Underwater robots that use sonar navigation in turbid and visually degraded environments face several challenges. This includes sensor blackouts, visually similar rock formations, and misleading canyon geometries (geometry-induced ambiguity). While operating sonar at high resolution can help in improving navigation accuracy, it can, however, impact the computational and energy costs. This paper implements a predictive scheduling system that anticipates moments when the navigation estimator may become unstable. This approach is able to select the suitable sonar processing mode to match the current environmental condition and predict ahead, thereby balancing the trade-offs between sensor fidelity and resource consumption. The proposed approach monitors scan-matching performance and estimator uncertainty to predict impending navigation degradation and escalates sensing fidelity only when necessary. The framework is evaluated in three simulated environments with distinct perceptual characteristics: sparse, cluttered, and corridor-like environments, as well as on a real forward-looking sonar dataset (UATD). Results showed that in the challenging environments characterized by fluctuating geometries, the scheduler was able to reduce its reliance on continuous high-fidelity sensing while maintaining estimator stability. In the structured corridor environment, the scheduler was conservative in its action as it switched only a few times, reflecting sustained observability conditions. With the real sonar data, the scheduler chose the high-fidelity operation as it sensed persistent degradation. It is able to prioritize reliability over energy efficiency. The study shows that integrating uncertainty-aware sensing adaptation proves to be effective for managing sensing degradation while reducing reliance on continuous high-fidelity operation in environments where lower-cost modes remain sufficient.
Training technicians to diagnose and resolve hydroelectric plant faults requires exposure to multi-stage system issues that are difficult, expensive, and hazardous to recreate in real facilities. This work presents a scenario-driven virtual reality (VR) hydroelectric training environment built around a static facility layout and hand-based low-fidelity interactions. The system models turbine bays, reservoir areas, drain gates, intake zones, and mechanical components, and each scenario is instrumented with fine-grained telemetry capturing object-manipulation events such as debris removal, reservoir cleaning, component replacement, alignment attempts, and disposal accuracy. The environment presents learners with fault sequences in which symptoms (e.g., low flow, unstable pressure, abnormal vibration) evolve logically from underlying causes such as reservoir debris accumulation or mechanical wear. Each episode includes structured cues, root-cause relationships, and expected multi-step procedures, allowing the system to collect performance metrics such as task order, alignment offsets, cleanup completeness, safety adherence, and timing. In addition to supporting realistic practice, the platform enables repeatable, controlled experimentation with trainee behavior across varied fault conditions. Preliminary demonstrations using scripted interaction traces show that this telemetry can differentiate correct diagnostic pathways from common errors and support data-driven evaluation of trainee proficiency. Overall, this work introduces a fully instrumented VR platform for hydroelectric scenario training and establishes a foundation for future adaptive, personalized, or automated assessment systems.
Human–robot collaboration (HRC) is increasingly used in industrial environments to improve flexibility and productivity. As humans and robots operate in shared workspaces, safety remains a central requirement during system design and operation. An important challenge in the planning phase is the placement of safety sensors, which affects the ability to detect humans and prevent hazardous situations. In current industrial practice, sensor layout design is often based on manual planning and expert experience. This process requires significant effort and may lead to blind spots caused by occlusions or complex cell geometries. This paper presents a systematic approach for generating sensor layouts in human–robot collaboration applications. The proposed method analyzes the geometry of an industrial cell to identify occluded regions that limit human detection. A heatmap representation of potential human movement paths toward hazardous zones is used to describe the spatial distribution of risk within the workspace. Based on this information, suitable sensor positions are derived to improve coverage of critical areas while reducing redundant sensing. The approach is evaluated using several industrial use cases with different layouts and collaboration modes. The results show that the generated sensor layouts provide improved coverage of safety-relevant areas compared to manually designed configurations, while reducing planning effort. In addition, the method can be integrated into digital twin environments, allowing sensor layouts to be assessed and adapted during the design phase. The proposed approach supports the development of safer and more efficient human–robot collaboration systems.
