This paper proposes a multiple unmanned aerial vehicle (multi-UAV) formation obstacle avoidance strategy based on the dynamic vector velocity obstacle (DVVO) method, aiming to address the design challenges of the VO method in three-dimensional (3D) dynamic space applications. By considering the relative position and velocity between the UAV and obstacle, a dynamic avoidance plane (DA-Plane) is constructed, which can accurately reflect the collision risk of UAVs in 3D space. On this basis, a dynamic vector velocity obstacle space is designed, and the optimal constant critical escape velocity is selected outside this space to accomplish the obstacle avoidance task, thereby ensuring the smoothness of the flight process. Subsequently, this paper extends the DVVO method to multi-UAV formation based on fuzzy control and the virtual spring algorithm, and designs a differentiated obstacle avoidance strategy to ensure the safe operation of multi-UAV formation in dynamic environments. Simulation and real-world experiments demonstrate that the proposed method enables multi-UAV formation to effectively avoid dynamic obstacles in 3D space and reach the target in a desired formation shape. Compared to the citation algorithm, the DVVO method shows superior performance in terms of smooth motion and computational efficiency. This study effectively simplifies the complex obstacle avoidance problems of multi-UAV formation in 3D dynamic environments, demonstrating significant potential for applications.

In the article, a position-verified (PV) kinodynamic planning method is proposed based on the extended kinodynamic rapid-exploration random tree (KRRT*) algorithm. The aim is enabling microaerial vehicles (MAVs) to generate efficient and feasible motion trajectories in unknown complex environments. An initial trajectory is generated first for analyzing obstacles between start and end points. Then, a sampling strategy with adaptive obstacle classification is designed, and is integrated into the KRRT* method for verifying position of sampling nodes. In collision-prone regions, cylindrical inflated spaces are constructed, so as to improve distribution of the sampling nodes. During the rewiring process, redundant sampling nodes and solving unnecessary boundary value problem (BVP) are eliminated to generate high quality front-end trajectories. Moreover, a trajectory curvature penalty term is designed and combined with penalty terms of trajectory smoothness and path topology deviation, and the quadratic programming (QP) model is constructed. Finally, comparative simulations and experiments are conducted to verify effectiveness and flexibility of the proposed method.

Fusing convolutional neural networks (CNNs) and Trans-former models, recent depth completion methods have shown satisfying performance by effectively capturing both local features and global contextual information. But the fusion patterns for depth completion demand high compu-tational cost and increase model complexity, hindering real-time deployment. This paper proposes a Lightweight CNN-TransFormer Fusion model (LWFusionFormer) for efficient depth completion. The proposed model integrates lightweight convolution and attention mechanism, and the convolution-Transformer fusion strategy is designed to enhance both efficiency and flexibility. By utilizing depth-wise separable convolution, LWFusionFormer significantly reduces floating-point operations (FLOPs) and improves inference speed. The Selective Fusion with Depthwise Separable Convolution (SF-DSC) module is designed to efficiently extract and fuse features, integrating spatial and contextual information through lightweight convolutional operation. Moreover, the hybrid decoder is designed to reduce computational demand, and to enhance model flexibility and practicality. Comparative results show that LWFusionFormer achieves high accuracy with lower computational costs. Note to Practitioners-The proposed method is motivated by requirements for depth perception in autonomous driving scenarios, where LiDAR measurements are sparse and onboard computational resources are limited. Existing depth completion approaches typically focus on convolutional modeling for local geometry and on Transformer-based global context aggregation, which often require trade-offs between computational cost and reconstruction accuracy. The lightweight depth completion framework LWFusionFormer is proposed in the paper, which integrates convolutional feature extraction for local geometric reconstruction. A compact global fusion mechanism is designed to aggregate contextual information. In experiments, the framework reduces computational complexity and maintains structural consistency, providing dense and accurate depth maps.

