The validation of autonomous driving systems benefits greatly from the ability to generate scenarios that are both realistic and precisely controllable. However, real-world test drives are costly and lack the flexibility to target edge cases. We propose a bi-level latent diffusion framework that first train a latent diffusion model (DM) and then fine tune it via reinforcement learning (RL). RL optimizes diffusion-based generation to meet user defined specific target objectives, enabling automatic generation of diverse, fine-grained and controllable driving scenarios for virtual testing. Different from traditional guided sampling of DM, our method delivers a generative model inherently for traffic simulation without the need for external guidance signals or handcrafted conditioning mechanisms during sampling. Experiments show that our method delivers substantially fast inference speed while maintaining strong overall quality across realism, safety and stability. These results indicate that bi-level latent diffusion with RL-guided training can effectively improve safety, stability, and controllability of generated scenarios, contributing to more rigorous and nuanced safety evaluation of autonomous vehicles.

Accurate state of charge estimation is critical for the success of electric vehicle battery management strategies, but it is well known that conventional estimators suffer from two fundamental shortcomings: cumulative errors that grow over time and reliance on simplified battery models that do not reflect real world dynamics. Therefore, this paper presents a novel hybrid approach combining Tucker tensor decomposition with LSTM networks, using full - lifecycle EV field data for SOC prediction. The inputs are charge status, mileage, voltage, current, cell differentials, and temporal features. Tucker decomposition is skillfully used to reduce dimensionality while maintaining the temporal structure, hence allowing a direct, fair comparison with standard LSTM. The result is unequivocal: Tucker - LSTM outperforms the baseline on all metrics, with MSE dropping 70.5% (from 21.07 to 6.22 ), MAE improving 48.7% (from 3.37% to 1.73%), RMSE falling from 4.59% to 2.49%, and R^2 rising from 0.918 to 0.976. Since the experimental results demonstrably demonstrate that tensor decomposition compresses high-dimensional battery data very well without loss of predictive fidelity, this paper naturally opens up a new direction for tensor-based analytics in electric vehicle battery management.

Point clouds provide discrete representations of 3D scenes. The relative positions and directions between points collectively describe the objects. Variations in sampling angles, distances, or noises can introduce perturbations, disrupting these spatial and directional relationships. These pose significant challenges for achieving robust feature representations. However, research on the robust representation of point clouds is limited. Although advanced models achieve impressive performance, they exhibit poor robustness to perturbations. To address this issue, we propose Geometry-enhanced Directional Consistency Representation (GDCR), a novel method designed to enhance robustness. In GDCR, we introduce Statistic-based Geometric Reasoning (SGR) to achieve precise spatial geometric estimation for discrete point sets, explicitly enriching spatial geometric information. Furthermore, GDCR vectorizes features embedded with SGR information and applies feature rotation and relative direction refinement in the expanded feature space for robust directional representation. GDCR improves the flexibility and directional expressiveness of point cloud features, significantly improving robustness against perturbations. Extensive experiments demonstrate that GDCR exhibits outstanding robustness while surpassing or matching the performance of state-of-the-art methods.

Behavioral heterogeneity in mixed traffic challenges the safety and comfort of autonomous vehicle (AV) lane-changing maneuvers. To address this, this paper proposes a trajectory planning framework that integrates dynamic risk awareness. First, a hybrid GMM-LightGBM approach identifies the driving styles of surrounding human-driven vehicles (HDVs). Subsequently, personalized trajectory predictions are generated based on the FVD car-following model. Second, by quantifying the spatiotemporal coupling relationships among traffic participants, a dynamic risk assessment system is constructed to adapt to the differentiated behavioral characteristics of HDVs. Building on this, an innovative Lane-Change-Specific Dual-Graph Spatio-Temporal VectorNet (LC-DGST-VNet) predicts the lane-change duration and displacement. By utilizing these predictions to narrow sampling scopes and dynamically adjust risk weights, the framework achieves real-time, risk-aware planning, executed via Model Predictive Control (MPC). Simulation results demonstrate that, compared with traditional methods that neglect behavioral heterogeneity and environmental risk levels, the framework significantly optimizes safety and smoothness. Specifically, the average collision risk is reduced from 9.2 to 1.8

Single positive multi-label learning (SPML) aims to recognize multiple categories with limited supervision from one positive label in an image. With the emergence of pre-trained visual-language models such as CLIP, recent studies focused on capturing label-to-label dependencies. However, hierarchies with deeper layers of labels or more branches in label-to-label relationships cannot be well expressed in Euclidean space. To address the challenge, we introduce a semantic hierarchy-aware hyperbolic representations framework for single positive multi-label learning. Specifically, drawing inspiration from semantic hierarchical information, we introduce a label relation prior strategy to map single labels to other labels. The semantic chain of labels is extracted along the hierarchical path from the child node to the parent node. Furthermore, hyperbolic entailment constraints are adopted to enforce the semantic similarity between image-text pairs and the hierarchical consistency among labels in hyperbolic space. Experimental results conducted on four SPML benchmark datasets demonstrate that our SHHNet achieves state-of-the-art performance.