This paper presents the design, implementation, and experimental validation of a novel underwater robotic quadrotor-float developed to address the challenges of remotely navigating, mapping, and sampling seafloor methane seeps. The proposed platform integrates the passive stability of a Lagrangian float with the active control authority of an underwater quadrotor, enabling precise depth tracking in dynamic underwater environments. The mechanical architecture adopts a modular design that combines a 3D-printed structural frame with a commercially available watertight enclosure to house onboard electronics, facilitating rapid prototyping and future upgrades. Actuation is achieved using four electric motor–propeller thrusters aligned along the heave axis, while depth feedback is provided by a high-resolution pressure sensor. A PID-based control strategy is implemented to regulate the vertical motion of the platform and achieve accurate depth tracking. The performance of the quadrotor-float is experimentally evaluated through a series of progressively complex tests conducted in a laboratory water tank, a deep-water swimming pool, and an outdoor natural water body. Experimental results demonstrate reliable tracking of both step and time-varying depth reference signals, with centimeter-level accuracy under nominal operating conditions and in the presence of environmental disturbances. These results validate the feasibility of the proposed quadrotor-float concept for near-seafloor operations and highlight its potential to improve maneuverability, reliability, and data or sample collection success in methane seep exploration. To the best of the authors’ knowledge, this work represents the first pool-scale experimental demonstration of closed-loop reference depth tracking using a robotic quadrotor-float in complex underwater scenarios.
Sea-level rise poses a growing threat to coastal structures, requiring building-level assessments that move beyond traditional 2D flood maps toward richer sensing and environment modeling. This study develops a 3D-mapping workflow to quantify building-level exposure to sea-level rise (SLR), using Old Dominion University (ODU) as a case study. Python, GeoPandas, rasterio, and Leafmap with a MapLibre backend are used for multisource sensor fusion, combining LiDAR-derived USGS DEM data, NASA GIBS satellite imagery, and OpenStreetMap footprints into a machine-readable 3D environment model. Building geometries are extruded by estimating ground elevations from the DEM and heights from available attributes. Exposure is evaluated by comparing building elevations to projected water levels for SLR scenarios between 1.5 m and 4.0 m. Results show an abrupt, nonlinear increase in flooded structures at higher SLR levels, revealing spatial clusters of vulnerability. These building-level risk maps provide actionable information for campus planners and local emergency managers to prioritize adaptation, retrofitting, and evacuation planning. The workflow shows how open-source tools and multi-sensor geospatial data can create a practical 3D sensing and sensor-fusion framework for assessing climate-resilient infrastructure. This framework can be expanded with real-time sensing. It can support autonomous navigation, robotic inspection, and environmental monitoring in flood-prone coastal areas. Such an approach enables more adaptive, cost-effective, and scalable solutions for infrastructure monitoring and disaster response.
Adverse environmental conditions such as fog, rain, or lens-based dust particles are still a major challenge in the development of robust detection systems. Robust camera-based object detection is essential for safe autonomous operation in agriculture, where missed detections of animals or obstacles can lead to severe damage. However, optical sensors experience substantial performance degradation under adverse environmental conditions and lens-proximal disturbances. This work introduces a disturbance-aware perception framework, the Multi-Expert Controller (MEC), that improves object detection robustness by classifying disturbance types and routing each image to a specialized object-detection expert. The MEC employs an EfficientNet-B4 top-level classifier to distinguish noiseless scenes, fog, rain, dust, and grass occlusions, complemented by a dust-intensity subclassifier. For each disturbance category, a dedicated YOLOv12 expert detector is trained on disturbance-specific samples generated using Python-based augmentation pipelines for dust, plant debris, rain and fog. In extensive test scenarios the MEC delivers substantially increased robustness relative to a model trained exclusively on clear images. It also outperforms an all-weather model trained on a same amount of data, while an all-weather model trained on a markedly larger and more diverse dataset achieves the highest absolute performance. These findings indicate that expert specialization governed by reliable disturbance classification provides an approach to improving the safety of autonomous machines.