Trajectory planning for dual unmanned ground vehicles (UGVs) in cooperative carrying remains challenging in complex environments. The relative distance constraint imposed by the shared payload significantly increases the difficulty of cooperative trajectory planning. To address this issue, this paper proposes a reinforcement learning-based dual-UGV cooperative trajectory planning method with adaptive target entropy regulation. Specifically, a task-oriented local observation representation and a cooperative reward function are jointly designed for target guidance, distance maintenance, and obstacle avoidance, so that the learned policy can better satisfy the cooperative control objectives. Moreover, a target entropy regulation mechanism driven by relative distance error is incorporated into the maximum-entropy policy optimization process, enabling adaptive adjustment of the exploration intensity according to the current distance error. These designs are unified within a Multi-Agent Soft Actor-Critic (MASAC) framework under the centralized training and decentralized execution (CTDE) paradigm, forming a complete learning-based solution for dual-UGV cooperative carrying. Simulation results, including validation, comparative, and ablation studies, demonstrate that the proposed method can generate safe coordinated trajectories and achieve high relative-distance maintenance accuracy.

Vector-thrust quadrotors have emerged as a versatile solution for complex aerial missions requiring full actuation and decoupled 6-DOF motion. However, the practical deployment of these 8-DOF platforms is severely hindered by intense nonlinearities arising from tilting mechanisms and the heavy computational burden of real-time control allocation under strict actuator constraints. Existing control strategies often suffer from asymptotic convergence lags or fail to maintain allocation precision during aggressive maneuvers. To bridge these gaps, a unified robust control and allocation framework is presented. Initially, a global dynamic model is established on the SO(3) manifold to avoid singularities, integrated with a Fixed-Time Extended State Observer (FxTESO). Unlike conventional observers, the FxTESO ensures that disturbance estimation errors converge within a predefined timeframe independent of initial states, providing superior rejection of non-vanishing aerodynamic torques and gyroscopic couplings. Furthermore, to resolve the accuracy-efficiency dilemma in over-actuated systems, a Hybrid ANN-SQP Control Allocation (HASCA) strategy is developed. By leveraging an offline-trained Artificial Neural Network to provide a high-quality "warm-start" initialization for an online Sequential Quadratic Programming solver, the framework significantly accelerates convergence while strictly adhering to physical actuator limits. Comparative simulations demonstrate that the proposed unified approach markedly outperforms existing methods in tracking precision, disturbance rejection stiffness, and computational responsiveness, rendering it highly suitable for high-performance vector-thrust UAV applications.

Visual-Inertial Odometry (VIO) estimates system pose by fusing visual and inertial measurements. Although line features can enhance structural perception, existing approaches still face challenges such as redundant short segments and weak geometric constraints. To address these, in the front end, we propose a complete geometric optimization pipeline for line features. This pipeline adopts a length-threshold-based filtering strategy and integrates the proposed geometric-consistency-based merging mechanism, endpoint-distance-based verification mechanism, and epipolar-constraint-based triangulation method, transforming fragmented short segments into structurally complete 3D spatial lines. In the back end, reprojection residuals of the optimized line features are jointly optimized with point residuals, IMU pre-integration residuals, and marginalization priors in a sliding-window framework. Experiments on the EuRoC dataset show that compared to VINS-Mono, PL-VINS, and EPLF-VINS, the proposed method reduces the Absolute Pose Error (APE) by 17.57%, 9.88%, and 6.65%, respectively. Additionally, compared to PL-VINS, it reduces the line feature processing time by 4.16% and the average per-frame processing time by 2.36%, validating the effectiveness of the proposed method.

An adaptive control strategy is proposed in the paper for trajectory tracking of stratospheric airships, capable of handling uncertainties and external disturbances effectively. Due to intricate flight traits of airships, it is a challenge to develop an adaptive controller that considers all inertia/mass parameters and separable aerodynamic coefficients. Firstly, a key structural property is established for stratospheric airship dynamics, namely that the airship dynamics can be linearly parameterized. Then, an adaptive feedforward controller is designed, which does not depend on precise knowledge of airship dynamics, and the adaptive updating law can effectively compensate for uncertain internal parameters and external disturbances. Then, a continuous nonlinear feedback entity is designed based on adaptive update law and desired regression matrix. Overall stability analysis is provided based on LaSalle-Yoshizawa corollary extension and Lyapunov technique, and trajectory tracking errors converge to zero asymptotically. Comparative tests are conducted to confirm efficacy of the proposed adaptive control strategy in executing various tasks.

