UAV imagery semantic segmentation underpins the expansion of the low-altitude economy, with immense practical application value and broad development prospects, yet its advancement is constrained by scarce pixel-level annotated datasets. Data augmentation is a core solution to this bottleneck. Although existing data augmentation methods improve model robustness to some extent, they still exhibit notable limitations. Traditional methods merely realize superficial data variation and fail to introduce novel semantic structures, while generative methods rely on single priors, leading to poor structural fidelity and semantic misalignment in generated UAV scenes. To address this, we propose a method based on controllable diffusion models, fusing pixel-level semantic masks, HED edge maps as visual priors and category-enriched text prompts as semantic guidance via a custom control adapter. Specifically, it excels in preserving UAV scene structural and semantic consistency; simultaneously, it enriches dataset texture and color diversity without re-annotation. Extensive experiments on UAV benchmarks verify its superiority, significantly boosting segmentation model mIoU and advancing generative UAV data augmentation.

Autonomous exploration with Unmanned Ground Vehicles (UGVs) demands a balance between exploration efficiency and navigation safety. However, existing methods typically focus on known regions or frontiers, neglecting the informative potential of unknown regions, resulting in redundant backtracking or suboptimal efficiency. To address this issue, we introduce HSURE, a hierarchical safety-aware autonomous exploration framework guided by unknown regions. We explicitly decompose unknown regions by topological connectivity and construct an observability-driven safety assessment using dynamic sparse graphs in known space. Leveraging these hybrid representations, we devise a coarse-to-fine safety-aware exploration planner. The global planner computes an optimal coverage path and the local planner refines this into a safe and executable trajectory. Extensive simulation and real-world experiments demonstrate the proposed method outperforms state-of-the-art baselines, significantly reducing exploration time and trajectory redundancy while maintaining navigational safety.

This review explores the intersection of visual sensing and decentralized Simultaneous Localization and Mapping (SLAM), motivating the development of decentralized visual SLAM. To the best of our knowledge, exclusive reviews of decentralized SLAM frameworks are limited. We first discuss the motivations and technical challenges, such as computational and communication scalability, that these systems encounter. We then outline the mathematical formulation of the collaborative SLAM problem and review recent end-to-end decentralized visual SLAM systems to highlight the latest advancements. After that, we break down a typical decentralized visual SLAM pipeline into key modules, briefly summarizing the progress in each with greater focus on decentralized elements. Following, we discuss the common evaluation techniques and public datasets for the systematic assessment of such systems. Last but not least, we highlight open problems and potential research directions to further advance this field.

Due to the ability to provide point clouds and Doppler velocity, as well as the adaptability in harsh weather conditions, 4D Radar has emerged as a new option for Simultaneous Localization and Mapping (SLAM). However, there is limited research on both robustness and computational efficiency, which are essential requirements for deploying 4D Radar as a specialized sensor in harsh weather. This paper proposes a fast and robust Radar-Inertial-Odometry (RIO) approach, named FAR-RIO. Specifically, we propose a dynamic point filtering based on full covariance propagation, along with an isotropic uncertainty model for accurate registration. In particular, leveraging the dual-observation capability of 4D radar, we design a novel dual-observation update pipeline for the iterated error-state Kalman filter (IESKF), coupled with a corresponding keyframe selection strategy, significantly reducing computational load while minimizing accuracy loss. Our method approaches state-of-the-art performance on various sensors and runs significantly faster than existing methods.

