Abstract Background Instrument-assisted soft tissue manipulation (IASTM) is widely used, yet clinicians mostly rely on their subjective perception of applied force which can lead to variability. No studies have determined intra- and inter-examiner reliability of three-dimensional (3D) IASTM force–motions applied to humans. The objective of this study was to evaluate whether real-time visual monitoring enhances consistency of applied IASTM force. Methods This study was an observational reliability study with blinded assessors. 45 healthy adults were enrolled between June to December 2021. Clinicians (two novice, two experienced) applied 1-inch IASTM linear strokes using two quantifiable soft tissue manipulation (QSTM) smart devices (localized; dispersive) to lumbar and calf regions under two conditions: (1) applying self-perceived “medium” force without visual monitoring, and (2) applying force guided by real-time visual monitoring from a graphic visual interface (GVI). Triaxial (3D) average peak force (primary variable), stroke frequency, and angle were measured. Linear mixed models and variance components evaluated repeatability and reproducibility within and between clinicians and across two sessions. Results Visual monitoring substantially reduced variability in average peak force across clinicians, regions, devices, and sessions. Standard deviations (SDs) were significantly larger and ranges meaningfully broader without monitoring compared to with monitoring. Monitoring improved intra- and inter-examiner consistency by ≥ 30% in nearly all conditions, with medium to very large effect sizes. Effects on stroke frequency were mixed, and influence on angle minimal. Conclusions Visual monitoring in real-time significantly improved the consistency of 3D IASTM force applications on humans. Optimal reliability is foundational to practice fidelity, training, and more rigorous investigation of dose–force response relationships in manual therapy. Trial Registration : The study was prospectively registered at ClinicalTrials.gov (Identifier: NCT04923633).

Understanding the 3D occupancy of surrounding obstacles and occupants is a critical perception task for autonomous navigation, and is particularly challenging in off-road environments where the absence of structured road geometry demands robust, sensor-efficient solutions. While conventional methods predominantly optimize for mean Average Precision (mAP) or NuScenes Detection Score (NDS), recall-aware architectures that prioritise foreground object coverage at real-time speeds remain underexplored. In this paper, we present YOLO-SemBEV, a semantic-guided Bird's-Eye-View (BEV) occupancy detector that operates from a single front-facing camera and a single front-facing radar. The architecture combines a detection-optimized YOLOv8 backbone with a semantically painted radar feature encoder and sparse semantic proposals that provide interpretable 3D spatial priors for BEV lifting. This design eliminates the multi-view synchronization overhead of surround-view systems while preserving strong foreground recall. Evaluated on the nuScenes validation set, YOLO-SemBEV achieves a foreground F1 of 0.684 with a Foreground Recall score of 0.779, with an end-to-end latency of 21 ms (47+ FPS), surpassing the LiDAR-assisted Minkowski Engine in recall by 23% and at least 1.9x speedup over both Minkowski Engine and CRN. The code is available at github.com/11rahi23/YOLO-SemBEV

The United States leads in corn production and consumption in the world with an estimated USD 50 billion per year. There is a pressing need for the development of novel and efficient techniques aimed at enhancing the identification and eradication of weeds in a manner that is both environmentally sustainable and economically advantageous. Weed classification for autonomous agricultural robots is a challenging task for a single-camera-based system due to noise, vibration, and occlusion. To address this issue, we present a multi-camera-based system with decision-level sensor fusion to improve the limitations of a single-camera-based system in this paper. This study involves the utilization of a convolutional neural network (CNN) that was pre-trained on the ImageNet dataset. The CNN subsequently underwent re-training using a limited weed dataset to facilitate the classification of three distinct weed species: Xanthium strumarium (Common Cocklebur), Amaranthus retroflexus (Redroot Pigweed), and Ambrosia trifida (Giant Ragweed). These weed species are frequently encountered within corn fields. The test results showed that the re-trained VGG16 with a transfer-learning-based classifier exhibited acceptable accuracy (99% training, 97% validation, 94% testing accuracy) and inference time for weed classification from the video feed was suitable for real-time implementation. But the accuracy of CNN-based classification from video feed from a single camera was found to deteriorate due to noise, vibration, and partial occlusion of weeds. Test results from a single-camera video feed show that weed classification accuracy is not always accurate for the spray system of an agricultural robot (AgBot). To improve the accuracy of the weed classification system and to overcome the shortcomings of single-sensor-based classification from CNN, an improved Dempster–Shafer (DS)-based decision-level multi-sensor fusion algorithm was developed and implemented. The proposed algorithm offers improvement on the CNN-based weed classification when the weed is partially occluded. This algorithm can also detect if a sensor is faulty within an array of sensors and improves the overall classification accuracy by penalizing the evidence from a faulty sensor. Overall, the proposed fusion algorithm showed robust results in challenging scenarios, overcoming the limitations of a single-sensor-based system.

