Artificial Life and Robotics最新文献

A multi-agent system is a system in which multiple agents cooperate to achieve a global objective. Each agent within the system can move independently and exchange information with other agents to collaborate in executing tasks. As an entertainment application of multi-agent systems, previous research has been conducted to realize a mass game in which a group of small robots displays static images. Additionally, with the recent advancements in text-to-image generative models (e.g., stable diffusion), image generative AI has rapidly gained popularity across various fields. In this paper, we propose ChatRAD, a system that integrates image generative AI with a multi-agent system, allowing a group of robots to display animations based on text input from users. We describe the detailed structure and functionality of the system, as well as the control laws governing the robot swarm. Numerical experiments are conducted using multiple target animations to evaluate the system’s performance. Furthermore, to enhance the quality of the displayed animations, we apply Bayesian optimization for gain tuning and verify its effectiveness using image evaluation indices.

This paper proposes a mobile underwater robotic manipulator equipped with a suction mechanism consisting of a marine thruster and a disk plate to ensure its position-keeping performance. The marine thruster induces water flow between the disk plate and a structure’s surface. The water flow produces negative pressure to maintain stable contact between the manipulator and the structure’s surface, thereby ensuring reliable manipulation. Even when hydrodynamic forces act on the manipulator, the suction force keeps it stable on the structure’s surface. For this work, a prototype underwater manipulator with a suction mechanism was designed and developed. Based on measurements of the force and moment generated by the suction mechanism, the stabilization performance of the manipulator was evaluated numerically when a disturbance water flow is present or when the manipulator moves dynamically. Results confirmed that the prototype manipulator can maintain its position and orientation even in a disturbance water flow and during dynamic motions of the manipulator.

The objective of a reinforcement learning agent is to discover the best actions through exploration. However, standard methods such as (epsilon)-greedy and intrinsic motivation may rely excessively on random selection or the pursuit of novelty. As a result, they can drive exploration toward states unrelated to the actual learning objective, potentially leading to inefficient learning and delayed convergence. We propose a novel deep reinforcement learning method that prioritizes achieving a target return over maximizing return. This method directly utilizes the target return as a guiding signal for exploration, prioritizing actions that are more likely to achieve the target return. In addition, when the target has not yet been achieved, our agent actively explores the less frequently selected actions to discover better ones. Through experiments on a motion control task and a navigation task, our method demonstrated more stable and robust performance than standard methods, achieving higher returns with fewer episodes. These findings suggest that exploration based on target return can be an effective approach for practical applications whose priority is achieving a specific performance level under limited resources.

This study presents a novel approach for objective chronic stress assessment using multi-wavelength facial imagery and independent component analysis (ICA). We investigated the relationship between facial images captured across distinct spectral bands (infrared, near-infrared (780–900 nm, 900–1700 nm), and visible (L*, a*, b*)) and participants’ chronic stress scores. Applying ICA, we identified statistically independent facial features and found significant correlations between their component weights and chronic stress levels. Notably, visible and infrared wavelengths exhibited robust and more generalized correlations across participants compared to near-infrared. Our analysis revealed stress-correlated features predominantly in the perioral and periorbital regions, intriguing given their association with acute stress responses, yet suggesting distinct underlying physiological adaptations for chronic stress. This ICA-based method offers an interpretable and data-driven framework for uncovering subtle, persistent manifestations of chronic stress in the face, paving the way for scalable non-contact health monitoring solutions.

Underwater robotic systems face significant challenges in object detection due to the complexity of marine environments. To address these challenges, our previous work introduced EAW-YOLO11, an enhanced object detection network based on the YOLO11 architecture and specifically designed for underwater scenarios. In this model, we propose two novel modules: the EC3k2 module with Efficient Multi-scale Attention (EMA) for improved feature extraction and the C2AIFI module for effective feature integration. In addition, we adopt the Wise-IoU v3 loss function to enhance localization performance. In this extended study, we further refine EAW-YOLO11 to address the overfitting issues observed in the initial version, specifically adjusting the momentum parameter during training. Experimental results on the URPC2019 dataset show that EAW-YOLO11 achieves a 2.1% increase in mAP@0.5 compared to the baseline YOLO11, demonstrating strong performance even in blurred and low-visibility conditions. Further ablation studies and qualitative evaluations confirm that EAW-YOLO11 is a promising solution for real-world underwater robotic applications, including marine exploration and autonomous navigation. The code will be released at https://github.com/successdang99/EAW-YOLO11.