Research in aerial swarms have gained traction in recent years and there appears to be a lack of user-friendly frameworks with a focus on bringing swarm UAVs from simulation to actual flight. Furthermore, spatial constraints and resource challenges hinder the validation of larger-scale swarm algorithms. To tackle these issues, we propose Gestelt, a relatively lightweight framework that accelerates the sim-to-real transition for swarm algorithms. First, the modular design of Gestelt is highlighted to illustrate it's generalization to multiple types of planning algorithms and paradigms. Next, we outline an approach to model any given quadrotor platform for use in our framework's simulation environment. Another unique feature of Gestelt is it's virtual-physical environment, which can simultaneously host both virtual and physical agents, thereby providing an intermediate platform for testing larger-scale swarm algorithms safely. Finally, we implement an asymptotically stable closed-loop control technique known as Robust Perfect Tracking (RPT) to track the reference trajectories of the swarm agents in the face of disturbances. We demonstrate our framework through physical experiments featuring our custom swarm platform, the NUSwarm drone, where we show a swarm navigation scenario for 3 physical and 3 virtual drones. A video demonstrating the virtual-physical environment can be found at https://youtu.be/FYlwz2yZxLE

This work presents FG-Net, a general deep learning framework for large-scale point cloud understanding without voxelizations, which achieves accurate and real-time performance with a single NVIDIA GTX 1080 8G GPU and an i7 CPU. First, a novel noise and outlier filtering method is designed to facilitate the subsequent high-level understanding tasks. For effective understanding purpose, we propose a novel plug-and-play module consisting of correlated feature mining and deformable convolution-based geometric-aware modeling, in which the local feature relationships and point cloud geometric structures can be fully extracted and exploited. For the efficiency issue, we put forward a new composite inverse density sampling (IDS)-based and learning-based operation and a feature pyramid-based residual learning strategy to save the computational cost and memory consumption, respectively. Compared with current methods which are only validated on limited datasets, we have done extensive experiments on eight real-world challenging benchmarks, which demonstrates that our approaches outperform state-of-the-art (SOTA) approaches in terms of accuracy, speed, and memory efficiency. Moreover, weakly supervised transfer learning is also conducted to demonstrate the generalization capacity of our method.

In this article, we present a comprehensive design and implementation for a micro aerial vehicle (MAV) that is able to perform 3-D autonomous navigation and obstacle avoidance in cluttered and realistic unknown environments without the aid of global positioning system and other external sensors or markers. To achieve these autonomous missions, modularized components are developed for the MAV, including visual-inertial odometry, 3-D occupancy mapping, and motion planning. The proposed system is implemented to run on a small embedded computer in real time. It is demonstrated to be robust in both simulation and real flight experiments. The demonstration video is available at: https://youtu.be/KUKzsnORm-4.

Below-canopy UAVs hold promise for automated forest surveys because their sensors can provide detailed information on below-canopy forest structures, especially in dense forests, which may be inaccessible to above-canopy UAVs, aircraft, and satellites. We present an end-to-end autonomous system for estimating tree diameters using a below-canopy UAV in parklands. We used simultaneous localization and mapping (SLAM) and LiDAR data produced at flight time as inputs to diameter-estimation algorithms in post-processing. The SLAM path was used for initial compilation of horizontal LiDAR scans into a 2D cross-sectional map, and then optimization algorithms aligned the scans for each tree within the 2D map to achieve a precision suitable for diameter measurement. The algorithms successfully identified 12 objects, 11 of which were trees and one a lamppost. For these, the estimated diameters from the autonomous survey were highly correlated with manual ground-truthed diameters (R2=0.92, root mean squared error = 30.6%, bias = 18.4%). Autonomous measurement was most effective for larger trees (>300 mm diameter) within 10 m of the UAV flight path, for medium trees (200–300 mm diameter) within 5 m, and for trees with regular cross sections. We conclude that fully automated below-canopy forest surveys are a promising, but still nascent, technology and suggest directions for future research.

This work presents a general deep learning framework for large-scale point clouds understanding without voxelizations, called FG-Conv, which achieves an accurate and real-time understanding of point clouds. Through our novel design combining feature level correlation mining and deformable convolutions based geometric aware modeling, the local feature relationships and geometric patterns can be captured. The attention mechanism is also adopted to enhance the global long-range feature correlations. Finally, the feature pyramid residual learning network is proposed to combine patterns at different resolutions in a memory-efficient way. Extensive experiments on real-world challenging datasets demonstrated that our approaches outperform state-of-the-art methods in terms of accuracy and efficiency. Weakly supervised transfer learning demonstrates the generalization capacity of our methods.