Robot motion speed selection is a key decision in human–robot collaboration, particularly in settings where a robot repeatedly executes tasks alongside a human collaborator. In many existing collaborative systems, robot speed is fixed or adjusted using manually designed rules, without explicitly accounting for changes in the human’s cognitive state during task execution. In practice, human cognitive fatigue can significantly influence comfort, safety, and overall task effectiveness, motivating the need for adaptive robot behavior. This paper presents a cognitive fatigue–aware speed selection framework for human– robot collaboration, formulated as a discrete Markov Decision Process. A simplified kitchen helper scenario is considered, in which a robot fetches small objects, such as a box of pasta, from one table and delivers them to a seated human collaborator. The robot follows a fixed reference trajectory that is divided into discrete phases corresponding to task segments such as reaching, transporting, and delivering the object. Cognitive fatigue is modeled as a binary state variable, and at each trajectory phase, the robot selects between two execution speeds. Reinforcement learning is employed to learn a speed-selection policy that balances task efficiency with fatigue-aware caution through interaction with the environment. All experiments are conducted in the PyBullet simulation environment using tabular Q-learning. The learned policies exhibit interpretable and consistent behavior, selecting faster execution when the human is non-fatigued and slower execution when fatigue is present. The results demonstrate that even a minimal fatigue-aware decision model can produce meaningful and predictable adaptations in robot motion speed, highlighting the potential of reinforcement learning for incorporating human cognitive state into collaborative robot control.
Most multi-agent area coverage methods in the literature rely on one or more restrictive assumptions: A1) a fully connected communication network, A2) a fixed and known number of agents, or A3) idealized onboard navigation capabilities. To overcome these limitations, this paper proposes an adaptive coverage control algorithm for teams of autonomous robots operating without any networked inter-agent communication. Each robot employs Light Detection and Ranging (LiDAR) to scan the environment, generate a dense point cloud, and detect neighboring robots using the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) clustering algorithm. Using positions estimated relative to a global frame given by an onboard localization system, the environment is partitioned using Voronoi tessellation. The robots then compute the circumcenters of their respective Voronoi cells, which serve as the target position to achieve optimal configuration. Notably, the proposed framework does not assume prior knowledge of the number of robots in the environment, making it highly applicable to scenarios such as disaster response, where robots may be deployed incrementally. We then validate the robustness and scalability of the proposed approach with computer experiments conducted in a commercial robot simulator, CoppeliaSim, emulating realistic deployment conditions. The absence of communication among robots enhances system reliability and scalability, making it suitable for both structured and unstructured environments.
In this work, we develop a control framework for a custom 4-DOF rigid-body manipulator by coupling Pontryagin’s Minimum Principle, which generates trajectories, with physics-informed gradient descent to determine the boundary conditions. Optimal control provides a closed-form control law for the joint accelerations, while the gradient descent module determines the corresponding time horizon and terminal state by minimizing a cost functional built directly from the full rigid-body dynamics. This control framework yields dynamically feasible trajectories that can efficiently drive the manipulator through a sequence of waypoints. In addition, we use quasi-static force and torque analysis to generate a set of feasible joint states to ensure that our proposed optimizer remains within the admissible region. We further enhance our gradient descent solver with a pseudo-inertia matrix and momentum term to improve convergence. The resulting kinematic trajectories and dynamically feasible time horizons are transformed to torque inputs using an analytic model of the inverse dynamics, which we derive using the Euler–Lagrange method. This pipeline preserves a strict control-theoretic structure while embedding the physical constraints and loading behavior of the manipulator in a computationally efficient way. We demonstrate the performance of our gradient descent-based optimizer and the overall pipeline in physically realistic simulation.
This paper investigates energy–efficient trajectory generation for a planar mobile robot navigating among polygonal obstacles and demonstrates its execution on both a point–mass model and a realistic Quanser QCar platform. The planning problem is formulated as a trajectory-generation problem with fixed time allocation and energy-based cost metrics for a double–integrator system subject to collision–avoidance constraints, using three cost measures: geometric distance, rest–to–rest control energy, and continuous spline energy. On a visibility graph of the free space, we introduce a four–stage hierarchy: (i) distance–weighted Dijkstra, (ii) energy–weighted Dijkstra using closed–form edge costs, (iii) geometric smoothing via arc–length splines, and (iv) time–parameterized cubic splines that minimize energy. We adopt Stage 4 as the final energy-efficient spline-based trajectory approximation. The resulting trajectories nearly preserve the shortest–path length (e.g., 11.65 vs. 11.84 units) while reducing energy consumption by more than an order of magnitude (from 12.73 to 1.01). In a cluttered map, Stage 4 also outperforms a Rapidly–exploring Random Tree (RRT*) baseline (with 50k iterations), yielding smoother profiles, lower energy, and shorter completion time. Finally, by inflating obstacles to account for the QCar dimensions, we obtain collision–free splines that track well under unicycle dynamics, with maximum errors of 4.7 × 10−2, respectively. Overall, the proposed pipeline integrates graph search, analytical energy models, and spline smoothing into a simple and hardware–feasible trajectory–generation framework.