When natural disasters occur, accurately assessing the situation in a changing environment requires rapid and persistent observation. This paper focuses on the problem of conducting persistent observations using heterogeneous multi-agents with different observation capabilities and maximum operational time. To enable long-term operations covering vast environments, agents must replenish energy resources such as batteries and fuel. In this research, we discuss a method for selecting the direction of movement for observation by calculating utility values based on mutual information and remaining energy. Since mutual information reflects the different observation capabilities of each agent type, agents can move to locations where they can obtain the most information through their own observations. Through utility value-based action determination, agents fundamentally decide their movement directions based on mutual information, but as their energy levels decrease, they become more likely to choose the shortest path to a supply station. To validate the proposed method, we conducted numerical experiments involving persistent observation simulations in two changing environments: a simplified artificial environment and an environment representing a large-scale disaster in progress. The experimental results confirm that our proposed method successfully enables heterogeneous multi-agents to perform reactive observations while resupplying energy.

This study proposes a novel method for controlling anisotropic texture in 3D food printing using a single type of food ink. By adjusting the layer pitch and internal structure during printing, localized variations in hardness were achieved. Specifically, by introducing a “variable pitch technique”, rupture tests revealed that the hardness in the Z-direction increased as the pitch decreased, with values of 3.34 ± 0.29 N at 0.3 mm, 3.00 ± 0.33 N at 0.6 mm, and 1.89 ± 0.37 N at 0.9 mm. Additionally, wheel-shaped samples with different internal structures exhibited distinct fracture behaviors, confirming that internal design is effective for texture modulation. This technique enables personalized control of food texture based on individual chewing abilities, contributing to the realization of customized food experiences for medical and other special dietary needs.

This research proposed a novel method for tracking and predicting the growth stages of two-spotted crickets, reared in a temperature-controlled box at different growth stages using the YOLOv5s model. The images of crickets feeding inside the rearing box were taken with an infrared camera above the feeding point every hour. Images of the cricket were used to train a YOLOv5s model to detect crickets for each growth stage in the rearing box. The experimental results showed that the trained deep learning had an average accuracy of 95.7%. The relationship between the ratio of crickets at each growth stage throughout the 45-day rearing period was plotted and discussed. The results also showed a clear relationship between the amount of food consumed by crickets per day and their growth stage, which could be useful for appropriately managing food consumption according to the growth stage of crickets.

In this study, we propose a noninvasive blood pressure estimation method using spatial remote photoplethysmography (rPPG). Multiple regions of interest were set to cover the entire face in the captured RGB facial video, and rPPG signals were extracted from each region. The pulse wave contour, pulse beat, and derivative features were obtained from the extracted data and input into a neural network with an attention feature fusion module to extract essential features while reducing bias among them. The relationship between the features and blood pressure was modeled using this network, and the blood pressure was estimated. Finally, an experiment-wise 5-fold cross-validation was performed on a dataset of six subjects. When compared to the ground truth, the correlation coefficients for the systolic and diastolic blood pressure were 0.79 and 0.75, respectively, with mean absolute errors of 6.25 mmHg and 3.45 mmHg, outperforming the conventional method.

In recent years, the prevalence of hypertension has increased, making routine blood pressure monitoring crucial for early detection and prevention. Our research group focused on facial skin temperature, a cardiovascular indicator that can be remotely measured using infrared thermography. It has been studied for non-contact blood pressure estimation based on the spatial characteristics of captured thermal facial images (TFIs). Previous studies have estimated the resting blood pressure based on spatial features extracted by applying independent component analysis (ICA) to acquired TFIs. In practical applications, the use of low-resolution TFIs is considered cost-effective and advantageous for reducing the burden of data management owing to the smaller data volume. However, the reduced resolution could potentially result in the loss of critical information necessary for accurate blood pressure estimation. Therefore, it is essential to evaluate whether low-resolution TFIs can maintain sufficient estimation accuracy for practical implementation. In this study, we developed a resting blood pressure estimation model using low-resolution TFIs with a resolution of 160 (times) 120 pixels. We compared its estimation accuracy with that of a conventional model employing higher-resolution TFIs (320 (times) 256 pixels). Following the procedure used in the previous studies, ICA was applied to the TFIs, and a linear support vector regression (SVR) model was constructed using the weights of the selected independent components as input features achieving a root-mean-square error (RMSE) of 13.1 mmHg and a correlation coefficient (r) of 0.369.