Electrochemical models of Lithium-Ion batteries are physics-based models that account for the kinetic and transport phenomena. One such model is the pseudo-two-dimensional (P2D) model by Newman. In this study, we utilized electrochemical modeling capabilities of the computational fluid dynamics (CFD) tool FLUENT for lithium-ion batteries. We developed an LFP (Lithium Iron Phosphate) battery electrochemical model within Ansys FLUENT that is compatible with the internal MSMD solution method. A 14.6 Ah LFP pouch cell was modeled and after meshing, various cases were simulated for the cell, including different C-rates at single cycles. We also updated the model utilizing the cell aging mechanisms or degradation models that include solid-electrolyte interphase (SEI) layer growth and irreversible lithium plating to study cell degradation over long usage. Simulation of the developed P2D model was performed at 0.5, 1, and 2 C rates and the thermal and electric characteristics of the cell were recorded. This model was then simulated with the addition of the degrading side reactions. This study showed that CFD modeling techniques could be successfully applied to model a Li-Ion battery cell to obtain degradation data. Metrics for state of charge, cell voltage, cell temperature, and lithium concentrations for different degradation mechanisms were shown to be consistent.

Lane-keeping systems are a major part of advanced driver assistance systems (ADAS). Existing lane detection algorithms are based on either Computer Vision (CV) models or deep learning techniques which are often vulnerable to unfamiliar routes, lane marking conditions, night-time driving, weather conditions, etc. To improve lane detection accuracy under various challenging conditions, we propose a framework that utilizes several lane detection models with different features to obtain a robust algorithm. The proposed Multiple Model Adaptive Estimation (MMAE) algorithm works with two cameras, one front camera and one rear camera. The front camera is used for lane offset estimates whereas the rear camera serves as a time-delayed reference for the estimated lane offsets. The offsets from front camera CV models (two) are used as inputs to the MMAE algorithm which compares the offset computed by the rear camera CV model (time-delayed) as the reference. The proposed MMAE algorithm then estimates the probability of lane offsets to match the time-delayed reference model lane offset and selects the offset with higher probability of matching with reference model. The offset from the time-delayed reference model cannot be used for the real-time lane keeping control system since it would produce erroneous steering output due to the time lag in offset estimated by the real camera model. Thus, the MMAE estimated offset offers a more accurate lane offset and hence used in a PID steering controller for the lane keeping system. The proposed algorithm is then deployed in an AirSim simulation environment for performance evaluation. The simulation results show that the proposed MMAE algorithm performed robustly even when one of the models performed poorly. The proposed lane detection algorithm was able to identify the poorly performing model and switch to the other model to ensure better lane detection performance.

Self-driving cars are dynamically complex systems whose behavior is highly sensitive to operating and environmental conditions. At racing speeds, aerodynamic effects, tire slip and grip, and heat accumulation significantly alter dynamics, making control a challenging task under high model uncertainty and limited knowledge of vehicle behavior. This paper addresses this challenge using the Active Disturbance Rejection Control (ADRC) framework. A path-planning strategy based on camera perception generates reference signals for two parallel ADRCs: one regulating forward velocity and the other controlling ego-vehicle heading. The controller was developed for the Indy Autonomous Challenge (IAC) on the Dallara AV-21 platform and validated using the official IAC simulator, Ansys VRXperience, which replicates both vehicle dynamics and sensor suites. The proposed approach leverages the disturbance estimation and rejection capabilities of ADRC and its Extended-State Observer (ESO), requiring only minimal model information. Testing on the simulator shows that the controller achieves consistent performance at speeds up to 70 m/s (157 mph) on the Indianapolis Motor Speedway, demonstrating its robustness for realistic high-speed autonomous racing scenarios.