Understanding how automated vehicles (AVs) influence human driving behaviors in mixed traffic environments is critical for enhancing transportation safety. Utilizing high-resolution trajectory data from the U.S. Department of Transportation's Third Generation Simulation (TGSIM) dataset, this study developed a multi-agent deep deterministic policy gradient (MA-DDPG) framework for modeling safety-critical car-following dynamics between AVs and human-driven vehicles (HDVs). The MA-DDPG framework, designed to learn sequential and interactive decision-making over continuous action spaces, effectively captured dynamic AV-HDV interactions in three- vehicle configurations. Compared to single-agent reinforcement learning and supervised learning approaches, it demonstrated superior performance in simulating realistic evasive behaviors, achieving velocity RMSE of 0.847 m/s for AVs and 1.212 m/s for HDVs. Trajectory reconstruction validated the framework's capability to replicate collision avoidance mechanisms. When time-to-collision dropped to 3.2 seconds, the model accurately reproduced the AV's deceleration response and coordinated speed adjustments across all vehicles, with strong alignment between predicted and observed trajectories. Counterfactual analysis revealed that drivers trained following AVs adopted significantly more conservative patterns when encountering human-driven lead vehicles (Kolmogorov-Smirnov test: p < 0.1). Mean speeds decreased from 5.7 m/s to 4.8 m/s, acceleration shifted from positive to negative, and braking frequency increased. This study develops multi-agent simulations of AV-HDV interactions that can inform collision avoidance strategies in autonomous and robotic systems.
The growing sophistication of image forgeries from conventional tampering techniques, such as splicing and copy-move, to AI-generated deepfakes poses a serious threat to digital trust, national security, and public safety. Active development of unified approaches that can jointly address traditional image forgeries and deepfakes remains less explored. Existing methods often focus on a single category, thereby compromising generalization across diverse forgery types. This highlights a need to develop robust unified frameworks capable of reliable detection of both conventional and deep forgeries. To combat these challenges, we propose a hybrid detection framework that incorporates local feature descriptors (Local phase quantization and Zernike moments) with deep semantic features to capture both fine-grained texture inconsistencies and global image context effectively. For deep features, we proposed a Forgery-Aware ConvNeXt (FA-ConvNeXt) by introducing Mish Activation and a lightweight Channel Attention Squeeze-and-Excitation in baseline ConvNeXt-T backbone. A vision transformer fusion layer is employed to further enhance the interaction of these local and deep features, followed by a lightweight classification head for final prediction. The proposed framework was tested on three standard and diverse datasets, namely CoMoFod, CASIA v2, and Faceforensics++. The experimental results on these datasets show that our method achieves an average accuracy of 96.83% and AUC of 0.95, outperforming state-of-the-art baselines. These results confirm the effectiveness of the suggested framework in tackling both traditional and synthetic forgeries in real-world applications.
Industrial training for hydroelectric maintenance requires learners to interpret the environment around them in real-time to make decisions in order to apply safety, regulations, and effective intervention under pressure. Traditional training approaches in a virtual reality environment fail to capitalize on the user’s own performance by not adapting to their training skills. This work presents an adaptive virtual reality environment that utilizes a reinforcement learning (RL) feedback agent to deliver context-specific prompts during maintenance scenarios. Of the studies that choose to evolve their static environment to an adaptive one, there is little work that systematically isolates design choices such as user skill tiers. It works by extracting state information from user interactions within a detailed hydroelectric plant simulation and uses a lightweight policy model to determine when and how feedback should be presented to the user. Prompt quality is then assessed using programmatic metrics for clarity, specificity, safety, actionability, and domain accuracy, as well as an overall scalar representative of both task performance and text quality. We compare three feedback prompts under the following circumstances: (1) adaptive RL prompts, (2) random prompts, (3) static, hand-crafted prompts, and run a set of ablation studies that disable axis adaption, tier adaption, and individual reward components. Across simulated trials, the fully adaptive Proximal Policy Optimization (PPO) configuration achieves higher overall reward and domain targeting.