A model predictive control (MPC) strategy is proposed based on state observation and updating for image-based visual servoing (IBVS) tasks of micro aerial vehicles (MAVs). This control strategy enables precise pose adjustment of MAVs without relying on the global positioning system (GPS). Specifically, image features are first defined on a virtual image plane to decouple the translational motion of the MAV. Subsequently, a linear velocity observer is developed to provide high-quality real-time velocity information for the MAV during IBVS execution. Furthermore, the image dynamics on the virtual image plane are linearized using a first-order Taylor expansion, and a linear MPC controller is formulated to efficiently compute the optimal control inputs. Moreover, the state inputs to the MPC controller are updated at each control cycle to eliminate errors accumulated during the rolling optimization based on the linearized dynamics, thereby ensuring the precision of IBVS. Simulation and experimental results demonstrate the performance of the proposed observer and control strategy.

Unsupervised domain adaptation (UDA) for semantic segmentation enables models trained on labeled source domains to generalize to unlabeled target domains. In autonomous driving, cross-domain shifts caused by weather, lighting, and object often degrade segmentation performance, especially for recognizing key objects. In the paper, a novel UDA framework is proposed as Domain Adaptation Encoder-Decoder Model (DAEDM). It is built upon the Vision Mamba architecture, and the state space modeling mechanism is used to capture long-range contextual dependencies. Then, the domain invariant context enhancement (DICE) module is designed, which introduces statistical perturbation to low-confidence target regions, so as to enhance feature diversity and robustness. Moreover, the decoder incorporates the channel attention mechanism and the Mamba structure, aiming to improve multi-scale feature fusion and detail recovery. Three domain adaptation tasks are conducted in experiments, that are GTA5 to Cityscapes, SYNTHIA to Cityscapes, and Cityscapes to ACDC. The results show thatDAEDM outperforms the compared UDA methods, particularly in segmenting small objects and rare categories. These results demonstrate superior cross-domain generalization and practicability of the proposed framework.

For a stratospheric airship under internal and external uncertainties, controller strategies with robust integral sign of the error (RISE) are proposed for asymptotic trajectory tracking control. Specifically, a fully connected neural network is constructed with adaptive weight updating to estimate internal uncertainties of the airship system. Then, a self-tuning RISE controller is designed to track trajectories with compensating external disturbance, which enables gain adjustment in real time according to system feedback. Based on the Lyapunov technique and LaSalle-Yoshizawa corollary extension, system stability is strictly analyzed, and asymptotic convergence performance is obtained for trajectory tracking by the designed controller. Moreover, a RISE controller is designed with fixed-gain matrices for the airship based on classic control strategies, and asymptotic tracking is also proved. Finally, numerical simulations are conducted with spiral and hovering control to verify the performance of the proposed strategies.

To enhance tracking capability of the robust integral of the sign of the error controller (RISE) structure, an RISE controller with time-varying gains is proposed in this paper, which is incorporated with feedforward estimation from a fully connected neural network (FCNN). Specifically, for uncertain Euler–Lagrange dynamic systems, two error filters are designed to obtain faster convergence for tracking errors. Then, an adaptive update law is constructed for adjusting FCNN weights online, so as to approximate unknown nonlinear component in dynamic systems. To efficiently track trajectories, a self-tuning RISE feedback controller is designed to decrease high gains and avoid chattering phenomenon. System stability is rigorously analyzed by utilizing Lyapunov technique and LaSalle–Yoshizawa corollary extension, and asymptotic tracking convergence is proved. Finally, regarding robotic mechanisms, comparative tests are conducted to validate tracking performance of the proposed strategy.

Generating an executable reference trajectory in realtime is a core of autonomous navigation for a quadrotor uncrewed aerial vehicle (UAV). In the paper, a time-adjusted optimization method is proposed for UAV motion planning to achieve safe and fast flight. Firstly, based on parameterization for uniform B-spline trajectories, relationship of knot spans and UAV speeds is analyzed, and an iterative velocity restrict algorithm is introduced to tackle nonexecutable flight. Subsequently, combining start and target points in local planning, an evaluation criterion function is designed with control points of the above trajectory and directions of collision avoidance. On this basis, risk assessment is conducted according to surrounding environments, where four phases of passing obstacles is constructed by considering risk level variation. Then, the weight function in each phase is determined by relationship of obstacle and planned local trajectory. Moreover, time intervals of B-spline trajectories are adjusted to obtain nonuniform trajectories, so as to affect UAV speeds. Simulation and experiment are conducted with comparison to verify performance of the proposed method.

