Predicting future trajectories of surrounding traffic agents is critical for safe autonomous navigation and collision avoidance. Despite all advances in the trajectory forecasting realm, the prediction models remains vulnerable to uncertainty caused by occlusions, limited sensing range, and perception errors. Collaborative vehicle-to-vehicle (V2V) approaches help reduce this uncertainty by sharing complementary information. Existing collaborative trajectory prediction methods typically fuse feature maps at the perception stage to construct a holistic scene view. Further this holistic representation is decoded into the future trajectories. Such design incurs substantial communication overhead due to the exchange of high-dimensional feature representations and often assumes idealized bandwidth and synchronization, limiting practical deployment. We address these limitations by shifting collaboration from perception to the prediction module and introducing a late-fusion framework for shared forecasts. The framework is model-agnostic and treats collaborating vehicles as independent asynchronous agents. We evaluate the approach on the OPV2V, V2V4Real, and DeepAccident datasets, comparing individual and collaborative forecasting. Across all datasets, late fusion consistently reduces miss rate and improves trajectory success rate (TSR_0.5), defined as the fraction of ground-truth agents with final displacement error below 0.5 m. On the real-world V2V4Real dataset, collaborative prediction improves the success rate by 1.69% and 1.22% for both intelligent vehicles, respectively, compared with individual forecasting.

This article presents a two-stage, biologically inspired neuromorphic framework for autonomous robot navigation in cluttered environments. The proposed architecture combines unsupervised spiking neural network (SNN)-based sensory abstraction with reward-modulated spike-timing-dependent plasticity (R-STDP) for decision-making. In the first stage, raw 360${}<^>{\circ }$ LiDAR measurements are transformed into a compact, interpretable obstacle state through a lateral-inhibition-driven STDP network, yielding a low-dimensional, behaviorally relevant perception of the environment. In the second stage, navigation actions are learned via reward-modulated STDP operating on this abstracted state, supporting long-horizon goal-directed behavior without backpropagation or deep reinforcement learning. Two navigation paradigms are investigated: conventional goal-oriented navigation using relative bearing information, and a probabilistic field-based formulation that enables source-seeking behavior under goal uncertainty. The proposed approach is evaluated extensively in Gazebo and NVIDIA Isaac Sim using a TurtleBot3 platform across static, dynamic, and near-realistic environments. Experimental results demonstrate reliable navigation performance, competitive success and collision rates compared to state-of-the-art SNN and hybrid SNN-RL methods, and substantially lower estimated energy consumption. These findings highlight the effectiveness of modular, biologically plausible neuromorphic architectures for energy-efficient autonomous navigation in complex environments.

This work addresses a critical limitation in state estimation and localization accuracy when applying the Kalman filter to 3D LiDAR point clouds for multi-object tracking (MOT), as well as the ongoing challenge of selecting appropriate motion models. Existing literature commonly employs constant-motion assumptions (e.g., constant velocity or constant acceleration), which neglect the complex and object-specific motion dynamics observed in real-world scenarios. Consequently, trajectory deviation and imprecise object localization arise, particularly under occlusion. We attribute these limitations to the fixed formulation of the conventional Kalman filter, which fails to account for the variability of motion dynamics as objects navigate their environments. To overcome this limitation, we introduce a novel Kalman filter formulation that explicitly incorporates motion dynamics, enabling the motion model to adapt online to changes in object-specific movement. The proposed approach substantially improves state estimation, localization accuracy, and trajectory prediction compared to the traditional Kalman filter. This improvement is reflected in tracking performance that surpasses recent benchmarks on the KITTI and Waymo Open Dataset (WOD), with gains of $0.56\%$ and $0.81\%$ in higher-order tracking accuracy (HOTA) and multi-object tracking accuracy (MOTA), respectively. Furthermore, the proposed Kalman filter consistently outperforms the baseline across a variety of detectors. It also demonstrates enhanced robustness to long-term occlusions, achieving improvements of $1.22\%$ in HOTA and $1.55\%$ in MOTA on the KITTI dataset. The formulation is computationally efficient, introducing only approximately $0.078$~ms of additional processing time per frame, thereby maintaining real-time applicability.

