Deep reinforcement learning (DRL), a vital branch of artificial intelligence, has shown great promise in mobile robot navigation within dynamic environments. However, existing studies mainly focus on simplified dynamic scenarios or the modeling of static environments, which results in trained models lacking sufficient generalization and adaptability when faced with real-world dynamic environments, particularly in handling complex task variations, dynamic obstacle interference, and multimodal data fusion. Addressing these gaps is essential for enhancing its real-time performance and versatility. Through a comparative analysis of classical DRL algorithms, this study highlights their advantages and limitations in handling real-time navigation tasks under dynamic environmental conditions. In particular, the paper systematically examines value-based, policy-based, and hybrid-based DRL methods, discussing their applicability to different navigation challenges. Additionally, by reviewing recent studies from 2021 to 2024, it identifies key trends in DRL-based navigation, revealing a strong focus on indoor environments while outdoor navigation and multi-robot collaboration remain underexplored. The analysis also highlights challenges in real-world deployment, particularly in sim-to-real transfer and sensor fusion. Based on these findings, this paper outlines future directions to enhance real-time adaptability, multimodal perception, and collaborative learning frameworks, providing theoretical and technical insights for advancing DRL in dynamic environments.

With the aging population rising, the decline in spatial cognitive ability has become a critical issue affecting the quality of life among the elderly. Electroencephalogram (EEG) signal analysis presents substantial potential in spatial cognitive assessments. However, conventional methods struggle to effectively classify spatial cognitive states, particularly in tasks requiring multi-class discrimination of pre- and post-training cognitive states. This study proposes a novel approach for EEG signal classification, utilizing Permutation Conditional Mutual Information (PCMI) for feature extraction and a Multi-Scale Squeezed Excitation Convolutional Neural Network (MSSECNN) model for classification. Specifically, the MSSECNN classifies spatial cognitive states into two classes—before and after cognitive training—based on EEG features. First, the PCMI extracts nonlinear spatial features, generating spatial feature matrices across different channels. SENet then adaptively weights these features, highlighting key channels. Finally, the MSCNN model captures local and global features using convolution kernels of varying sizes, enhancing classification accuracy and robustness. This study systematically validates the model using cognitive training data from a brain-controlled car and manually operated UAV tasks, with cognitive state assessments performed through spatial cognition games combined with EEG signals. The experimental findings demonstrate that the proposed model significantly outperforms traditional methods, offering superior classification accuracy, robustness, and feature extraction capabilities. The MSSECNN model’s advantages in spatial cognitive state classification provide valuable technical support for early identification and intervention in cognitive decline.

Episodic memory training plays a crucial role in cognitive enhancement, particularly in addressing age-related memory decline and cognitive disorders. Accurately assessing the effectiveness of such training requires reliable methods to capture changes in memory function. Electroencephalography (EEG) offers an objective way of evaluating neural activity before and after training. However, EEG classification in episodic memory assessment remains challenging due to the variability in brain responses, individual differences, and the complex temporal–spatial dynamics of neural signals. Traditional EEG classification methods, such as Support Vector Machines (SVMs) and Convolutional Neural Networks (CNNs), face limitations when applied to episodic memory training assessment, struggling to extract meaningful features and handle the inherent uncertainty in EEG signals. To address these issues, this paper introduces DU-former, which improves feature extraction and enhances the model’s robustness against noise. Specifically, data uncertainty (DU) explicitly handles data uncertainty by modeling input features as Gaussian distributions within the reparameterization module. One branch predicts the mean through convolution and normalization, while the other estimates the variance via average pooling and normalization. These values are then used for Gaussian reparameterization, enabling the model to learn more robust feature representations. This approach allows the model to remain stable when dealing with complex or noisy data. To validate the method, an episodic memory training experiment was designed with 17 participants who underwent 28 days of training. Behavioral data showed a significant reduction in task completion time. Object recognition accuracy also improved, as indicated by the higher proportion of correctly identified target items in the episodic memory testing game. Furthermore, EEG data collected before and after the training were used to evaluate the DU-former’s performance, demonstrating significant improvements in classification accuracy. This paper contributes by introducing uncertainty learning and proposing a more efficient and robust method for EEG signal classification, demonstrating superior performance in episodic memory assessment.