Advanced driver assistance systems (ADAS) require controlled and repeatable testing with vulnerable road users, but current robotic targets depend on GNSS or indoor positioning infrastructure, restricting experiments to instrumented proving grounds and preventing realistic indoor evaluation. This paper introduces a vision-based autonomous mobile robot (AMoR) that performs repeatable ADAS scenario execution in GPS-denied environments using only onboard perception. The robot integrates multi-camera visual SLAM, volumetric mapping, and model-predictive navigation on an embedded Jetson AGX Orin to build reusable maps and navigate in real time. The system is validated in a multi-story parking garage through Euro NCAP-inspired child pedestrian scenarios executed with a prototype test vehicle. Results show that the AMoR achieves real-time operation and accurate trajectories while executing complex vehicle-to-pedestrian interactions. The proposed platform demonstrates, to our knowledge, the first vision-only, perception-driven ADAS testing robot and establishes a scalable foundation for future infrastructure-free safety validation.
In this study, we propose a methodology that integrates domain knowledge with data-driven techniques to improve anomaly detection in autonomous Cyber-Physical Systems (CPS). The approach yields a unified pipeline designed to reinforce the security and reliability of real-world CPS deployments. A central contribution of the method is the efficient integration of domain-informed multimodal fusion with variational autoencoder (VAE)-based latent modeling of lidar-derived features combined with wheel-velocity information for identifying adversarial activity in autonomous CPS. This integration helps preserve informative multimodal representations across changing environments and attack conditions. By incorporating spatial and temporal patterns from multimodal data, such as range-based lidar observations and wheel velocities, the model learns representations that remain discriminative and robust to environmental variation. Experimental evaluations on a Clearpath Husky robot using the Army Research Laboratory (ARL) Phoenix autonomy stack, including attacks carried out through Robot Operating System (ROS)-based middleware, show that the proposed pipeline achieves above 99% accuracy in several evaluated settings while maintaining strong performance across the evaluated attack types, including data flooding, spoofing, and man-in-the-middle attacks. The results highlight the effectiveness of the proposed multimodal feature learning strategy in improving the security and reliability of anomaly detection systems for autonomous CPS operating under dynamic, adversarial, and often unpredictable real-world conditions.
Modeling collision structure in configuration space is essential for efficient manipulator planning, where sharp free–obstacle boundaries and narrow passages make geometric queries a major computational cost. This paper presents a unified comparison of three surrogate-modeling approaches for continuous goodness fields: Fast Fourier Transform (FFT) smoothing, Gaussian Process (GP) regression, and Radial Basis Function (RBF) interpolation. Using identical sampling budgets in a cluttered 2-DOF environment, we evaluate surrogate fidelity, boundary accuracy, sample efficiency, and downstream performance in A*, PRM*, RRT*, and Informed RRT*. GP surrogates achieve the highest F1, IoU, and lowest RMSE scores, providing sharply resolved boundaries and enabling consistently high planning success rates. RBF surrogates track GP performance but exhibit reduced confidence in free space, while FFT-based fields blur discontinuities and degrade planner reliability. We additionally study several goodness-guided refinement strategies and find that, in low-dimensional spaces with well-distributed initial samples, refinement yields negligible benefit compared to the choice of surrogate model. Overall, the results show that regression-based surrogates provide the fidelity required for robust surrogate-guided planning and motivate extensions to higher-dimensional manipulators and scalable hybrid models.
