Stem removal from harvested fruits remains one of the most labor-intensive tasks in fruit harvesting, directly affecting the fruit quality and marketability. Accurate and rapid fruit and stem segmentation techniques are essential for automating this process. This study proposes an enhanced You Only Look Once (YOLO) model, AppleStem (AS)-YOLO, which uses a ghost bottleneck and global attention mechanism to segment apple stems. The proposed model reduces the number of parameters and enhances the computational efficiency using the ghost bottleneck while improving feature extraction capabilities using the global attention mechanism. The model was evaluated using both a custom-built and an open dataset, which will be later released to ensure reproducibility. Experimental results demonstrated that the AS-YOLO model achieved high accuracy, with a mean average precision (mAP)@50 of 0.956 and mAP@50–95 of 0.782 across all classes, along with a real-time inference speed of 129.8 frames per second (FPS). Compared with state-of-the-art segmentation models, AS-YOLO exhibited superior performance. The proposed AS-YOLO model demonstrates the potential for real-time application in automated fruit-harvesting systems, contributing to the advancement of agricultural automation.

This paper proposes DB-LIO (database-driven LiDAR-inertial odometry), a simultaneous localization and mapping (SLAM) system that addresses memory scalability challenges in extended autonomous operation. Existing LiDAR-SLAM systems accumulate keyframe history in memory, leading to $O\left(N\right)$ growth and out-of-memory failures during extended operation. To overcome this limitation, DB-LIO introduces three core design elements. First, it proposes a spatially indexed keyframe management scheme that persistently stores keyframes in SQLite with R-Tree spatial indexing, enabling $O(\log N+k)$ spatial queries that tightly couple cache eviction with factor-graph optimization requirements—a design that ensures every keyframe potentially involved in the next optimization cycle resides in cache. Second, it presents a four-level memory bounding architecture—SLAM-engine keyframe trimming with transparent on-demand reloading, a DB-level least recently used (LRU) cache with a spatial active window, Scan Context descriptor pool bounding, and iSAM2 sliding window compaction with a sparse global anchor graph—that collectively bounds the dominant memory consumers to $O\left(C\right)$. Third, the DB-based persistent storage enables a localization mode that can reload previously built maps—including full point clouds, six-degree-of-freedom poses, timestamps, and inter-keyframe relationships—and perform pose estimation using the stored map, which is particularly valuable for agricultural robots and other autonomous systems requiring map reuse. Experiments on a custom orchard dataset demonstrate an 81.9% reduction in memory usage compared with that of the in-memory baseline (2888 MB → 524 MB), while preserving equivalent trajectory accuracy (absolute trajectory error (ATE) root mean square error (RMSE) 0.305 ± 0.001 m vs. 0.296 m). Validation on the KITTI odometry benchmark confirms that the proposed localization mode generalizes across different LiDAR types (Livox Mid360, Velodyne HDL-64E) and environments (orchard, urban driving).

Constructing high-quality datasets for precision agriculture applications is costly and time-consuming owing to severe object occlusion and background similarity inherent in agricultural environments. The state-of-the-art zero-shot models or artificial intelligence (AI) tools in existing interactive labeling tools often exhibit limitations such as fragmentation, where objects are split into multiple pieces owing to the arbitrary re-interpretation of the model in these complex environments, or frequent failures in precise detection and tracking. Consequently, this study proposes a semi-automated labeling system, DLMI-SAM labeler, which combines Segment Anything Model 3-based video object tracking with a Human-in-the-Loop strategy. The proposed system introduces direct latent memory injection technology, which allows high-quality mask information corrected by the user to be directly injected into the memory bank in the form of binary-deterministic logit without passing through the encoder of the model, thereby fundamentally blocking re-interpretation errors of the model and logically unifying fragmented objects. Experimental results using a smart farm tomato dataset demonstrate that the proposed system effectively resolves the fragmentation problem and precision issues that existing models fail to address, achieving an accuracy comparable to that of manually labeled datasets while achieving high efficiency. Furthermore, we confirmed that unnecessary iterative corrections can be reduced to drastically improve labeling efficiency when DLMI is applied, and that the dataset constructed using this method substantially contributes to the improvement of supervised learning AI instance segmentation performance.

