Artificial Life and Robotics最新文献

Early detection and accurate diagnosis of neurocognitive disorders (NCDs) are essential for enabling timely interventions to slow disease progression, preserve cognitive function, and enhance quality of life. This study introduces an EEG-based classification framework utilizing four AI-driven deep learning models—ResNet50-V2, EfficientNetB0, NasNetMobile, and MobileNetV2—to classify four neurocognitive groups: healthy controls (HC), mild cognitive impairment (MCI), Alzheimer’s disease (AD), and epilepsy (Ep) under eyes-closed (EC) conditions. To reduce complexity and improve accuracy, Fisher’s score was used to select 16 significant EEG channels from frontal, parietal, occipital, temporal, and central regions. Phase-amplitude coupling (PAC) images were generated from EC EEG signals to capture cross-frequency interactions, specifically how delta (0.5–4Hz) and theta (4–8Hz) phase modulate alpha (8–12Hz) and beta (12–35Hz) amplitudes—revealing functional brain connectivity. HC exhibited strong delta-alpha and reduced theta-beta coupling, while MCI and AD showed weakened delta-alpha and increased theta-beta interactions, reflecting cognitive decline. In contrast, Ep displayed elevated delta-alpha and reduced delta-beta coupling, linked to neural hyperexcitability. Among all models, MobileNetV2 achieved the best performance with 98.25% accuracy and 98.40% F score, attributed to its lightweight design and effective image feature extraction. This study’s novelty lies in its PAC image-based approach for NCD classification using MobileNetV2, providing valuable insights into non-linear hemispheric dynamics related to cognitive decline.

Robust monocular 3D object detection remains a pivotal challenge for intelligent robotic systems due to the absence of explicit depth information in single RGB images. In this paper, we propose a novel depth-guided attention enhancement (DGAE) module, integrated into the MonoDTR framework, to address its limitations in handling noisy depth supervision and refining spatial inconsistencies. DGAE leverages coarse depth maps as attention priors to guide visual feature refinement through temperature-scaled softmax and Gaussian smoothing, enabling enhanced spatial reasoning and robustness in cluttered scenes. To support this attention mechanism, we generate high-quality depth maps by projecting LiDAR points into the image plane and interpolating missing regions using a nearest-neighbor approach, followed by bilateral filtering and block downsampling to preserve edge details while reducing noise. This depth estimation pipeline improves the quality and coherence of the fused features used by DGAE. Extensive experiments on the KITTI 3D object detection benchmark show that our approach achieves state-of-the-art performance in moderate and hard detection scenarios for cars, pedestrians, and cyclists, while maintaining real-time inference speeds. These results underscore the effectiveness and practicality of our DGAE module for real-world 3D perception in autonomous driving applications.

Most humanoid robots use high-performance CPUs to process huge amounts of numerical calculations at high speed and control joint angles with servomotors. Humans, on the other hand, use neural networks to generate signals to contract and relax multiple muscles for efficient joint movement. We have created a musculoskeletal humanoid model that mimics the human musculoskeletal structure and have conducted dynamics analysis of walking. In the analysis, we obtained the time-specific generated force and contraction displacement of 12 different muscles during one walking cycle. In this study, based on dynamics analysis, artificial muscles were fabricated using elastomers and shape memory alloys for a total of 12 muscles: gluteus maximus, iliopsoas, rectus femoris, long head of biceps femoris, short head, vastus medialis and lateralis, gastrocnemius medialis and lateralis, tibialis anterior, and soleus muscles, and attached to the legs of a musculoskeletal humanoid. Using electrical signals from an external power source, the artificial muscles were driven based on the timing of muscle contraction during one cycle of walking, and the motion of the hip, knee, and ankle joints was reproduced in a stationary system. To validate the parameters of the artificial muscles obtained from the dynamics analysis of walking, the leg movements of the musculoskeletal humanoid were compared with the simulation results from the forward dynamics analysis.