By improving observer estimation accuracy, an active disturbance rejection control (ADRC) strategy is proposed for an unmanned ground vehicle (UGV), where dynamic performance is enhanced by disturbance suppression during trajectory tracking. In the strategy, part information of the lateral control model is added to the state matrix, so as to compensate estimation bias of UGV states from an extended state observer (ESO). Then, a free variable for estimating higher-order disturbance states is integrated into the highest-order variable of the ESO, and ESO accuracy is thus improved for total disturbance estimation. To achieve accurate suppression of complex disturbance, estimates of UGV total disturbance are added to the control law according to feedback linearization theories. System stability is proved by theoretical analysis, and estimation accuracy for system states and total disturbance are enhanced by above improvement. Finally, comparative tests are conducted under multiple cases to verify the performance of the proposed strategy.

An output feedback control method is proposed in the article with state observers for robust image-based visual servoing (IBVS) regarding micro aerial vehicles (MAVs). Specifically, image features are first selected from the perspective image moments and are projected onto a suitably defined image plane so as to provide decoupled kinematics for translational motion. In order to address the problem of low quality and even unavailability of linear velocity information, the image moment error observer is designed as a foundation based on an image filter, and then the MAV body velocities are estimated by designing a robust observer to obtain translational optical flow of virtual image features. An adaptive law is proposed for the unknown depth information of the visual target and is combined with observers in output feedback controller design to enhance system robustness. Stability analysis shows that all states of the system are uniformly ultimately bounded (UUB) and converge to zero asymptotically. Comparative simulation and experimental results demonstrate the performance of the proposed method.

Depth completion is a technique to densify the sparse depth maps acquired by depth sensors (e.g., RGB-D cameras, LiDAR) to generate complete and accurate depth maps. This technique has important application value in autonomous driving, robot navigation, and virtual reality. Currently, deep learning has become a mainstream method for depth completion. Therefore, we propose an edge-enhanced dynamically routed adaptive depth completion network, EDRNet, to achieve efficient and accurate depth completion through lightweight design and boundary optimisation. Firstly, we introduce the Canny operator (a classical image processing technique) to explicitly extract and fuse the object contour information and fuse the acquired edge maps with RGB images and sparse depth map inputs to provide the network with clear edge-structure information. Secondly, we design a Sparse Adaptive Dynamic Routing Transformer block called SADRT, which can effectively combine the global modelling capability of the Transformer and the local feature extraction capability of CNN. The dynamic routing mechanism introduced in this block can dynamically select key regions for efficient feature extraction, and the amount of redundant computation is significantly reduced compared with the traditional Transformer. In addition, we design a loss function with additional penalties for the depth error of the object edges, which further enhances the constraints on the edges. The experimental results demonstrate that the method presented in this paper achieves significant performance improvements on the public datasets KITTI DC and NYU Depth v2, especially in the edge region’s depth prediction accuracy and computational efficiency.

Research on adaptive control for time-varying parameters (TVPs) in nonlinear systems has gained considerable interest. The trajectory tracking problem for nonlinear Euler-Lagrange (EL) systems is investigated in the paper under TVPs and external disturbance, and we propose a continuous projection adaptive controller to improve tracking efficiency and reduce control amplitude. Firstly, two filtered tracking errors are designed to obtain faster convergence for tracking errors. Then, the adaptive update law is designed with continuous projection regarding the TVPs, especially for non-linearly-parameterized systems. Subsequently, the controller is designed by constructing a continuous auxiliary term to realize trajectory tracking. System stability is rigorously analyzed by utilizing Lyapunov technique and LaSalle-Yoshizawa corollary extension, and theoretical analysis demonstrates that all signals of the closed-loop error system remain bounded and tracking errors converge to zero. Comparative tests are performed to evaluate the effectiveness of the proposed strategy.

Velocity saturation constraints are a significant issue for wheeled mobile robots (WMRs) when designing kinematics-based control laws. To handle the problem of velocity saturation constraints, a novel monocular visual servoing controller is developed for WMRs to solve tracking problems and enable unknown depth estimation. By analyzing the kinematic model of the robot system and employing the homography decomposition technique, measurable signals are obtained to develop a visual tracking error model for non-holonomic mobile robots. To ensure that the velocity commands are consistently constrained within the allowed limits, a saturation function is employed in the designed visual servoing control law. Furthermore, an adaptive updating law is designed to estimate the unknown depth information. The boundedness of the velocity commands is analyzed to evaluate the saturation performance of the developed visual servoing controller. With the aid of Lyapunov techniques and Barbalat’s lemma, the stability of this scheme is demonstrated. The simulation and experiment verify the performance of the proposed method.

