Autonomous navigation of Unmanned Aerial Vehicles (UAVs) based on deep reinforcement learning (DRL) has made great progress. However, most studies assume relatively simple task scenarios and do not consider the impact of complex task scenarios on UAV flight performance. This paper proposes a DRL-based autonomous navigation algorithm for UAVs, which enables autonomous path planning for UAVs in high-density and highly dynamic environments. This algorithm proposes a state space representation method that contains position information and angle information by analyzing the impact of UAV position changes and angle changes on navigation performance in complex environments. In addition, a dynamic reward function is constructed based on a non-sparse reward function to balance the agent’s conservative behavior and exploratory behavior during the model training process. The results of multiple comparative experiments show that the proposed algorithm not only has the best autonomous navigation performance but also has the optimal flight efficiency in complex environments.

Safe and efficient cooperative flight in cluttered, dynamic multi-target environments requires UAV swarms to repeatedly reassign targets while maintaining collision-free trajectories under partial observability. Existing methods often decouple target assignment and path planning into hand-engineered pipelines, which are effective in static, fully known settings but become brittle when targets move and perception is uncertain, leading to reduced cooperation efficiency and poor adaptability. This paper proposes DA-MAPPO, a dynamic assignment-aware multi-agent reinforcement learning framework built on PPO. DA-MAPPO integrates an online minimum-cost target allocation module into a unified perception–allocation–decision-making loop: real-time assignment outcomes are embedded into each agent’s local observation to form assignment-augmented states, enabling step-by-step coordination reconfiguration. In addition, a hierarchical cooperative reward is designed to jointly encourage individual safety (collision avoidance) and team-level mission efficiency. High-fidelity simulation results in Gazebo show that DA-MAPPO achieves 90%–99% mission success across dynamic multi-target scenarios with varying obstacle densities, outperforming representative baselines by up to 25 percentage points. When transferring from static to dynamic targets, DA-MAPPO exhibits negligible degradation in low- and medium-density environments and only about a 2% drop in the densest setting, while baselines degrade substantially. DA-MAPPO also produces shorter average trajectories and fewer decision steps, indicating higher mission efficiency with lower collision risk.

Range–Doppler (RD) spectrogram-based maritime target recognition requires effective feature learning when labeled training samples are limited and the range and Doppler axes convey different physical information. We propose DAMD-Net, which combines Asymmetric Dimension-Aware Fusion (ADF) with a Dual-Directional Attention Module (DDAM). ADF integrates range-wise scattering features, Doppler-wise motion features, and local two-dimensional context, while DDAM recalibrates the fused representation through axis-specific descriptors and learnable fusion weights. With 5% of the simulated dataset used for training, DAMD-Net achieves $86.15\pm 0.43$% accuracy and $83.57\pm 0.41$% macro-F1. With 10% used for training, the corresponding results are $89.55\pm 0.39$% accuracy and $87.48\pm 0.51$% macro-F1. At the smaller 1% and 2% training budgets, DAMD-Net improves macro-F1 over VGG16 by 1.64 and 1.82 percentage points, respectively, with positive unadjusted 95% paired confidence intervals. Each training budget uses an additional 4049 labeled validation samples. Component ablations, kernel and pooling sensitivity studies, and condition-holdout evaluations examine the contribution of directional modeling; DAMD-Net has higher mean accuracy than VGG16 in all three elevation-holdout intervals. Controlled perturbation tests show advantages under partial Doppler masking and structured synthetic clutter: at a baseline-RD-to-added-clutter ratio of $-10$ dB, accuracy reaches 62.74%, compared with 36.68% for VGG16. On the external MTDSP dataset, DAMD-Net achieves 97.13% accuracy; on the measured LSS-DAUR dataset, it achieves $75.00\pm 14.03$% accuracy under five-fold cross-validation. DAMD-Net also reduces parameters by 89.9% and registered FLOPs by 52.6% relative to VGG16. Together, these findings support explicit RD anisotropy modeling for recognition with restricted training data, with complementary benefits in the evaluated disturbance conditions and model storage.

Unsupervised temporal action segmentation is a critical yet challenging task in high-level video understanding, which aims to partition continuous, untrimmed videos into distinct action segments without relying on manual annotations during the training phase. Existing methods typically exploit pre-extracted features for clustering or boundary detection but often suffer from limited accuracy and scalability. In this paper, we reframe unsupervised temporal action segmentation as a video alignment learning task and propose a novel framework with a Frame-Segment Contrastive (FSC) loss. The FSC loss simultaneously encourages fine-grained frame-level alignment and global segment-level coherence, effectively capturing temporal structures in a fully label-free manner. To map the learned representations to semantic categories, we introduce an inference-time exemplar-based label transfer strategy. This approach utilizes the alignment path from a single labeled reference video to generate predictions for target videos. Extensive experiments conducted on five benchmark datasets—Pouring, Penn Action, Breakfast, IKEA Assembly, and Desktop Assembly—demonstrate that our method outperforms state-of-the-art approaches, validating its robustness and scalability in modeling complex temporal dependencies.