In industrial environments, robots are confronted with constantly changing working conditions and manipulation tasks. Traditional approaches that require robots to learn skills from scratch for new tasks are often slow and heavily dependent on human intervention. To address inefficiency and enhance robots’ adaptability, we propose a general Intelligent Transfer System (ITS) that enables autonomous and rapid new skill learning in dynamic environments. ITS integrates Large Language Models (LLMs) with Transfer Reinforcement Learning (TRL), harnessing both the advanced comprehension and generative capabilities of LLMs and the pre-acquired skill knowledge of robotic systems. First, to enable robots to comprehend unseen task commands and learn skills autonomously, we propose a reward function generation method based on task-specific reward components. This approach improves time efficiency and accuracy while eliminating the need for manual design. Secondly, to accelerate the learning speed of new robotic skills, we propose an Intelligent Transfer Network (ITN) within the ITS. Unlike traditional methods that merely reuse or adapt existing skills, ITN intelligently integrates related skill features, enhancing learning efficiency through knowledge fusion. We evaluate our method in simulation, demonstrating that our method enables the system to learn skills autonomously without pre-programmed behaviors, achieving 72.22% and 65.17% faster learning speeds for two major tasks compared to learning from scratch. Supplementary materials are accessible via our project page:https://jkx-yy.github.io/

Ride comfort is an emerging focus in autonomous driving, yet integrating passenger behavior into vehicle motion planning remains challenging due to the impracticality of real-time passenger state feedback and high computational demands. This paper proposes a solution that approximates Ideal Motion Planning (IMP)—which traditionally requires nonlinear model predictive control (NMPC) and passenger state feedback—using linear model predictive control (LMPC) with weight tuning via Bayesian optimization. Our method eliminates the need for passenger state feedback and reduces the computation time, allowing real-time implementation. Although simplifying the cost function to prioritize tracking and stability, we achieve vehicle motion that maintains ride comfort equivalent to IMP, as demonstrated in simulations. This approach offers a practical pathway to improve passenger comfort in autonomous vehicles without additional sensory input.

Real-time posture estimation based on incomplete three-dimensional (3D) measurements is crucial for vision systems used in industrial robots. Conventional systems rely on manual pre-registering of multiple partial point cloud models for each workpiece and often fail when the 3D sensor is repositioned or its viewpoint changes. To overcome this bottleneck, we extend a fast MaskNet + singular‑value‑decomposition framework. However, its time‑series estimates still fluctuate owing to sensor and inference noise. To improve accuracy under realistic conditions, MaskNet was retrained on a large ray‑casting‑augmented CAD dataset that simulates random sensor viewpoints, and a Kalman filter was introduced to suppress temporal noise. The new training enhances mask‑vector accuracy, and the Kalman filter suppresses temporal fluctuations. Experiments confirm that the proposed method operates in real time on standard hardware, requires no pre‑registration after sensor movement, and can be seamlessly incorporated into a robot vision system for reliable target‑picking tasks.

Effective swarm management requires detecting individuals that negatively impact overall performance. This paper proposes a method using Explainable AI (XAI) to both detect swarm-level anomalies and identify the causative agents. We train a convolutional neural network (CNN)-based neural network to classify swarms as normal or abnormal, then use Grad-CAM (an XAI technique) to pinpoint the responsible individuals. We simulate swarms using the Boid model, introducing “anomaly agents” with slightly altered parameters for Alignment, Cohesion, and Separation. A novel model, adding a Lambda Layer to VGG16, is proposed and compared with standard CNNs (VGG16, ResNet50, DenseNet121, EfficientNetB0). The Lambda Layer model achieved the highest accuracy in both anomaly detection and agent identification. Experimental results show high accuracy in identifying agents with altered Alignment and Separation parameters. However, identifying agents with altered Cohesion is more challenging due to their proximity to normal agents, leading to increased misidentifications. The results demonstrate the effectiveness of combining CNNs and XAI for anomaly detection and root cause analysis in swarms.

