Waseem Akram

Autonomous surface vehicles (ASVs) offer a scalable solution for marine debris collection, but their dynamics change as trash accumulates. This work explores adaptive control for underactuated water-cleaning ASVs using model-predictive control (MPC) with real-time parameter estimation via moving horizon estimation (MHE). While MHE effectively tracked linear drag variations, it required sufficient excitation for added mass estimation and was computationally expensive. We also analyzed the sensitivity of explicit hydrodynamic estimation to measurement noise and control inputs, revealing potential inaccuracies. A thorough sensitivity study on nominal MPC under varying hydrodynamic conditions, based on the assumption of simulated trash accumulation effects on system dynamics, motivated the development of an adaptive MPC controller, which showed better performance in extreme scenarios (e.g., heavy trash accumulation and strong currents). Field experiments further indicated that wind-induced currents had a greater impact on MPC performance than trash accumulation. In addition, we developed a comprehensive Python simulation platform to incorporate real-time disturbances, highlighting MPC’s limitations when the model is inaccurate. These findings underscore the need for robust control strategies in dynamic marine environments, paving the way for improved autonomy in water-cleaning ASVs.

Sign language is a vital mode of communication for the deaf and hard-of-hearing community, Pakistan Sign Language (PSL) is a natural language; in this study, we focus on the PSL manual alphabet (fingerspelling) used to represent Urdu letters. While several approaches have been proposed for PSL recognition, challenges related to limited publicly available datasets, cost-effective acquisition setups, and efficient feature representation remain insufficiently explored. This paper presents a color-glove-assisted framework for the recognition of 34 static PSL alphabets. The proposed approach involves video-based dataset collection, extraction of geometrical features, and classification using Support Vector Machines (SVM) with an optimized kernel selection strategy. Unlike earlier glove-assisted PSL works that relied mainly on fuzzy inference or pixel-level features, this study introduces geometrical feature extraction (centroid, distances, and angles) combined with optimized SVM kernel selection, offering both accuracy and efficiency. In addition, a new video-based PSL alphabet dataset is introduced, contributing to the limited resources available for PSL research. Experimental results demonstrate that the SVM with a Gaussian RBF kernel achieves the highest performance, attaining an accuracy of 93.62%. To further improve model generalization and reduce reliance on handcrafted features, a deep learning model employing the VGG16 network with fine-tuning was introduced alongside machine learning classifiers, yielding enhanced recognition performance on the custom PSL dataset. Furthermore, a lightweight MobileNet-V2 model was also evaluated to investigate computationally efficient deep learning performance. The key novelties of this work include (i) a new video-based dataset of 34 PSL static alphabets,and (ii) a cost-effective recognition framework leveraging color gloves and geometrical features. The findings validate the efficacy of this approach for 34 static alphabets and provide a foundation for future PSL alphabet recognition.

Thruster allocation is critical for the reliable operation of underwater vehicles, particularly under actuator degradation, power limitations, and thermal stress. Existing methods, such as pseudo-inverse or standard quadratic programming (QP)-based approaches, mainly minimize allocation error or energy consumption but often overlook real-time degradation and resource constraints. In this paper, we propose an adaptive fault-tolerant thrust allocation framework integrated with a PID plus Sliding Mode Control (PID+SMC) law for robust trajectory tracking. The approach leverages convex optimization to simultaneously enforce: (i) residual-driven health adaptation that down-weights degraded thrusters online, (ii) power-aware allocation ensuring operation within a global energy budget, and (iii) thermal-aware constraints that actively prevent overheating. A lightweight residual filter continuously updates thruster health indices, enabling rapid reallocation under faults and efficiency loss. Simulation results across nominal, power-limited, thermal-limited, faulted, and combined scenarios show that the proposed method reduces trajectory tracking error by up to 4.3% and completely eliminates power and thermal violations compared to conventional baselines. This unified framework establishes a foundation for real-time, safety-aware thruster management in marine robotics.

Aquaculture is vital in ensuring a sustainable protein supply for economic purposes. Timely identification of damage to net pens poses a notable challenge in the aquaculture environment. Detecting damaged net pens in underwater environments using remotely operated vehicles (ROVs) offers a safe and efficient solution, eliminating the need for human divers to face potential dangers. Underwater inspections with optical cameras face challenges due to reduced visibility caused by floating particles and light attenuation. So, selecting the best available robot view is vital for clear images for net inspection. To address this issue, we track the mean gradient feature across partial or entire images and actively control the robot’s pose to get the best available images. To simplify setting the desired set-point for distance control, we train a convolutional neural network (CNN) offline using supervised learning and integrate it with a Proportional Integral Derivative (PID) controller. This combined approach enables the ROV to maintain a consistent relative pose with respect to the fishnet, thereby obtaining clear net images in dynamic marine environments and effectively identifying any damage. Experimental results obtained in both the garden pool and the real fish farm environments validate the efficacy of the proposed method.

Inspection of aquaculture net pens is essential for maintaining the structural integrity, biosecurity, and operational efficiency of fish farming systems. Traditional inspection approaches rely on pre-programmed missions or manual control, offering limited adaptability to dynamic underwater conditions and user-specific demands. In this study, we propose AquaChat, a novel Remotely Operated Vehicle (ROV) framework that integrates Large Language Models (LLMs) for intelligent and adaptive net pen inspection. The system features a multi-layered architecture: (1) a high-level planning layer that interprets natural language user commands using an LLM to generate symbolic task plans; (2) a mid-level task manager that translates plans into ROV control sequences; and (3) a low-level motion control layer that executes navigation and inspection tasks with precision. Real-time feedback and event-triggered replanning enhance robustness in challenging aquaculture environments. The framework is validated through experiments in both simulated and controlled aquatic environments representative of aquaculture net pens. Results demonstrate improved task flexibility, inspection accuracy, and operational efficiency. AquaChat illustrates the potential of integrating language-based AI with marine robotics to enable intelligent, user-interactive inspection systems for sustainable aquaculture operations.