While YOLO’s efficiency and accuracy have made it a popular choice for object detection and tracking in real-world applications, models trained on smaller datasets often suffer from intermittent detection failures, where objects remain undetected across multiple consecutive frames, significantly degrading tracking performance in practical scenarios. To address this challenge, we propose PrED (Predictive Enhancement of Detection), a novel framework that enhances object detection and aids in tracking by integrating low-confidence detections with multiple similarity metrics—including Intersection over Union (IoU), spatial distance similarity, and template similarity, and predicts the locations of undetected objects based on a parameter called predictability index. By maintaining object continuity during missed detections, PrED ensures robust tracking performance even when the underlying detection model experiences failures. Extensive evaluations across multiple benchmark datasets demonstrate PrED’s superior performance, achieving over 11% higher DetA with at least 6.9% MOTA improvement in our test scenarios, 17% higher detection accuracy (DetA) and 12.3% higher Multiple Object Tracking Accuracy (MOTA) on the KITTI training dataset, 8% higher DetA and 2.6% higher MOTA on the MOT17 training dataset, compared to ByteTrack, establishing PrED as an effective solution for enhancing tracking robustness in scenarios with suboptimal detection performance.

Autonomous navigation in off-road environments presents significant challenges due to the diverse and unpredictable characteristics of natural terrains. Due to the class imbalances prevalent in the existing datasets, the current models exhibit difficulties adapting to different environmental conditions. In this paper, we present an approach to address these challenges through the development of a deep semantic segmentation model tailored specifically to aid in navigating diverse off-road scenarios. Our methodology consists of two primary components: dataset development and segmentation model construction. The dataset comprises 1,414 images derived from the Yamaha-CMU Off-Road dataset (YCOR) through accuracy enhancements and augmentation techniques. Subsequently, a segmentation model is developed employing an encoder-decoder architecture using Resnet34 as the feature extractor and U-Net as the decoder. The proposed model demonstrates a notably high segmentation accuracy and attains a micro- $F_{1}$ score of 90% or higher on benchmark datasets such as YCOR, Rellis-3D, and RUGD with minimal or no transfer learning, highlighting its versatility and adaptability across various environmental settings. The model also exhibits a per-frame inference time of 40ms, rendering it feasible for real-time application.

This paper presents the preliminary results of our investigation in detecting negative road objects such as potholes and cracks in real-time for autonomous vehicles in order to improve safety. It is important to not only detect the existence of the negative road object from a safe distance, but also to identify the size, depth, and shape of such objects which plays a critical role in determining if the object must be avoided for safety or it can be ignored. We utilized open-source image databases such as ROBOFLOW (pothole) and RDD2020 of negative road objects for training a single layer convolutional neural network (CNN), such as YOLO. To test the trained algorithm, a small data set taken on the city streets in Indianapolis, IN was used in order to evaluate the accuracy of detection of the negative road objects. While the initial results show the ability of detecting the negative road objects through the proposed algorithm, more work needs to be done to improve the overall accuracy of detection with respect to size and shape of these negative road objects.

The Energy Management Strategy (EMS) of PHEV controls the motor power & engine power to reduce overall energy consumption and manage the battery state of charge (SOC). EMS strategies for PHEV are typically divided into two categories rule-based and optimization based. Rule-based EMS are easy to implement and simple to understand. However, their application is limited to specific driving cycles which makes it difficult to achieve the optimal fuel economy. Optimization based on the other hand requires a system model to compute the optimal solution with given constraints. Here we propose a value-based reinforcement learning approach called Deep Q learning-based (DQN) strategy to determine the optimal action to maximize the fuel economy, without the need to have a system model. A DQN algorithm is trained on a UDDS cycle using a PHEV simulation model. The agent interacts with the PHEV environment during training and stores the transactions s(t), a(t), r(t), s(t+1), then random mini-batches are sampled to calculate the state-action values Q(s, a;. theta(i)), and the target network is used to generate target Q value. These two are then used to compute loss function to update neural network weights. Fuel economy and convergence characteristics of DQN are discussed and compared with rulebased EMS and performance is analyzed on unseen HWFET cycle.

