To addresses the attitude tracking control problem of reusable launch vehicle under the influence of model uncertainties， an adaptive fixed-time reinforcement learning-based optimal backstepping control strategy is proposed in this paper. Firstly， a fixed-time controller based on the hyperbolic tangent function is designed to avoid the singularity issue in traditional fractional-order fixed-time controllers. Secondly， regarding handling system model uncertainties， a radial basis function neural network is employed for approximation， and a series-parallel estimation model is introduced to enhance estimation accuracy and convergence speed. On this basis， an "Identifier-Actor-Critic" framework is established， and an improved weight adaptive update law is designed to solve the Hamilton-Jacobi-Bellman equation， and a trade-off is achieved between tracking performance and control input. Finally， the fixed-time stability of this proposed method based on Lyapunov theory is proven， and numerical simulations are conducted to verify the effectiveness of the algorithm.
Regarding a guidance scenario of the multiple projectiles that simultaneously attack a stationary target at a preset time in three-dimensional space， a cooperative guidance law satisfying fixed time convergence is proposed. Firstly， a three-dimensional guidance constant velocity model for missiles and guidance error variables are established. Next， an input variable consisting of a fixed time consistency controller and an error convergence controller is designed to establish a first-order multi-agent system that satisfies time constraints. Then， the collaborative guidance law of proportional guidance term and time synergy term is derived from the system state equation. Finally， the effectiveness of the proposed derivative law is verified through numerical simulation.
Regarding the strong noise interference easily caused by high-frequency vibrations and complex illumination during the flight visual telemetry of launch vehicle and ground object recognition， and the deployment difficulties of high-precision models due to the constrains of launch vehicle-borne embedded platform power consumption and computing resource， a layer-wise sensitivity-aware adaptive interpolation quantization method is proposed. Firstly， a data-driven metric is established to identify accuracy-bottleneck and non-sensitive layers by analyzing feature drift under quantization noise in each layer. Secondly， a lightweight Alpha controller is jointly trained with the backbone network， by which dynamic interpolation coefficients for sensitive layers can be generated in real-time by characterics of input image textures. Finally， a weighted interpolation mechanism of static quantization parameters and dynamic statistics values is established， where static priors are utilized online to correct dynamic parameters in order to suppress outlier fluctuations. Simulation results of proposed method on YOLOv8s demonstrate that an effective balance between inference speed and detection accuracy is achieved， and by keeping low computational overhead， in the meantime the model's robustness in complex aerospace environments is significantly improved， which is applicable to real-time visual monitoring and target recognition tasks on resource-constrained launch vehicle-borne platforms.
To address the issues of local minima and parameter sensitivity in the traditional artificial potential field （APF）， as well as the slow convergence and low sample efficiency of deep reinforcement learning in UAV path planning in complex unknown environments， an adaptive path planning algorithm fusing proximal policy optimization （PPO） and improved artificial potential field （PPO-APF） is proposed. A hierarchical architecture is applied in this algorithm. The repulsive force coefficient， effective distance， and rate adjustment coefficient are adaptively outputted by the upper-layer PPO network based on local perception， and flight commands are generated by the lower-layer improved APF model， where a target distance factor and kinematic constraints are introduced. Meantime， a composite reward function incorporating path smoothness with parameter stability is designed to suppress path jitter. Simulation results demonstrate that in complex scenarios with 22% obstacle coverage， the planning success rate of PPO-APF reaches 97.33%， which is significantly superior to the improved APF by 53.67% and pure reinforcement learning based PPO by 85%. The proposed algorithm takes advantage of a significant "cold start" mode， which effectively combines the adaptability of DRL with the stability of APF to realize UAV's autonomous learning of slow-down crossing on narrow passage and preemptive avoidance strategies， etc.
Regarding the performance degradation and computational redundancy in existing multi-modal object tracking tasks caused by incomplete multi-modal information fusion methods and inefficient interaction processes among multi-source modalities， the star adapter for multi-modal tracking （SAMT） is proposed. A multi-level interaction architecture is employed in this method and the more efficient star-operation is introduced to achieve full compact fusion of multi-modal information in the feature space. Meanwhile， by investigating the similarity of multi-source modalities' features across different network layers， an efficient adapter architecture configuration is designed. The experimental results show that， compared with existing methods， the proposed approach achieves certain extent improvements in both precision and success rate on the RGBT234 and LasHeR datasets. Moreover， it reduces computational cost to 24.87% and model parameters to 48.74% of their original values.
To address the issue of limitations of navigation efficiency and susceptibility to local minima in unknown dynamic 3D environments for unmanned aerial vehicles， an improved DWA-APF path planning algorithm is proposed， which integrates temporal dynamic potential field prediction and an adaptive evaluation mechanism. Firstly， a temporal dynamic potential field model based on velocity-acceleration estimation is established to proactively avoid obstacles by predicting their spatiotemporal trajectories. Secondly， an adaptive weight adjustment strategy based on local environment density is designed to dynamically balance the planner's efficiency and obstacle avoidance safety. Thirdly， an active trap avoidance mechanism based on cost space topology analysis is proposed to effectively overcome local minima problems such as U-shaped obstacles. Finally， second-order difference constraints on the trajectory are incorporated into the sampling evaluation system to ensure the dynamic feasibility of the generated trajectory. Simulation results show that， compared with traditional methods， this flight time， path length and obstacle avoidance stability are significantly improved in complex dynamic scenarios by using the proposed algorithm that demonstrates excellent real-time performance and robustness.