UAVs increasingly rely on accurate SLAM for aerial mapping and inspection in GPS-denied environments. 3D Gaussian Splatting (3DGS) has opened a new direction for UAV mapping by allowing SLAM systems to build dense, photorealistic, and renderable maps. Yet in 3DGS SLAM the map is optimized from the pose graph, so a false loop closure can deform both the UAV trajectory and the Gaussian map consumed by downstream UAV autonomy. Reliable loop admission is therefore relevant to safe GPS-denied operation because it protects the state and map estimates on which autonomous functions depend. The present work evaluated this upstream estimation-integrity problem; it did not measure closed-loop guidance, control, or navigation-safety outcomes. We address the loop-admission problem that arises after a place-recognition (PR) module proposes a candidate loop and relative-pose seed. GSSeq is a rendered-reference sequential verifier that uses the current Gaussian map as active evidence before inserting a loop factor. It renders RGB-D references with the PR seed, checks LiDAR/rendered-depth consistency and image/rendered-reference consistency over active support, and propagates the seed through a short query trajectory window. A loop is admitted only when this evidence remains geometrically supported and photometrically stable. On fixed LiDAR-PR candidate sets spanning MARS-LVIG, MUN-FRL, and independent NTU-VIRAL aerial sequences together with ground-mobility benchmarks, GSSeq provides a competitive precision-oriented operating point while suppressing false loop admissions. Thresholds calibrated only on NTU-VIRAL spms_01 combine rendered RGB agreement with LiDAR-submap geometry and are then frozen for spms_02. On this held-out sequence, GSSeq rejects all seven false-positive BTC factors while retaining one of three true-positive factors. The trajectory-to-map experiment reduced ATE RMSE from $2.609\mathrm{m}$ to $1.417\mathrm{m}$ and improved selected-view PSNR from $13.80\mathrm{dB}$ to $16.46\mathrm{dB}$. These results show that rendered verification can preserve an aligned, renderable UAV trajectory-map pair before unsupported loop factors reshape the SLAM map.

Warm-season greenhouse crops such as tomato (Solanum lycopersicum L.) and paprika (Capsicum annuum L.) require nutrient supply strategies that respond dynamically to temperature, solar radiation, vapor pressure deficit (VPD), root zone water status, and electrical conductivity (EC). This systematic review assessed four operational strategies: (1) solar radiation-proportional irrigation; (2) radiation-rate-based volume and EC adjustment; (3) supply–drainage and substrate-feedback control; and (4) temperature, humidity, and VPD-based supervisory correction. Scopus, Web of Science, IEEE Xplore, ScienceDirect, and MDPI were searched on 15 August 2026 for studies published from 2010 onward, yielding 4088 records (2326 after removal of duplicates). Studies were eligible if they addressed a controlled-environment production system, a warm-season crop, a nutrient-management component, and the use of sensing, automated control, models or decision-support systems, or IoT/AI tools. After title/abstract screening and full-text assessment, 90 studies were included: 29 studies from the database search and 61 studies identified by manual searching of publisher platforms, the three other references were from textbooks. Radiation-based approaches provide practical feedforward estimation but incompletely represent humidity and root zone dynamics. Supply–drainage feedback offers stronger closed-loop information, while temperature and VPD-based correction improves responsiveness to atmospheric demand. Tomato requires careful calcium and potassium management during reproductive stages, whereas paprika demands stable substrate moisture, EC, and drainage. A conceptual hierarchical framework is proposed in which solar radiation establishes the baseline nutrient-solution supply, radiation-rate information provides short-term adjustment, root-zone and drainage measurements provide feedback correction, and temperature or VPD serves as a supervisory layer. Ion-specific sensing, mechanistic models, and AI/ML approaches can further support nutrient-specific and predictive management. Few included studies evaluated these strategies under explicitly imposed high-temperature conditions; therefore crop, cultivar, and growth-stage-specific validation under commercial high-temperature greenhouse conditions is required before operational thresholds can be generalized.

Vision–Language–Action (VLA) models unify visual perception, natural-language understanding, and action generation within a single foundation model, allowing a robot to follow instructions such as “fold the towel” or “fly to the red building” directly from camera images. Because VLAs inherit world knowledge from internet-scale pre-training, they have become the dominant framework for learning-based manipulation, with bimanual coordination serving as the most demanding testbed: two arms with 7+ degrees of freedom each must move in concert to fold, assemble, and reorient objects. Unmanned aerial robotics faces a structurally similar challenge: a drone must coordinate thrust, attitude, and increasingly gripper commands from visual observations under strict latency and payload constraints. This review covers 183 contributions spanning 2017–2026 and organized along seven dimensions: VLA architectures, training recipes, action representations, bimanual coordination (2022–2026), unmanned aerial vehicle (UAV) navigation and control (2017–2026), language grounding, and cross-cutting concerns including memory and world models. We show that the coordination strategies, training recipes, and action representations developed for bimanual VLAs transfer to unmanned aerial systems and identify fourteen research directions across both domains.