Lithium-Ion batteries have become the principal battery technology for EVs to date. However, one of the principal factors limiting the widespread usage of the EVs is the length of charging times for the lithium-ion battery packs. The appropriate charging algorithm is critical to shorten the battery charging times while keeping the battery safe. In our earlier work, we proposed a novel optimal strategy for charging the lithium-ion battery based on electrochemical battery model using A performance index that aimed at achieving a faster charging rate while maintaining safe limits for various battery parameters. A more realistic model, based on battery electro-chemistry has been used for the design of the optimal charging algorithm as opposed to the conventional equivalent circuit models. Simulation results showed that the proposed optimal charging algorithm is capable of shortening the charging time of a lithium-ion cell by as much as 30% when compared with the standard constant current charging. Here we present the results from a number of experiments using Lithium-Ion cylindrical cells that were charged using the proposed algorithm and compared the charging times with the standard constant current-constant voltage (CC-CV) charging algorithms. A Maccor Series 4300 battery testing system was used to carry out the experiments. The experimental results showed that the proposed algorithm offered shorter charging times by up to 16% when compared to the CC-CV charging algorithms under the same battery initial conditions such as SOC and temperature of the cells.

Rigorous testing of automotive sensors for accuracy and precision is essential before finalizing designs and proceeding with mass production. This testing process plays a crucial role in identifying these sensors’ potential shortcomings, which are essential components in the Advanced Driver Assistance Systems (ADAS). A virtual environment for sensor model simulation and testing provides a significant advantage in experimenting with the sensor in a vast array of realistic scenarios, which may be cost-prohibitive in physical testing. Thus, radar sensors, a commonly used sensor in ADAS, have been modeled using various methodologies in virtual environments. However, there is a need for a radar model that is both computationally efficient and accurate under different combinations of radar reflectivity, directivity, and signal-to-noise ratio (SNR). In this study, we propose a model of a Delphi-ESR 2.5 medium-range automotive radar in the Unreal Engine environment that leverages grid-based DBSCAN for point-cloud clustering and radar equations to calculate the SNR of grid points, facilitating object detection and tracking. The proposed model incorporates the reflectivity and directivity concept of Radar Cross-Section (RCS) to emulate the behavior of an actual radar. The sensor model exhibits superior accuracy and low computational requirement, with a mean error rate of 10.8% and an output rate of 8.7 Hz. This approach provides a reliable means of simulating and testing automotive radar sensors in a virtual environment, contributing to the further advancements of the ADAS systems.

The goal of achieving autonomous navigation for agricultural robots poses significant challenges, mostly arising from the substantial natural variations in crop row images as a result of weather conditions and the growth stages of crops. The processing of the detection algorithm also must be significantly low for real-time applications. In order to address the aforementioned requirements, we propose a crop row detection algorithm that has the following features: Firstly, a projective transformation is applied to transform the camera view and a color-based segmentation is employed to distinguish crop and weed from the background. Secondly, a clustering algorithm is used to differentiate between the crop and weed pixels. Lastly, a robust line-fitting approach is implemented to detect crop rows. The proposed algorithm is evaluated throughout a diverse range of scenarios, and its efficacy is assessed in comparison to four distinct existing solutions. The algorithm achieves an overall intersection over union (IOU) of 0.73 and exhibits robustness in challenging scenarios with high weed growth. The experiments conducted on real-time video featuring challenging scenarios show that our proposed algorithm exhibits a detection accuracy of over 90% and is a viable option for real-time implementation. With the high accuracy and low inference time, the proposed methodology offers a viable solution for autonomous navigation of agricultural robots in a crop field without damaging the crop and thus can serve as a foundation for future research.