To address the issue of severe fluctuations and singular peaks in two-dimensional resolution during the terminal guidance of spaceborne-airborne bistatic synthetic aperture radar， a two-dimensional resolution optimization method that involves both configuration optimization and terminal guidance feasibility is proposed. In the configuration planning stage， the non-dominated sorting genetic algorithm Ⅱ is used to implement multi-objective optimization， and the average level and uniformity of the scene's two-dimensional resolution are token as optimization index， and a forward-shift constraint based on the statistical quantile of the resolution angle is introduced to suppress high-risk configurations. In the trajectory generation stage， an online fine-tuning correction strategy is designed， which is based on the proportional navigation guidance model and enables real-time suppression of resolution degradation trends during flight. The results of ablation comparison simulations show that the proposed forward-shift constraint and online correction can decrease resolution peaks and fluctuation indexes， and thereby further improve the stability of two-dimensional resolution. The optimization approach can serve as an engineering-feasible reference for configuration planning and online guidance in spaceborne-air-borne bistatic SAR systems.
A robust adaptive collaborative positioning algorithm is proposed to address the issue of ultra-wideband （UWB） ranging that is susceptible to interference from non-line-of-sight and multi-path effects on unmanned vehicle cluster collaborative positioning within global navigation satellite system-denied environments. Firstly， a robust random cut forest method is applied to perform anomaly detection by using the ranging innovation and the UWB first-path power ratio as input features， and a Huber weighting matrix based on UWB signal strength is established to correct abnormal ranging measurements. In addition， a variational Bayesian adaptive mechanism is introduced to estimate the measurement noise covariance online， which results in the formation of proposed RRCF-Huber-VBSRCKF algorithm. Experimental results show that， compared with the standard SRCKF， the average localization error is reduced from 0.23 m to 0.14 m and improved by 39.1%； compared with the RRCF-Huber-SRCKF， the average localization error is reduced from 0.18 m to 0.14 m and improved by 22.2%， and the maximum error is reduced from 0.5 m to 0.39 m， which demonstrate the superiority of the proposed algorithm.
Regarding the insufficient control robustness of deep reinforcement learning methods in complex dynamic environments caused by external noise interference， sensor failures and other factors， a kind of robust deep reinforcement learning control algorithm framework based on integration of spiking neural networks is proposed. A spiking actor network consists of kernel on spatio-temporal dynamic neurons is established， and the proposed robust deep reinforcement learning framework is designed by facilitating non-stationary signal filtering with a TD3-based optimization scheme. By leveraging the spatio-temporal accumulation and threshold-triggering characteristics of spiking neurons and integrating with multiple neural coding mechanisms， a biologically inspired fault-tolerant mechanism for resolving fluctuations in input signals is established. The experimental results on the BipedalWalker-v3 continuous control task indicate that the accessed SNN-based framework significantly outperforms the standard TD3 algorithm in terms of robustness against Gaussian noise， sensor dropout and fast gradient sign method attacks， which provides a new technical path for the development of high-reliability DRL systems in complex and uncertain environments.
According to tight coupling of DDS and flight control software in the scenarios of aerospace software， difficulty in perceiving the application layer state， and inflexibility of static QoS policy for data distribution service， the TSM-DDS framework is proposed， which is a kind of DDS-based communication service architecture for aerospace equipment applications. This architecture guarantees the high reliability of aerospace communication through a triple mechanism： 1） a dual-task inter-process communication architecture that enables the asynchronous decoupling of the control and communication domains； 2） a transaction management-based handshake and retransmission mechanism that achieves perceiving execution state at the application layer； 3） a dynamic traffic control policy that ensures the high reliability of control information in high-concurrency transmission scenarios. The simulation results demonstrate that the TSM-layer handshake and retransmission mechanism can accurately perceive the execution state of the peer application layer and effectively guarantees the execution reliability of control messages. Furthermore， in high-concurrency packet loss-prone environments， the transmission reliability of control messages can be effectively safeguarded by deploying the dynamic traffic control policy. Ultimately， the complex QoS configurations of DDS is simplified by the proposed framework that provides a reusable technical paradigm for the engineering implementation of DDS in the aerospace control domain.
In response to the demand of intelligence and integration in ground-based testing and launching control systems for launch vehicles， an intelligent and integrated architecture is proposed， which is comprised of test， launch and control functionality， auxiliary decision-making functionality and integrated storage-access-computation functionality. Regarding the test and launch control functionality， a systematic approach by featuring comprehensivization， modularization， and integratization is proposed， and the designs of module functions， system functions and the software architecture are completed. Regarding auxiliary decision-making functionality， a multi-agent collaborative auxiliary decision-making system based on a knowledge base of historical faults and contingency plans is developed， which achieves the objectives of rapid response and precise decision-making. Regarding the integrated storage-access-computation functionality， a lightweight intelligent computing platform based on large language models is proposed， which is realizes the integration and intelligence of data storage， data analysis and related capabilities.
Aiming at the problems of resource heterogeneity， complex constraints and difficult dynamic scheduling in hybrid-architecture measurement and control systems， a dynamic resource scheduling method named MA-DRS based on multi-agent deep reinforcement learning is proposed. The multi-objective， multi-constraint scheduling problem is modeled as a Markov decision process， and a multi-agent collaborative framework based on centralized training and decentralized execution is designed. The key techniques such as heterogeneous feature embedding， prioritized experience replay and dynamic action masking are introduced to improve learning efficiency and decision-making rationality. Simulation experiment results show that， compared with fixed combination， rule-based scheduling， genetic algorithm and single-agent DQN， resource utilization rate， task completion rate， and response speed are significantly imporved by using this proposed MA-DRS method in the scenarios of normal， high-concurrency and fault-injection which demonstrates excellent adaptability and robustness. It can serve as an effective solution for intelligent scheduling in complex measurement and control systems.