Traditional approaches to off-road autonomy rely on separate models for terrain classification, height estimation, and quantifying slip or slope conditions. Utilizing several models requires training each component separately, having task specific datasets, and fine-tuning. In this work, we present a zero-shot approach leveraging SAM2 for environment segmentation and a vision-language model (VLM) to reason about drivable areas. Our approach involves passing to the VLM both the original image and the segmented image annotated with numeric labels for each mask. The VLM is then prompted to identify which regions, represented by these numeric labels, are drivable. Combined with planning and control modules, this unified framework eliminates the need for explicit terrain-specific models and relies instead on the inherent reasoning capabilities of the VLM. Our approach surpasses state-of-the-art trainable models on high resolution segmentation datasets and enables full stack navigation in our Isaac Sim offroad environment.

As the potential for autonomous vehicles to be integrated on a large scale into modern traffic systems continues to grow, ensuring safe navigation in dynamic environments is crucial for smooth integration. To guarantee safety and prevent collisions, autonomous vehicles must be capable of accurately predicting the trajectories of surrounding traffic agents. Over the past decade, significant efforts from both academia and industry have been dedicated to designing solutions for precise trajectory forecasting. These efforts have produced a diverse range of approaches, raising questions about the differences between these methods and whether trajectory prediction challenges have been fully addressed. This paper reviews a substantial portion of recent trajectory prediction methods proposing a taxonomy to classify existing solutions. A general overview of the prediction pipeline is also provided, covering input and output modalities, modeling features, and prediction paradigms existing in the literature. In addition, the paper discusses active research areas within trajectory prediction, addresses the posed research questions, and highlights the remaining research gaps and challenges.

Reliable visual tracking and localization remain significant challenges in underwater swarm robotics due to limited visibility, frequent occlusions, and stringent computational constraints. This technical communication presents a lightweight visual perception framework designed and validated for deployment on the edge of underwater robotic swarms, using the heterogeneous swarm of underwater robotic fish as the experimental platform. Rather than proposing new algorithms, the work focuses on integrating and adapting existing monocular vision-based keypoint detection and multiobject tracking methods for real-time operation on resource-constrained hardware. The framework achieves a balance between accuracy and computational efficiency, demonstrating practical feasibility for edge deployment. Experimental validation showed an average heading error of 17.2(degrees )and an average depth error of 24.4 cm, confirming its robustness in relative localization and orientation estimation under realistic underwater conditions. The ByteTrack-YOLOv5n.ONNX combination achieved 2.01 fps on a Raspberry Pi 4, representing a 70.3% improvement over the PyTorch baseline, while ByteTrack-YOLOv8n achieved the highest tracking metrics with higher order tracking accuracy (HOTA) at 50.4%, multi-object tracking accuracy (MOTA) at 77.9%, and identity F1 score (IDF1) at 63.6%. Overall, this work contributes a validated reference for efficient visual perception and localization in underwater swarm robotics, supporting future research on scalable, energy-aware robotic networks for marine applications.

The tracking-by-detection paradigm in multi-object tracking (MOT) typically relies on static appearance descriptors to complement motion estimation. However, these descriptors are frame-independent, limiting their robustness as visual cues. Since such descriptors are often obtained from computationally intensive pretrained backbones, real-time MOT systems frequently abandon appearance cues altogether and rely solely on motion prediction and geometric association. In this work, we introduce Polycepta, an object-centric appearance state estimation framework that reformulates appearance modeling as a recursive estimation problem rather than a frame-wise matching task. Polycepta constructs and continuously updates an independent appearance state for each tracked object, enabling future appearance representations to be estimated from accumulated observations. Polycepta is encouraged to learn the appearance-state construction of object-specific representations rather than memorize them through a proposed learning strategy, enabling appearance estimation for unseen classes. A key property of Polycepta is that the quality of appearance estimation improves as object states evolve during inference. While conventional appearance descriptors remain static or degrade over time, Polycepta progressively refines appearance estimates as additional observations are accumulated. Extensive experiments on KITTI, the Waymo Open Dataset, and MOT17 demonstrate consistent reductions in identity switches and improvements in tracking performance when integrated into the tracking-by-detection pipelines. Polycepta operates at 90.57 Hz and delivers state-of-the-art performance on the KITTI benchmark when integrated into the RobMOT framework, achieving a MOTA of 92.27%.