Challenges in motion planning for multiple quadrotors in complex environments lie in overall flight efficiency and the avoidance of obstacles, deadlock, and collisions among themselves. In this paper, we present a gradient-free trajectory generation method for multiple quadrotors in dynamic obstacle-dense environments with the consideration of time consumption. A model predictive control (MPC)-based approach for each quadrotor is proposed to achieve distributed and asynchronous cooperative motion planning. First, the motion primitives of each quadrotor are formulated as the boundary state constrained primitives (BSCPs) which are constructed with jerk limited trajectory (JLT) generation method, a boundary value problem (BVP) solver, to obtain time-optimal trajectories. They are then approximated with a neural network (NN), pre-trained using this solver to reduce the computational burden. The NN is used for fast evaluation with the guidance of a navigation function during optimization to guarantee flight safety without deadlock. Finally, the reference trajectories are generated using the same BVP solver. Our simulation and experimental results demonstrate the superior performance of the proposed method.

We present in this paper an all-in-parallel mapping and perception framework for robotic navigation. When performing motion planning in cluttered environments, a local map is required to be updated real time. It is often necessary for a gradient-based trajectory optimization method to retrieve both occupancy information and distance to obstacles, in which occupancy grid maps (OGMs) and Euclidean signed distance fields (ESDFs) are often employed as a bridge between robotic perception and planning. To build the OGMs and ESDFs fast and efficiently, we propose a complete map construction system that is compatible with the commonly used on-board sensors. The proposed system constructs the OGMs and ESDFs in parallel on GPUs. The map increases with the movement of the vehicle by adapting a hashing data structure. The new data are fused continuously to the existing map by employing the parallel wavefront algorithm. Our approach can achieve real-time performance in limited on-board computing resources. It has been showed to outperform many existing real-time mapping systems that generate ESDFs on micro aerial vehicles.

An unconventional drone design and controller for fast and accurate trajectory tracking are proposed in this paper. A dynamic model was developed, the controllability is analysed and physical parameters to maximise controllability are identified. A multiple-input multiple-output (MIMO) sliding mode controller via the unit vector approach was considered. A novel discrete sliding mode controller was designed via Lyapunov theory to guarantee convergence to the sliding surface. A discrete sliding mode controller with model-reference tracking was further developed to improve performance. Finally simulations of step, tracking response and disturbance rejection with significant non-idealities and non-linearities were performed to illustrate the performance of the design and controllers.

In recent years, consensus problem in multiagent systems (MASs) has been studied extensively. However, when a fault occurs in a MAS, this problem becomes more challenging. In this paper, we present several fault accommodation control schemes for a class of MASs. First, actuator faults are considered in our MASs. Second, a fault detection is proposed, which is constructed based on a residual and threshold value. Third, both passive and active fault accommodation control schemes are presented. Finally, a simulation example is given to show the effectiveness of the proposed control method.

In this paper, we propose a real-time disparity map estimation framework on FPGA, combined with an effective post-processing method. Given the input stream of stereo image pairs, a semi-global matching based framework is implemented on the FPGA to estimate the disparity in real-time. The generated disparity map is refined with a weighted least square (WLS) filtering method. The experiments show that the disparity map can be reconstructed in real-time. In addition, the weighted least square filtering based post-processing can significantly improve the accuracy of the disparity map and remove large estimation errors.

A hierarchical composite nonlinear feedback (HCNF) control technique is proposed in which the auxiliary integrators are introduced to remove the steady-state bias due to the disturbances. The main idea of HCNF is to separate the control design of the auxiliary integrators and the composite nonlinear feedback (CNF) control design. The auxiliary integrators are applied only to remove the steady-state bias by automatically adjusting the commands to be tracked by the CNF control law in which the CNF control design does not involve with the auxiliary integrators. The HCNF control technique is applied to design the attitude control law of a maneuvering flight control law for a quadrotor. The resulting closed-loop system is verified in simulation successfully to complete the maneuvering flight.