Abstract Surgical Haptics is an emergent field of research to integrate and advance the sense of robotic touch in laparoscopic tools in robot-assisted minimally invasive surgery. Haptic feedback from the tooltip and soft tissue surface interaction during robotic palpation can be leveraged to detect the texture and contour of subsurface geometry. However, precise force modulation of the robotic palpating probe is necessary to determine stiff inclusions of the anatomy and maneuver successive manipulation tasks during surgery. This paper focuses on investigating the layered deformations associated with different force profiles involved in manipulating the superficial anatomy of soft tissues during dynamic robotic palpation to determine the underlying anomaly. A realistic three-dimensional (3D) cross-sectional soft tissue phantom with anatomical layers and tumor, as an anomaly, is designed, modeled, and analyzed to examine the effects of oriented palpating forces (0–5 N) of a 7 DOF robot arm equipped with a contoured palpation probe. Finite element static structural analysis of oriented robotic palpation on the developed 3D soft tissue phantoms (with and without anomaly) reveals the soft tissue layer deformations and associated strains needed to identify presence of stiffer inclusions or anomaly during Robotic palpation. The finite element analysis study shows that the difference in deformations of soft tissue layers (e.g., underlying myofascial layers) under stiffer inclusions at different force levels can facilitate haptic feedback to acquire information about subsurface tumors. The deformation variations are further compared to assess better palpation orientations for subsurface anomaly detection.

As autonomous vehicle technologies continue to advance in the automotive industry, sensors such as Infra-Red (IR) cameras are becoming increasingly crucial during night-time and low visibility scenarios. However, developing and validating object detection and recognition algorithms with these sensors can be expensive and time-consuming. Therefore, simulation of autonomous vehicles with appropriate sensor models provides an efficient and cost-effective way of testing and validating such algorithms. The use of virtual environments for developing and validating sensor models offers experimentation with various scenarios that may be cost-prohibitive or impossible to recreate in a physical environment including different weather conditions, lighting conditions, and other environmental factors that may affect the performance of the sensor. In this paper, we present a novel IR sensor model that approximates the output of an actual IR camera. The model utilizes Unreal Engine’s Custom Stencil and Post-Processing Materials to assign greyscale color values based on the object’s temperature property and an assigned temperature range for the camera. The IR sensor model outputs are compared with the outputs of an actual IR camera under similar conditions. The results show that the proposed IR camera model captures the basic features of an IR camera very well. However, further improvements to this innovative IR camera model is needed to capture additional features of an IR camera such as the rejection of non-thermal IR reflections, creation of a reflection buffer for real-time implementations, etc.

The advent of mobile ubiquitous computing enabled sensor informatics of human movements to be used in modeling and building deep learning classifiers for cognitive AI. Expanding deep learning approaches for classifying instrumented hand manipulation tasks, especially the art of manual therapy and soft tissue manipulation, can potentially augment practitioner’s performance and enhance fidelity with computer assisted guidelines. This paper introduces a dataset of 3D force profiles and manipulation motion sequences of controlled soft tissue manipulation stroke pattern applications in thoracolumbar, upper thigh and calf regions of a single human subject performed by five experienced manual therapists. The multimodal 3D force, 3D accelerometer and resultant gyro raw data were preprocessed and experimentally fed into a multilayered Long Short-Term Memory (LSTM) based Recurrent Neural Network (RNN) deep learning model to observe sequence classifications of two manipulation motion techniques (Linear "Strumming" motion and curvilinear "J-Stroke" arched motion) of manual therapy performed using a handheld, localizing Quantifiable Soft Tissue Manipulation (QSTM) medical tool. Each of these motion sequences were further labeled with corresponding best practice technique from validated video tapes and reclassified into "Correct" and "Incorrect" practice based on defined criteria. The deep learning model resulted in 90-95% classification accuracy for individual intra-therapist reduced dataset. The classification accuracy varied between 78%-93% range, when trained with multivariate characteristic feature set combinations for the complete spectrum of inter-therapist dataset.Clinical Relevance — AI informed online therapeutic guidelines can be leveraged to minimize practice inconsistencies, optimize educational training of therapy using data informed protocols, and study progression of pain and healing towards advancing manual therapy.