In this paper, we introduce snnTrans-DHZ, a lightweight spiking transformer architecture that incorporates a learnable threshold membrane potential mechanism for underwater image dehazing. Despite its compact design with 0.57 M parameters, the method substantially improves underwater image clarity and visibility. Leveraging the temporal dynamics of spiking neural networks, snnTrans-DHZ efficiently processes time-dependent raw image sequences while maintaining low power consumption. The raw underwater images are first converted into time-dependent image sequences by repeatedly passing the static image to a user-defined timestep value. The RGB sequences are then converted into LAB color space representations and processed simultaneously. The architecture integrates three primary modules: (i) K estimator module to extract features from different color space representations, (ii) background light estimator module to jointly estimate the background light component from the RGB-LAB color space representations, and (iii) soft image reconstruction to reconstruct the haze-free, visibility-enhanced image. The snnTrans-DHZ model is directly trained using surrogate gradient-based backpropagation through time strategy. In this research, a combined loss function is designed and used. Our model is trained and tested on the UIEB and EUVP, the two publicly available benchmark dataset for image dehazing. Our algorithm achieves a PSNR of 21.6773dB and SSIM of 0.8795 on UIEB and, on EUVP, it achieves 23.4562dB and 0.8439. snnTrans-DHZ algorithm achieves this algorithmic performance with fewer operations (7.42 GSOPs) and lower energy consumption of 0.0151 J compared to existing state-of-the-art image enhancement methods. It provides a 3.3 x improvement in energy efficiency over the lightest state-of-the-art transformer-based method, making it suitable for underwater robotics and environmental monitoring. The source code is available at snnTrans-DHZ.

High-speed autonomous racing presents extreme perception challenges, including large relative velocities and substantial domain shifts from conventional urban-driving datasets. Existing benchmarks do not adequately capture these high-dynamic conditions. We introduce EagleVision, a unified LiDAR-based multi-task benchmark for 3D detection and trajectory prediction in high-speed racing, providing newly annotated 3D bounding boxes for the Indy Autonomous Challenge dataset (14,893 frames) and the A2RL Real competition dataset (1,163 frames), together with 12,000 simulator-generated annotated frames, all standardized under a common evaluation protocol. Using a dataset-centric transfer framework, we quantify cross-domain generalization across urban, simulator, and real racing domains. Urban pretraining improves detection over scratch training (NDS 0.72 vs. 0.69), while intermediate pretraining on real racing data achieves the best transfer to A2RL (NDS 0.726), outperforming simulator-only adaptation. For trajectory prediction, Indy-trained models surpass in-domain A2RL training on A2RL test sequences (FDE 0.947 vs. 1.250), highlighting the role of motion-distribution coverage in cross-domain forecasting. EagleVision enables systematic study of perception generalization under extreme high-speed dynamics. The dataset and benchmark are publicly available at https://avlab.io/EagleVision