Photochemical experiments commonly rely on manual positioning of samples and manual control of light exposure, resulting in labor-intensive workflows, limited repeatability, and reduced experimental throughput. To address these limitations, this paper presents AutoPCRS, an Automated Photochemical Reaction System that integrates precise linear actuation with microcontroller-based timing and web-enabled control. The system employs an ESP32-S3 microcontroller, a stepper motor–driven ball screw linear actuator, and an IoT-controlled light source to automate test tube positioning and exposure timing. A web-based interface allows users to configure exposure durations and initiate batch experiments remotely via Wi-Fi. In addition, the platform is designed for batch-based photochemical irradiation, supporting discrete exposure control for individual samples using standard laboratory test tubes. An inductive proximity sensor enables repeatable homing and accurate alignment of samples with the fixed light source. Experimental validation demonstrates reliable positioning, consistent exposure timing, and fully unattended batch operation across multiple samples. The system prototype supports a rack-based configuration accommodating up to ten test tubes, allowing sequential irradiation within a single experimental run. By reducing manual intervention and standardizing experimental conditions, AutoPCRS improves repeatability and operational efficiency in photochemical workflows while reducing researcher workload. Overall, the proposed system provides a low-cost platform for laboratory automation in research applications.
Recent advances in large language models (LLMs) have enabled software developers to create functional applications from a single prompt. Analogous capabilities, however, for building physical electronic circuits remain elusive. Here, we introduce Impromptu, an end-to-end, prompt-to-hardware circuit prototyping platform that automates the design, verification, and assembly of electronic circuits. Impromptu combines an LLM-based circuit synthesis module with a simulation-driven validation pipeline based on industry-standard SPICE tools, and a Cartesian gantry equipped with a pick-and-place toolhead. The system models the breadboard and component library as a discretized set of capsules and uses a placement algorithm to select electrically valid layouts in this discrete space. We evaluate Impromptu on benchmark circuits, including LED–resistor networks, voltage dividers, RLC circuits, and a 555 astable timer, and compare a lightweight and a higher-capacity LLM. Our results demonstrate that prompt-to-circuit prototyping is feasible, laying the groundwork for more complex, fully automated end-to-end hardware design and assembly systems that could not only accelerate circuit-design learning in academic settings but also streamline prototyping workflows in professional environments.
Robotic surface polishing remains difficult to deploy in industrial settings due to the gap between process expertise and robot-specific programming skills. While intuitive teaching approaches such as hand-guiding reduce interaction complexity, they typically fail to encode essential process knowledge, leading to trajectories that are inefficient or poorly aligned with surface-processing requirements. This paper presents a process-aligned, no-code teaching framework for robotic surface polishing that integrates sensor-based workpiece reconstruction, explicit process modeling, and coverage-aware motion planning. Users specify polishing intent by selecting a two-dimensional region of interest (ROI), which is automatically mapped onto the reconstructed three-dimensional workpiece geometry. Based on a tool–workpiece interaction model, the system synthesizes normal-aligned polishing trajectories by jointly optimizing traverse length and surface coverage uniformity. Candidate trajectories are evaluated using a weighted multi-objective cost function that balances efficiency and coverage quality, enabling stable tool–surface interaction without requiring explicit force control during execution. The proposed framework is validated through a comparative user study against kinesthetic hand-guiding on a free-form workpiece. Experimental results demonstrate a reduction in programming time by approximately one order of magnitude, while producing smoother robot motions and significantly more consistent contact-force distributions. These findings indicate that the observed performance gains stem not merely from simplified interaction, but from embedding process constraints directly into the motion planning stage. By combining no-code interaction with sensor-based adaptation to geometric variability, the proposed approach lowers deployment barriers and supports scalable robotic surface processing in high-mix, low-volume manufacturing environments.
The deployment of autonomous robots in harsh and unpredictable environments has gained significant attention in recent years, driven by the need to improve safety and efficiency in tasks where human involvement poses significant risks. Technological advances in sensors and communication technologies, as well as the proliferation of AI, have catalysed the development of highly capable and robust robotic systems in diverse scenarios, optimising efficiency, particularly in the area of disaster response. However, despite the progress made in the field of autonomous Search and Rescue (SAR) operations in harsh environments, several challenges remain unresolved. This paper presents a comprehensive overview of the key technological advancements in components, algorithms and communication systems and how they facilitate highly efficient and robust autonomous missions in extreme subterranean conditions and SAR operations. Additionally, this paper aims to tackle some of the key challenges encountered during emergency response operations through a novel autonomous robotic design that leverages locomotion, perception, communication, and control capabilities to achieve effective exploration tasks in harsh and hazardous environments. Finally, it sets the scene for future enhancements in the area of autonomous and intelligent robots, suggesting potential avenues for further research and development in enhancing their capabilities and adaptability.
