With the deep expansion of informatization into multi-dimensional physical space, dominance over spatiotemporal information has become a core area of strategic competition among major countries around the world. In this paper, addressing the development needs of the national comprehensive positioning, navigation, and timing(PNT) system, the connotation and technological evolution of dominance competition systems are explored, the cross-domain collaborative development from command of the sea, air, space to electromagnetic and information dominance are reviewed, and the technological evolution path of navigation dominance in adversarial environments is specifically examined. By summarizing the strategies and development trends of the United States, Russia, and other countries in constructing technological systems for navigation countermeasures, it conducts a critical analysis of the vulnerabilities existing in current satellite navigation systems at both the service and application levels. Furthermore, it finds out the breakthrough technological directions such as space-based resilient PNT and intelligent multi-source autonomous navigation. Finally, the future trends of navigation dominance technology from dimensions including system confrontation and the cognitive domain is prospected, aiming to provide technological support for the development of new-generation comprehensive PNT system.

Autonomous navigation technology serves as an indispensable core capability for critical platforms such as unmanned systems, with its strategic value and application prospects becoming increasingly prominent. However, in GNSS-denied environments like urban canyons, indoor spaces, and underwater settings, traditional navigation methods reliant on GNSS are susceptible to interference, leading to severe accuracy degradation or positioning failure. Recent rapid advancements in deep learning technology have provided novel approaches for constructing high-precision, highly robust, and fully autonomous navigation systems. Focusing on deep learning-assisted autonomous navigation technology under GNSS-denied conditions, an in-depth review and analysis of research progress in three key areas is conducted: deep learning-assisted inertial navigation technology, multi-source intelligent navigation technology under GNSS-denied environments, and deep learning-enhanced filtering and fusion techniques. Finally, the future development trends of deep learning-assisted autonomous navigation technology are outlined.

Autonomous navigation is a core capability for evaluating the robot’s level of intelligence. Traditional navigation frameworks heavily rely on continuous and precise positioning information, which often leads to system collapse in perception-degraded environments such as long corridors due to localization failure. Meanwhile, a single planning strategy is insufficient to balance efficiency and safety across diverse environments. To address these challenges, an adaptive navigation framework based on point cloud scene understanding and topological planning is proposed. A navigation strategy switching method based on SPVCNN scene understanding is developed, which effectively recognizes spatial structures such as open areas, narrow corridors, and rooms, designing an adaptive switching approach for scene-feature-oriented navigation strategies. An improved Zhang-Suen skeleton extraction method is introduced, combined with a skeleton-based pruning strategy to remove redundant nodes and branches, thereby enhancing the ability of the topological map to represent environmental spatial layouts. Furthermore, a heuristic A^* algorithm is designed, leveraging the extracted skeleton topology to generate path guidance aligned with corridor structures, improving the robot’s stability and safety margin in confined spaces. Experimental results show that, in narrow environments, the proposed method reduces navigation time by an average of 13.1% and improves average path smoothness by 34.6% compared to mainstream local path planning methods, while maintaining stable and safe operation even under localization failure.

The behavioral learning of embodied intelligence relies on high-quality robot manipulation data. However, real-world robotic data collection is costly and limited in scale, while internet-scale video data, though abundant, lacks action and state annotations. To address the challenge of extracting state representations from unlabeled videos, a video pre-training-based behavioral learning method for embodied intelligence is proposed. Firstly, an unsupervised video pre-training framework is constructed to achieve latent state extraction through feature extraction encoding, static-dynamic feature separation, and cross-frame consistency constraints. Secondly, a multimodal Transformer architecture is designed, integrating patch-wise attention mechanisms with dynamic action heads to accomplish multimodal information fusion and adaptive action generation. Simulation results demonstrate that the proposed method achieves up to 32.96% performance improvement over the baseline method Moto in task execution on CALVIN and SIMPLER simulation environments. It also exhibits significant advantages in both unknown environment generalization and environmental robustness testing, effectively enhancing the behavioral learning capabilities of embodied intelligence.

GNSS spoofing attacks can enable attackers to covertly control autonomous vehicles, posing a serious threat to road traffic safety. Addressing limitations in existing detection methods, such as restricted application scenarios and insufficient real-time performance, a spoofing detection method based on LSTM and attention mechanisms is proposed. By constructing an LSTM-Attention model, the approach achieves multi-sensor data fusion and vehicle motion state estimation, thereby performing dead reckoning to generate a reference trajectory and employing trajectory consistency verification to detect spoofing. To mitigate the impact of cumulative errors on positioning accuracy, a sliding time window mechanism is introduced to correct errors within the window while comparing reference trajectory with GNSS navigation solutions. Experimental results demonstrate that the proposed method achieves 97.5% detection rate for abrupt spoofing attacks while maintaining low false alarm rates, outperforming existing baseline methods and meeting real-time detection requirements.