This article introduces NeuroVIO, a hybrid end-to-end neuromorphic architecture that combines conventional deep learning with spiking neural computation for visual–inertial odometry in underwater mobile robots. Accurate and energy-efficient pose estimation in a marine environment remains challenging due to low visibility, dynamic lighting, and limited on-board computational resources. NeuroVIO addresses these challenges by extracting visual features through a convolutional neural backbone, transforming them into temporally encoded spike sequences processed by adaptive leaky-integrate-and-fire neurons with learnable thresholds. Inertial signals are simultaneously embedded by a dedicated spiking feature encoder. The fused multimodal representation is passed through a spiking long short-term memory module to model temporal dependencies, followed by a spiking regression head for six degrees-of-freedom pose estimation. Evaluated on the AQUALOC data set, NeuroVIO with adaptive thresholds reduces energy consumption by 80.4% compared with its nonspiking counterpart while maintaining competitive pose accuracy. These results demonstrate the ability of neuromorphic processing to substantially lower power demands without sacrificing state estimation fidelity. By leveraging biologically inspired computation, NeuroVIO enables scalable and persistent autonomy for resource-constrained marine platforms, establishing a promising direction for next-generation underwater navigation systems.

Precise 3D state estimation in multi-object tracking (MOT) is critical for self-driving cars, particularly for objects occluded. Motion modeling in the Kalman filter with a constant motion assumption is widely used in MOT methods, but it neglects the continuous changes in objects' motion caused by traffic in urban environments. Although recent research introduces a multimodel Kalman filter that incorporates multiple motion models, these approaches incur significant computational overhead from the simultaneous processing of multiple models. To this end, this work introduces a motion-dynamics Kalman filter (MD-KF) that overcomes the constant-motion assumption while preserving the singularity of the motion model. MD-KF models the changes in objects' motion over successive measurements as Gaussian distributions, and adaptively adjusts a weighted motion model to account for these variations. MD-KF consistently outperforms constant and multimodel KF across multiple datasets with a significant reduction in computation latency compared to multimodel approaches. The proposed approach demonstrates its superiority in trajectory estimation during occlusion and state estimation stability for stationary objects.

While crucial to customer satisfaction, last-mile delivery (LMD) remains the most time-consuming and costly stage of the shipping process. Pressing environmental concerns and the recent surge in e-commerce sales have sparked renewed interest in automating and electrifying last-mile logistics. To address the limitations of current robotic couriers, this paper presents a customer-centric and safety-conscious LMD system designed for small urban communities. Built on AI-assisted autonomous delivery robots, the proposed system enables end-to-end automation and optimization of the logistics process, with added considerations for pedestrian safety. The optimization component is modeled as a robust variant of the Cumulative Capacitated Vehicle Routing Problem with Time Windows (RCCVRPTW), accommodating real-world operational uncertainties, customer satisfaction, and delivery preferences. Specifically, RCCVRPTW constructs routes under uncertain travel times with the objective of minimizing total delivery latency (i.e., customers’ waiting time), directly addressing factors linked to customer dissatisfaction. To validate the effectiveness of the system, real-world proof-of-concept tests are conducted at a university campus using a single robotic courier. Detailed discussions cover key implementation aspects, findings, and insights gained from these deployment experiments. Finally, the scalability of the RCCVRPTW model is investigated in the Gurobi solver with varying numbers of robotic vehicles and customers.

The release of large-scale autonomous vehicles (AV) datasets has become essential for advancing perception models capable of generalizing across diverse environments. However, visual data captured in public settings often contains identifiable information about pedestrians and vehicle owners, raising privacy concerns and compliance issues with data protection regulations. Since AV datasets are inherently non-consensual, anonymization becomes a necessary step before public data distribution. In this paper, we introduce Priv-DATA, an open-source tool for AV dataset anonymization. It supports: (1) Automatic anonymization of privacy-sensitive content (e.g., faces, license plates) using state-of-the-art detection models with Gaussian blurring in a systematic way that while anonymization, it minimizes any resultant visual artifacts; and (2) User-friendly validation via a GUI for manual correction and artifact removal. We validate Priv-DATA on the Emirates Multi-task dataset and discuss challenges in balancing privacy and data utility. The code is publicly available on GitHub11https://github.com/AV-Lab/Dataset_Anonymization.

