Multi-UAV relative localization using onboard 3D LiDAR is challenged by intermittent target detections caused by sparse point returns and field-of-view limitations. In the outdoor experiments considered, valid detections are available in only 13–41% of scan frames, resulting in detection outages of up to 11 s, during which standard kinematic predictors accumulate substantial drift. To address this problem, a semi-cooperative dual-inertial measurement unit (IMU) framework is proposed, in which the target UAV transmits inertial measurements and attitude estimates to the observing UAV. An Extended Kalman Filter (EKF) fuses inertial data from both platforms to propagate the relative state through detection outages, while a lightweight spatiotemporal consistency mechanism rejects clutter-induced false detections prior to filter updates. Six EKF configurations are evaluated, including kinematic baselines based on constant velocity (CV) and constant turn rate and velocity (CTRV), as well as four dual-IMU formulations with different bias parameterizations. Results over 13 outdoor flight segments show that dual-IMU propagation consistently outperforms kinematic prediction, and that a minimal 9-DOF formulation provides the best trade-off between estimation accuracy, robustness, and computational cost. In particular, the proposed 9-DOF model reduces relative position root mean square error by 64% compared with the CV baseline while maintaining real-time performance during extended LiDAR detection outages.
As humanoid robots become increasingly integrated into daily environments, developing systems capable of socially intelligent and intuitive Human–Robot Interaction (HRI) is essential. Mutual gaze, the shared exchange of eye contact, is a critical nonverbal cue for attention and intent recognition. However, gaze estimation in real-world, multi-person environments remains challenging due to varying illumination, occlusions, and computational constraints.This work presents a dual-Field-of-View (FOV) gaze estimation framework for the Vizzy humanoid robot, combining wide-angle and zoom Red-Green-Blue (RGB) cameras to balance coverage and precision. A lightweight gaze estimation model based on L2CS-Net [1] was optimized through Knowledge Distillation (KD), transferring knowledge from a teacher to a compact student network. The system achieves real-time inference, ≈18 Frames per Second (FPS), with less than 3° mean Angular Error (AE) degradation compared to the teacher. Experiments conducted in Gazebo simulation and real-world scenarios confirm that the dual-FOV setup significantly improves mutual gaze recognition in dynamic, multi-person environments, enabling robots to detect human attention cues more efficiently.
This work presents a complete autonomous racing stack for a 1/10-scale F1TENTH vehicle, addressing perception, planning, and control on embedded hardware. The perception module fuses LiDAR and IMU data via an Extended Kalman Filter (EKF) and Cartographer-based SLAM for drift-compensated localization. A planner was implemented that extracts smooth centerlines from occupancy grids, coupled with Kinematic and Dynamic Model Predictive Control (MPC) in curvilinear coordinates. Experiments in simulation and in a real scenario demonstrated real-time feasibility, stable trajectory tracking, and bounded errors. This framework offers a reproducible approach for deploying optimization-based control near the limits of handling on resource-constrained systems 1.
Automated weld inspection is a critical task in industrial manufacturing, as undetected defects can compromise structural integrity, safety, and product quality. This paper investigates computer vision and deep learning models for the detection and classification of weld surface defects. A custom dataset of weld images was acquired using a dedicated image acquisition system, and several detection and classification model configurations were investigated. A hybrid detection–classification strategy is proposed, in which YOLOv8n performs defect localization and ConvNeXt-Small is employed for defect classification, with particular emphasis on underrepresented defect classes. Experimental results demonstrate that the proposed hybrid approach outperforms single-model strategies, highlighting the importance of dataset design choices and the effectiveness of deep learning-based methods for automated weld inspection.
Synthetic data has become a crucial enabler for training deep learning models in robotic perception tasks. This work presents a configurable synthetic data generation pipeline, built on BlenderProc, that leverages CAD model annotations to produce part-based labels directly at the geometry level. The proposed approach enables the automatic generation of Oriented Bounding Box (OBB) annotations and supports the definition of multiple regions of interest through a vertex-level annotation tool. The pipeline is validated in an warehouse environment use case involving a storage container for automotive parts for autonomous forklift operations. Experimental results show that YOLO11m-Oriented Bounding Box model trained exclusively on synthetic data achieve strong performance on real-world images, reaching up to 0.768 mAP@0.5:0.95. Furthermore, increasing synthetic data diversity consistently improves detection performance, and hybrid training with a small amount of real data provides an effective trade-off between annotation effort and accuracy.