The efficiency of hybrid electric powertrains is heavily dependent on energy and power management strategies, which are sensitive to the dynamics of the powertrain components that they use. In this study, a Modified Particle Swarm Optimization (Modified PSO) methodology, which incorporates novel concepts such as the Vector Particle concept and the Seeded Particle concept, has been developed to minimize the fuel consumption and NOx emissions for an extended-range electric vehicle (EREV). An optimization problem is formulated such that the battery state of charge (SOC) trajectory over the entire driving cycle, a vector of size 50, is to be optimized via a control lever consisting of 50 engine/generator speed points spread over the same 2 h cycle. Thus, the vector particle consisted of the battery SOC trajectory, having 50 elements, and 50 engine/generator speed points, resulting in a 100-D optimization problem. To improve the convergence of the vector particle PSO, the concept of seeding the vector particles was introduced. Additionally, further improvements were accomplished by adapting the Time-Varying Acceleration Coefficients (TVAC) PSO and Frankenstein's PSO features to the vector particles. The MATLAB/SIMULINK platform was used to validate the developed commercial vehicle hybrid powertrain model against a similar ADVISOR powertrain model using a standard rule-based PMS algorithm. The validated model was then used for the simulation of the developed, modified PSO algorithms through a multi-objective optimization strategy using a weighted sum fitness function. Simulation results show that a fuel consumption reduction of 12% and a NOx emission reduction of 35% were achieved individually by deploying the developed algorithms. When the multi-objective optimization was applied, a simultaneous reduction of 9.4% fuel consumption and 7.9% NOx emission was achieved when compared to the baseline model with the rule-based PMS algorithm.

The objective of this paper is to formulate and analyze the benefits of a predictive non-linear multi objective optimization method for a platoon of mild-hybrid line haul trucks. In this study a group of three trucks with hybrid electric powertrain are considered in a platoon formation where each truck has a predictive optimal control to save fuel with out any loss of trip time. While the controller on each truck uses the look ahead knowledge of the entire route in terms of road grade, the overall platoon controller used a multi agent method (Metropolis algorithm) to define coordination between the trucks. While the individual trucks, showed significant improvement in fuel economy when running on predictive mode, the true savings came from the entire platoon and showed promising results in terms of absolute fuel economy without trading off on total trip time. The proposed algorithm also proved to be significantly emission efficient. A platoon of 3 trucks achieved an average of 10% fuel savings while cutting back 13% on engine out NOx emissions for engine off coasting and 9.3% fuel saving with 8% emissions reduction for engine idle coast configuration when compared to non-predictive non-platoon configuration.

Accurately testing sensors used in vehicles prior to physical production is crucial for identifying potential shortcomings and configuring effective vehicle driver assistance systems and automations. The use of virtual environments for sensor simulation and testing offers extensive freedom and allows for experimentation with scenarios that are cost-prohibitive for physical testing. In this study, we developed a sensor model using the Unreal Engine to approximate the output of an actual Delphi ESR 2.5 medium-range radar. We utilized the LineBatcher function to create point clouds that simulate radar reflections and employed 3D grid based DBSCAN for object detection. Subsequently, we implemented centroid tracking to track objects and determine range rate, acceleration, and other outputs of the Delphi ESR Radar. In our test environment, the sensor performed close to identical to a Physical Delphi ESR 2.5 Radar with accuracy of 91% and at a data rate of 10-20 Hz.

Crop row detection is a vital task in precision agriculture. Earlier works for solving this task follow traditional computer vision based methodologies. However, in recent years deep learning based approaches are being adopted for solving this task. Among various deep learning methodologies, semantic segmentation has found to be most successful for obtaining meaningful representation of images in a plethora of domains, such as, medical image analysis and autonomous driving. Scene parsing is a subcategory of semantic segmentation where all objects of interest in a scene are color coded as a way to simultaneously classify and localize their presence. In this way, scene parsing technique is a very good fit for solving crop row detection task; However, no existing research has yet ventured this direction. In this work we investigate the performance of five latest semantic segmentation methodologies on real-life crop row datasets for solving the crop row detection task. Our experimental results validate that most of the semantic segmentation methods provide substantially good results for solving the crop row detection task; Importantly, LinkNet architecture provides the best results among the competitors. We also discuss various reallife challenges for solving crop row detection in real-life scenarios.