Accurately measuring combustion efficiency is critical for optimizing industrial flare performance, reducing emissions, and ensuring regulatory compliance. However, existing methods rely on expensive gas analyzers and optical sensors that require frequent calibration and maintenance, making them unsuitable for large-scale or remote deployment. Moreover, no current solutions can directly estimate combustion efficiency from visual video data. To address this gap, we present a novel multimodal framework for estimating combustion efficiency using synchronized RGB, infrared (IR), and thermal video data, complemented by real-time gas analyzer measurements. Due to the lack of public datasets for this task, we constructed a custom flare setup and captured multimodal recordings under controlled environmental conditions. To address the temporal and spatial disparities among different camera streams, we implemented frame rate alignment and affine video registration. Timestamp-synchronized gas sensor readings were used to compute combustion efficiency for each frame, providing ground truth labels. These labels were transformed into semantic text prompts ("high," "medium," "low" efficiency) for fine-tuning visual-language model on image-text alignment. High-confidence image-text pairs were subsequently used to train a deep regression model for continuous efficiency estimation. Comparative experiments across multiple vision backbones reveal that the visual-language encoder, particularly when enhanced with frame averaging and sensor fusion, achieves superior performance (MAE: 0.039, R2: 0.822), outperforming traditional CNNs and transformers. The results demonstrate the potential of combining multimodal vision and sensor data for accurate, scalable combustion monitoring.

Prostate pathology analysis has been handled remarkably well by deep learning models. However, deep learning models in histopathology still face challenges in effectively modeling fine-grained, object-level features that are critical for a robust cancer assessment. Topological Data Analysis (TDA) has shown promise in addressing these issues but remains underexplored, particularly for whole-slide pathology applications. This is manifested in the nonexistent application of TDA on the slide level; an issue that significantly undermines their practical usability, and the non-utilization of the multi-magnification nature of Whole-slide Images (WSIs). Furthermore, existing studies are limited to small-scale datasets, thus challenging the validity of TDA in histopathology. In this work, we address these gaps by introducing Persistent Homology in Multiple Instance Learning (PMIL), the first adaptable TDA-based module within the MIL framework. We further propose the Cubic version cPMIL, utilizing the magnification factor in pathology images for filtration and unlocking an improved object-level modeling capability. We validate our approach on two prostate subtyping datasets, comparing against multiple state-of-the-art methods. Our proposed modules are developed offline, can be seamlessly plugged in any MIL pipeline, do not incur any additional expert annotation, operational within clinical limitations, user-friendly, and open-source. We provide a detailed guide for implementing the modules at github.com/ahmadobeid/PMIL.

Underwater image enhancement (UIE) is fundamental for marine applications, including autonomous vision-based navigation. Deep learning methods using convolutional neural networks (CNN) and vision transformers advanced UIE performance. Recently, spiking neural networks (SNN) have gained attention for their lightweight design, energy efficiency, and scalability. This paper introduces UIE-SNN, the first SNN-based UIE algorithm to improve visibility of underwater images. UIE-SNN is a 19- layered convolutional spiking encoder-decoder framework with skip connections, directly trained using surrogate gradient-based backpropagation through time (BPTT) strategy. We explore and validate the influence of training datasets on energy reduction, a unique advantage of UIE-SNN architecture, in contrast to the conventional learning-based architectures, where energy consumption is model-dependent. UIE-SNN optimizes the loss function in latent space representation to reconstruct clear underwater images. Our algorithm performs on par with its non-spiking counterpart methods in terms of PSNR and structural similarity index (SSIM) at reduced timesteps ($T=5$) and energy consumption of $85\%$. The algorithm is trained on two publicly available benchmark datasets, UIEB and EUVP, and tested on unseen images from UIEB, EUVP, LSUI, U45, and our custom UIE dataset. The UIE-SNN algorithm achieves PSNR of \(17.7801~dB\) and SSIM of \(0.7454\) on UIEB, and PSNR of \(23.1725~dB\) and SSIM of \(0.7890\) on EUVP. UIE-SNN achieves this algorithmic performance with fewer operators (\(147.49\) GSOPs) and energy (\(0.1327~J\)) compared to its non-spiking counterpart (GFLOPs = \(218.88\) and Energy=\(1.0068~J\)). Compared with existing SOTA UIE methods, UIE-SNN achieves an average of \(6.5\times\) improvement in energy efficiency. The source code is available at \href{https://github.com/vidya-rejul/UIE-SNN.git}{UIE-SNN}.

