ABSTRACT Task‐oriented grasping (TOG) plays a crucial role in enabling robots to perform complex semantic tasks within embodied AI research. However, most existing TOG frameworks adopt an open‐loop, serial reasoning architecture that disconnects task decision making from grasp perception, hindering joint reasoning and closed‐loop refinement from execution feedback, thus limiting TOG performance in cluttered scenes. To address this issue, we propose a closed‐loop cross‐level joint reasoning for TOG in cluttered environments. In the high‐level phase, a vision–text encoder maps the task to a pixel‐wise mask, which serves as a spatial prior for reasoning in a vision‐language model and yields high‐level knowledge priors. In the low‐level phase, a point‐cloud encoder predicts high‐confidence grasp regions under this prior as intermediate representations connecting high and low level reasoning. We then integrate high‐level priors with low‐level perception to evaluate grasp candidates. Finally, we establish a vision‐based feedback loop where the high‐level module verifies execution outcomes and updates the state, thereby achieving closed‐loop control of “perception–redecision.” We evaluate CLCR‐Grasp on the public TaskGrasp dataset against GCNGrasp, GraspGPT and VLA‐Grasp. Results show that CLCR‐Grasp improves the TOG success rate by 35.56%, 9.46% and 4.82% over these methods, respectively, demonstrating its effectiveness on benchmark task‐oriented grasping. The real‐time robot validation video is available at: https://github.com/naiyouzhang11/CLCR‐Grasp‐RealRobot‐Demo .
ABSTRACT Short‐term electricity load forecasting (STLF) is important for smart grid operations but challenging due to non‐linear electricity demand. A stacked ensemble electricity load forecasting (ST‐ELF) machine learning (ML) model is proposed by stacking support vector extreme gradient boosting regression (XGBR) and random forest regression (RFR) as first‐level learning models and linear regression (LR) as a second‐level aggregation model. The LR, RFR, XGBR and SVR are also trained as baseline models. SHapley Additive exPlanations (SHAP) and principal component analysis (PCA) are used for feature analysis. The training and evaluation of the models using historical and weather‐based datasets from Panama revealed that the ST‐ELF model outperformed all the baseline models against the fivefold cross‐validation with a coefficient of determination () of 0.93 and root mean squared error (RMSE) of 49.45 MW, and on an independent test set with an of 0.89 and RMSE of 61.48 MW. The proposed ST‐ELF showed an 18.2% RMSE reduction over previous studies with interpretability for weather‐sensitive power systems.
ABSTRACT Underwater fixed sonar systems are often constrained by limited observability when only bearing measurements are available, posing significant challenges for reliable target detection and tracking. To overcome this limitation, this paper proposes a generalised Radon transform (GRT)‐based sonar signal processing framework that introduces radial‐velocity‐time records (RVTRs) for target motion parameter estimation. By applying the GRT to RVTRs, target velocity and heading angle are jointly estimated through a global energy accumulation process, eliminating the need for initial‐state information and iterative optimisation. Furthermore, a closed‐form solution for range estimation is derived by integrating the estimated motion parameters with bearing measurements. Sea‐trial results demonstrate that the proposed method achieves accurate range estimation with strong robustness and practical applicability for underwater target tracking and surveillance.
ABSTRACT Blind emitter candidate generation from power‐spectrum measurements is a fundamental task in regulatory spectrum monitoring, where only spectral information is available, and signal components must be distinguished from background spectral variations. This letter proposes a machine learning–assisted approach to blind emitter‐candidate generation that operates directly on averaged power spectra. The method extracts local spectral features and applies supervised classification to distinguish signal‐bearing from noise‐only frequency bins. Performance is evaluated using a simulation framework comprising 100 randomised frequency‐modulated (FM) broadcast band monitoring scenarios under additive white Gaussian noise, with multiple simultaneously active emitters. Several supervised classifiers are evaluated and compared using detection and false alarm metrics. Experimental results show that the proposed approach achieves detection probabilities comparable to conventional peak‐based methods while significantly reducing the false alarm probability. Statistical analysis indicates that random forest classifier (RFC), support vector machine with radial basis function (SVM‐RBF) kernel and linear discriminant analysis (LDA) provide the best overall trade‐offs. The results demonstrate that machine learning–assisted candidate generation can effectively reduce false alarms without sacrificing detection performance, thereby alleviating the burden on subsequent filtering stages in blind spectrum monitoring systems. Source code and data will be available at https://github.com/eliezersflores/ml‐assisted‐emitter‐candidate‐generation .
This letter investigates the spatial separability of closely-located users in cell-free massive multiple-input multiple-output (CF-mMIMO) systems using high-diversity real-world channel measurements. We evaluate the orthogonality of four closely-located users under various CF topologies. To ensure a fair comparison, a constant aggregate number of 64 transmit antennas is maintained across all configurations. Our results demonstrate that transitioning from a co-located to a distributed architecture significantly enhances spatial multiplexing: increasing the number of access points (APs) from one to eight reduces both the median channel correlation and the condition number by approximately 50%. However, scaling beyond this threshold results in a performance saturation, as additional APs contribute negligible power due to decreasing large-scale fading coefficients. These findings suggest that excessive AP densification leads to marginal improvements, highlighting a practical trade-off between spatial diversification and the effective signal strength of additional distributed APs.