For the visual-inertial odometry (VIO) in a robotic system, the performance of localisation can be improved by taking into account uncertainties of sensors. In this paper, an initialisation strategy is proposed with respect to monocular visual-inertial localisation, which optimises the temporal offset between a camera and an inertial measurement unit (IMU). Moreover, persistent refinement is involved in front-end processes, and specific states are optimised in back-end processes. Firstly, with assumption that IMU timestamps are valid, influence of the temporal offset between camera and IMU timestamps is analysed, and the temporal offset is derived from reprojection errors so as to align camera and IMU samplings. Then, taking into account camera samplings, both translation and rotation movement is calculated for projected 3D points, and variation of reprojected image features is analysed with respect to the corresponding camera keyframe. At last, the temporal offset and specific states are estimated by optimising reprojection errors, making that influence of the uncertain temporal offset is eliminated for visual-inertial localisation. Comparative experiments are conducted to validate the performance of the proposed approach.

Purpose The purpose of this paper is to propose a new velocity prediction navigation algorithm to develop a conflict-free path for robots in dynamic crowded environments. The algorithm BP-prediction and reciprocal velocity obstacle (PRVO) combines the BP neural network for velocity PRVO to accomplish dynamic collision avoidance. Design/methodology/approach This presented method exhibits innovation by anticipating ahead velocities using BP neural networks to reconstruct the velocity obstacle region; determining the optimized velocity corresponding to the robot’s scalable radius range from the error generated by the non-holonomic robot tracking the desired trajectory; and considering acceleration constraints, determining the set of multi-step reachable velocities of non-holonomic robot in the space of velocity variations. Findings The method is validated using three commonly used metrics of collision rate, travel time and average distance in a comparison between simulation experiments including multiple differential drive robots and physical experiments using the Turtkebot3 robot. The experimental results show that our method outperforms other RVO extension methods on the three metrics. Originality/value In this paper, the authors propose navigation algorithms capable of adaptively selecting the optimal speed for a multi-robot system to avoid robot collisions during dynamic crowded interactions.

Three types of heat transfer structures were constructed base on a LaNi5-type rare earth-based hydrogen storage alloy as hydrogen storage medium by the method of strengthen heat conduction and mass transfer,and the heat transfer charac-teristics of each structure and their loading characteristics in solid state hydrogen storage tanks were measured.Importantly,we conducted the key research on hydrogen charging/discharging of solid state hydrogen storage tanks,then compared and evaluated the effect of three types of heat transfer structures on the hydrogen charging/discharging performance of the filled sol-id state hydrogen storage tanks,and thoroughly analyzed the heat conduction behavior of solid state hydrogen storage tanks dur-ing the hydrogen charging/discharging process using simulation methods finally.The results show that:The incorporation of a-luminum and graphite,which are high thermal conductivity materials,as well as the unique design of heat transfer structures,extremely improve the thermal conductivity and hydrogen charging/discharging thermodynamic and kinetic level of solid state hydrogen storage tanks.Among the three types of heat transfer structures,RE09 particles/Al scrap/Al fins composite structure(structure-Ⅱ)exhibits the optimum hydrogen charging/discharging performance because of its stable structure and distin-guished heat and mass transfer efficiency.Simulation results show that the temperature of three types of heat transfer structures decreasing gradually outward along the radius,which means that the core of the structures turn into the hottest zone,and the maximum temperature of RE09 particles/Al scrap/Al fins composite structure(structure-Ⅱ)and RE09 powder/expanded graphite powder compressed disks structure(structure-Ⅲ)is almost 400 ℃ lower than RE09 particles structure(structure-I).Heat flux distribution connects closely with heat transfer structure form,excellent heat transfer structure design stimu-lates the heat flux flowing to the higher thermal conductive medium rather than gathering in alloy,and elevate the hydrogen charging/discharging kinetic level of solid state hydrogen storage tanks.