This paper describes the development of a novel algorithm to tackle the problem of real-time video stabilization for unmanned aerial vehicles (UAVs). There are two main components in the algorithm: (1) By designing a suitable model for the global motion of UAV, the proposed algorithm avoids the necessity of estimating the most general motion model, projective transformation, and considers simpler motion models, such as rigid transformation and similarity transformation. (2) To achieve a high processing speed, optical-flow based tracking is employed in lieu of conventional tracking and matching methods used by state-of-the-art algorithms. These two new ideas resulted in a real-time stabilization algorithm, developed over two phases. Stage I considers processing the whole sequence of frames in the video while achieving an average processing speed of 50fps on several publicly available benchmark videos. Next, Stage II undertakes the task of real-time video stabilization using a multi-threading implementation of the algorithm designed in Stage I.

A set of algorithms are developed to estimate the air velocity of the quadrotors by fusing the plant model of the three-dimensional aerodynamic force and the IMU (inertial measurement unit) measurement. The main ideas are to compute out the air velocity of the quadrotors at first with the measured load acceleration and the model of the aerodynamic force, and then to smooth the computed results with the extended Kalman filtering against the model uncertainties. The developed algorithms are verified in the closed-loop flight control system. The simulation results demonstrate that the air velocity of the quadrotors are estimated well in the large airspeed and the resulting closed-loop system with the estimated air velocity can achieve almost the same flight performances as that with the real air velocity.

In this paper, a novel motion planning approach is proposed for high-speed flight of Micro Air Vehicles (MAVs) in unknown cluttered environment using a stereo camera as unique sensor. The proposed motion planning works directly on a depth map which represents partial surrounding environment due to limited field of view (FOV) of the stereo camera. The depth map updates at each sampling time. An optimal local target is selected from a set of local targets to which line-of-sight paths are collision-free. Then an optimal collision-free trajectory satisfying kinematic constraints is generated. The simulation results demonstrate the effectiveness of the proposed motion planning approach.

In this paper, we present a novel integrated vehicular system using collaborative unmanned aerial vehicles (UAVs) and unmanned ground vehicles (UGVs) for autonomous exploration, mapping, and navigation in GPS-denied 3-dimensional (3-D) unknown environments. The system implements a novel two-layered exploration strategy to decompose the perception task into a coarse exploration layer and a fine mapping layer. The coarse exploration makes use of a UGV to carry out fast autonomous exploration and active 2.5D simultaneous localization and mapping (SLAM) to generate a coarse environment model, which serves as a navigation reference for subsequent complementary 3-D fine mapping conducted by a UAV. The two layers share a novel optimized exploration path planning and navigation framework, which provides optimal exploration paths and integrates the collaborative exploration and mapping efforts through an OctoMap-based volumetric motion planning interface. The proposed system provides an efficient pipeline of fast environment perception taking advantages of the agility of the UAVs as well as the powerful computation resource aboard UGVs, also allowing assistive local perception with augmented object information when necessary. The effectiveness of our system is verified by both simulations and experiments, which demonstrate its capability of implementing heterogeneous UAV and UGV collaborative exploration and structural reconstruction of the environments through active SLAM, providing optimized perception for navigation tasks.

In this paper, a multi-plane infrared visual odometry (VO) algorithm is proposed for the navigation system of unmanned aerial vehicles (UAV) flying at a medium height for both day and night. It uses a monocular downward-looking and stabilized thermal camera, aided by a single laser beam to restore scale factor. The scale recovery is further improved by a series of comprehensive logics of taking different measurements into a multi-plane homography decomposition process, which makes the algorithm practicable and robust for the platform flying over various complex terrains with multiple planar surfaces. With the help of some sophisticated techniques, like keyframe replacement and subregion adaptive KLT tracker, the performance of motion estimation is further improved. This visual odometry algorithm allows UAV platforms to take various missions from a medium altitude when GPS is not available and illumination condition is poor.