Achieving efficient obstacle avoidance and stable navigation in complex dynamic environments remains a critical challenge for the development of service robots. Traditional methods often rely on static maps or frequent iterative computations of locally optimal trajectories, which are prone to local optima or collisions in dynamic and narrow corridors. To address these limitations, an end-to-end navigation framework integrating spatio-temporal perception with knowledge distillation is proposed. Specifically, a PredRNN-based structure is employed to model image sequences and capture spatio-temporal features. It further incorporates a Teacher-Student architecture. The Teacher network, which utilizes LiDAR and prior knowledge of dynamic obstacles, generates high-quality policies to distill both perception and behavior into the Student network, which takes depth images as input. This enables the Student network to inherit Teacher knowledge while achieving robust decision-making. Experimental results demonstrate that the proposed method improves the success rate of reaching target destinations by 13% compared to the best-performing baseline in complex dynamic scenarios. Overall, this framework overcomes the limitations of traditional approaches in dynamic, narrow, and perception-constrained environments, exhibiting stronger generalization and adaptability, and offering a novel perspective for ensuring safe and efficient operation of service robots in real-world complex environments.

Inertial measurement units play a crucial role in autonomous navigation and positioning, but their measurement errors increase exponentially over time. A pedestrian inertial navigation method based on time-frequency feature encoding neural networks is proposed. The time series sequence of inertial data and the frequency domain sequence obtained through Haar transformation are respectively used as imputs to the neural network. The time-domain and frequency-domain features are extracted separately by the inertial time-frequency feature encoder, and the dependencies between different time steps and frequency components are adaptively fused and learned through the multi-head attention mechanism. Then, the prediction results of the neural network are integrated with the inertial motion model through the extended Kalman filter framework to further optimize the state estimation. Experimental results on the public datasets TLIO and RoNIN show that, compared with the benchmark method TLIO, the proposed method reduces the ATE, RTE, and DR by 10.8%, 17.7%, and 12.9% respectively, demonstrating high accuracy and robustness in complex pedestrian motion scenarios.

With the continuous expansion of international trade and maritime transportation, ship trajectory prediction based on the AIS faces increasingly stringent demands for accuracy and robustness as a core technology for smart shipping and maritime supervision. To address common observational disturbances and insufficient predictive performance in complex navigation environments, a ship trajectory prediction framework combining high-precision forecasting capabilities with strong noise robustness is developed. Specifically, a novel prediction model, GRU-MSAformer, is proposed, integrating GRU and multi-scale causal self-attention mechanisms. The model first captures local temporal dependency features via GRU, then employs a multi-scale self-attention mechanism to model trajectory behavior across different time scales, thereby enabling adaptive noise filtering. Experimental results demonstrate that GRU-MSAformer achieves superior performance under both noise-free and Gaussian noise conditions. It maintains low prediction errors across 10 to 40 minutes forecasting tasks while sustaining stable prediction accuracy under varying noise intensities.

With the widespread application of nuclear technology, the use of mobile robots to replace human operators in executing nuclear emergency tasks within unknown radiation environments has become increasingly important. However, due to limitations in detection time and sensor performance, robots can only obtain sparse radiation data. Nonetheless, in order to facilitate nuclear safety monitoring, it is essential to obtain the radiation fields distribution and the locations of radioactive sources in the environment. To address the above problems, a source localization method integrating two-dimensional laser SLAM and radiation features is proposed. This method uses mobile robots equipped with nuclear radiation detectors, LiDAR, and other sensors to collect radiation data and construct environmental maps. Subsequently, it uses Gaussian process regression method to invert the regional radiation field and integrates the inverted radiation field into the SLAM environmental map. Finally, the Hough transform method is applied to locate unknown radioactive sources. In addition, experimental verification is conducted in real environments where radioactive sources are present. The experimental results show that based on occupancy grid maps constructed using three laser SLAM algorithms (Gmapping, Hector, and Cartographer), the fusion of global radiation environment maps can be completed in both open space and factory environments, with localization accuracy exceeding 0.29 m.

Axis misalignment, scale factor deviations, and time-varying noise present in low-cost IMUs significantly degrade attitude estimation accuracy. Existing neural network-based denoising methods exhibit clear limitations in multidimensional error modeling. To address this, a gyroscope adaptive calibration network integrating temporal and channel attention mechanisms is proposed: a convolutional neural network performs feature extraction, channel attention optimizes multi-axis feature weighting, and the temporal attention balances feature contributions over time, thereby enhancing network accuracy and robustness. Experimental results on the EuRoC dataset demonstrate that channel attention substantially improves dynamic compensation accuracy, while temporal attention balances accuracy across the three axes without overall performance gains, combining both mechanisms yields certain improvements, their interaction may hinder optimal performance in certain scenarios. These results validate the effectiveness of multi-attention mechanisms in inertial sensor errors modeling and provide new insights for designing low-cost gyroscope dynamic compensation algorithms.