been developed, combining an external environment simulator with a robot simulator that fully replicates internal module communications using the ROS 2 interface protocol. The functionality of the proposed approach is demonstrated by prototyping the mobile robot and the effectiveness is evaluated through simulation-based comparison study in dynamic scenes featuring multiple pedestrians.

This study proposes a method to predict wave-based acoustic behavior from geometrical acoustics simulations using a pix2pix image-to-image translation model. By training a conditional GAN with ray tracing outputs as inputs and FDTD simulation results as targets, the model learns to infer complex wave phenomena, such as interference and diffraction. Validation on training data demonstrates the model's ability to reconstruct wavefields from ray-based features. This approach aims to enable high-fidelity wide-area acoustic simulation with reduced computational cost, contributing to future applications in outdoor sound field prediction and acoustic signal processing.

This paper proposes a framework for high-precision, real-time alignment of point clouds from multiple RGB-D cameras in dynamic environments. First, our method performs a parallel-accelerated, robust pairwise Iterative Closest Point (ICP) using geometric and color information, followed by a global pose graph optimization (PGO). Next, an Extended Kalman Filter (EKF) is employed to compensate for computational latency, achieving a balance between accuracy and real-time performance. We demonstrate the effectiveness of the proposed method through real-world experiments with three robot arms and three RGB-D cameras.