Developing efficient transportation infrastructure networks capable of accommodating increases in population and demand is essential in urban planning. The conventional approaches to urban planning involve simulations using mathematical models that incorporate temporal changes. The current models are often based on static factors like existing land and road networks. However, land use and road networks need to be adapted to environmental and systemic changes to better capture urban dynamics. In this study, we aimed to address this by proposing a novel synergistic development model of population growth and infrastructure networks inspired by the adaptive network formation of slime mold Physarum polycephalum. The proposed model builds on the Physarum solver by incorporating two dynamic processes: adding new source points and deleting sink points with low flow. Adding source points simulates population growth and increases infrastructure demand, whereas deleting sink points enhances network efficiency by removing redundant paths. The numerical simulations were conducted under various conditions to evaluate the effect of these processes on network formation. The results indicate that deleting sink points accelerates the convergence of the network by eliminating unnecessary paths. However, an increased flow can result in higher energy loss if the number of paths is insufficient. These findings indicate that adaptive feedback mechanisms, inspired by biological systems, play a crucial role in optimizing infrastructure networks in response to population growth, offering insights for flexible urban development strategies.

With the proliferation of security cameras, image content protection is a major challenge. Image scrambling has increasingly been used for content protection as it does not degrade the quality of image. However, security cameras pose challenges of real-time implementation and target specific content protection. To this end, this paper presents a target specific, linear transform-based multi-image scrambling algorithm. The algorithm can scramble the image in 2D and 3D. Scrambling in 3D enables inter-image pixel scrambling which prevents brute-force attacks. The algorithm can be implemented using MMA (matrix–matrix multiply add) operation for parallel computing. A faster algorithm is proposed for serial computation. Both square and rectangular images can be scrambled. Along with complete image, targeted areas of the image can be scrambled in real time. The quality of scrambling is evaluated using PSNR (peak-signal-to-noise ratio) parameter. Experiment results with actual security cameras with motion detection feature shows that the proposed algorithm can be used in real time with high pixel irregularity for content protection.

Large Language Models (LLMs) are a type of machine learning model trained on vast amounts of natural language that have demonstrated novel capabilities in tasks such as text prediction and generation. These tasks allow LLMs to be remarkably suited for understanding the semantics of natural language, which in turn enables applications such as planning real world tasks, writing code for computers, and translating between human languages. Even though LLMs could provide more flexibility in interpreting user requests and have shown to possess some commonsense knowledge, their capabilities for translating natural language instructions into code to control robot actions is only starting to be explored. More specifically, in this paper we are interested in the control of robots tasked with preparing cocktails. Within this context, it is assumed that the LLM has access to a repository of well-formatted recipes. This means that each recipe is written according to the following layout: a list of ingredients, then a subsequent description of how to prepare and mix the various items. Moreover, a set of low-level modules responsible for robot manipulation and vision-related tasks is also provided to the LLM in the shape of an application programming interface (API). Consequently, the main focus of the LLM is on generating a sequence of calls to the API, along with the right parameters, to produce the cocktail requested by users in natural language. Here, we show that it is feasible for LLMs to perform this type of translation on a small number of custom modules, and that certain techniques provide a measurable benefit to the accuracy and consistency of this task without fine-tuning. We found in particular that the use of an ensemble-voting strategy, where multiple trials are repeated and the most common answer is selected, increases accuracy to a certain extent. In addition, there is moderate support for the use of natural language parsing to adjust the prompt of the LLM prior to translation. Lastly, building on previous knowledge we also provide a set of guidelines to help design prompts to improve the accuracy of the resulting sequence of actions. In general, these results suggest that while LLMs can be used as translators of robot instructions, they are best applied in conjunction with these other strategies. The impact of these findings could influence future robotics development, as it provides directions for implementing LLMs more effectively and broadening the accessibility of robotic control to users without an extensive software background.