In underwater missions, energy-efficient operation is critical to maximize the operability of an autonomous robots and mission duration. At the same time, accurate pose estimation at reduced computational cost is essential for various underwater robotics and exploration applications. This paper presents an energy-efficient model for pose regression tasks based on direct Spike-Coding-Decoding (SCD) with adaptive threshold mechanism. The proposed approach enables efficient processing of multimodal continuous-valued images and IMU data. The effectiveness of the proposed module is tested using an open-sourced underwater simulator. In underwater robotics domain, real-world testing remains critical due to limited access to actual underwater environment, making simulators invaluable for rapid algorithm development and refinement. Accordingly, HoloOcean simulator is utilized to collect images and IMU sensor data from various underwater scenarios, which are then fed into a pose estimation framework. The SCD module uses 2D and 1D convolutional layers followed by LIF neurons with adaptive threshold to convert continuous-valued data into sparse spike-based signals. During spike decoding, the spike-based signals are converted back to continuous-values by utilizing the accumulated membrane potential of the output LIF neurons. The feasibility of SCD module is evaluated by its ability to introduce sparsity in spike-based representations with minimal information loss that enables efficient signal reconstruction. Custom multimodal data is collected using the HoloOcean simulator and is used to train the network. We compared the performance of the adaptive learnable threshold method with fixed threshold methods. The experimental results show that spike coding with an adaptive threshold mechanism is highly effective compared to a fixed threshold mechanism, as it can encode the continuous-valued input to its spike-based representation with approximate to 2x higher sparsity and a lesser spike rate.

Acoustic communication allows underwater Remotely Operated Vehicles (ROVs) to be controlled wirelessly, improving their maneuverability by removing the need for a physical tether connection to the surface. However, acoustic communication suffers from severe delays, degrading the operator's performance due to the lag between the given command and the vehicle's corresponding movement. In this paper, a novel delay mitigation approach is introduced, utilizing a Predictive Display (PD) enhanced by a Long Short-Term Memory (LSTM) neural network. The PD predicts the robot's delayed movements using dynamic modeling and LSTM-based time series forecasting. It then displays a virtual robot in real-time based on the operator's commands, which the real robot follows after a delay. Experiments are conducted on an acoustically-teleoperated, bioinspired robotic fish in our facility's pool with time delays reaching 2.5 seconds. A Visual Positioning System (VPS) is utilized to observe the robot's poses as ground truth. The LSTM network is trained to minimize the error between the robot's pose from its dynamic model and the ground truth. The corrected pose is displayed to the operator on the PD, with a haptic cueing device alerting the operator to potential collisions with obstacles. The results show average reductions of 82.3% and 65.24% in pose and angular errors, respectively, when utilizing the LSTM network compared to solely relying on the dynamic model. The code and dataset are available at: https://github.com/MahmoudElMezain/LSTM_PD

As the automotive industry advances toward higher levels of autonomy, decision-making frameworks must evolve to address increasingly complex and dynamic environments. This paper presents a novel approach called Multi-Grid Markov Decision Processes (mg-MDP), designed to enhance scalability, robustness, and efficiency in autonomous vehicle decision-making. Building on the foundations of traditional Markov Decision Processes (MDPs), mg-MDP utilize a hierarchical multi-layer grid structure to better represent distinct aspects of the environment. Through extensive simulations, we show that mg-MDP incrementally adjusts decision-making across multiple grid- based layers, efficiently handling dynamic traffic scenarios such as intersections, lane merging, and obstacle avoidance. This approach intends to reduce the computational effort while improving decision accuracy. This paper also discusses how mg-MDP can be applied in Cyber-Physical Systems for better real-world modeling that will leverage intelligent transportation.