Change detection (CD) technique has wide applications in remote sensing-based application. Most existing works regard CD as a task of pixel-level binary classification to distinguish changed areas from unchanged ones. In order to segment the changed areas and classify ground objects before and after the change, an attention mechanism augmented semantic change detection (SCD) model, Multi-Task learning and Attention mechanism based Semantic Change Detection (MTASCD), is developed, which translates SCD into the combination of binary-classification CD task and multi-classification object segmentation (MCOS) task, and fuses the semantic features from CD and MCOS for reliable segmentation and reduction of training complexity of model. In the proposed model, a Siamese network with shared weights is designed to extract the semantic features of ground objects and separate the false changes caused by environment like season change or lighting condition. The prediction result of the CD task feeds the MCOS task, and then fuses with the bi-temporal image features at the higher feature level so as to retain more detailed context information. Attention mechanism is introduced to promote the generalization and robustness of the model. Sufficient experiments are carried out on the open Land-CD dataset, and comparative results show that MTASCD can improve the accuracy of semantic change detection, especially for small-sized objects.

Conventional pruning paradigm typically determines the pruned network structure while identifying and removing filters, requiring iterative pruning and fine-tuning, which incurs substantial time and computational costs. Moreover, existing methods overly emphasize the importance of individual filters while neglecting the optimization of the overall network structure, resulting in performance degradation. In this letter, a new Coarse-to-Fine Fast Filter Pruning (CF-FFP) framework is proposed, which decomposes the conventional pruning paradigm into two offline learning stages to achieve fast and efficient model compression. Specifically, the pruned network structure is coarsely determined based on the importance of weights, and an adaptive balancing strategy is proposed to address the issue of significant pruning rate differences across layers. Then, a dual redundancy screening criterion is proposed to finely identify and prune redundant filters based on their similarity and contribution, thereby initializing the pruned network structure. Thanks to CF-FFP’s two-stage offline pruning process, which progresses from coarse to fine, the pruning inference time is significantly reduced. Extensive experiments show that our method outperforms the state-of-the-art methods on CIFAR-10 and ImageNet datasets. For instance, CF-FFP prunes 51.2% FLOPs of ResNet50 on the ImageNet dataset with only 0.67% drop in Top-1 accuracy.

Conventional pruning paradigm typically determines the pruned network structure while identifying and removing filters, requiring iterative pruning and fine-tuning, which incurs substantial time and computational costs. Moreover, existing methods overly emphasize the importance of individual filters while neglecting the optimization of the overall network structure, resulting in performance degradation. In this letter, a new Coarse-to-Fine Fast Filter Pruning (CF-FFP) framework is proposed, which decomposes the conventional pruning paradigm into two offline learning stages to achieve fast and efficient model compression. Specifically, the pruned network structure is coarsely determined based on the importance of weights, and an adaptive balancing strategy is proposed to address the issue of significant pruning rate differences across layers. Then, dual redundancy screening criterion is proposed to finely identify and prune redundant filters based on their similarity and contribution, thereby initializing the pruned network structure. Thanks to CF-FFP's two-stage offline pruning process, which progresses from coarse to fine, the pruning inference time is significantly reduced. Extensive experiments show that our method outperforms the state-of-the-art methods on CIFAR-10 and ImageNet datasets. For instance, CF-FFP prunes 51.2% FLOPs of ResNet50 on the ImageNet dataset with only 0.67% drop in Top-1 accuracy.

In this article, we propose a dual-driven modeling method for vehicle-mounted LiDAR sensors. The method has a two-stage modeling framework comprising point cloud position modeling and intensity modeling, respectively. In the spatial position modeling stage, we generate ideal LiDAR point clouds by employing the geometric relationships of spatial rays. Subsequently, we integrate the LiDAR detection equation with a probability model to simulate the attenuation and noise. In the intensity modeling stage, a network InPNet is leveraged to train an intensity model. This model is able to predict accurate reflection intensity for point clouds generated in the previous stage. Experiments have shown that the proposed method is able to simulate point clouds within +/- 10% errors on various driving scenarios. We also mix the simulated point clouds as augmentation data with the real scanning point clouds and employ them to train several representative 3-D object detection models under clear and adverse weather conditions. The results demonstrate promising prospects for engineering applications.

Light Detection And Ranging (LiDAR) sensors can generate a number of sequential 3D point clouds, which are widely deployed in many real-world systems. 3D object detection in point clouds, is one of the most fundamental tasks. Unfortunately, the existing 3D object detection methods degrade in snowy weather, because in that situation the annotated samples are difficult to collect. To solve this issue, we propose a novel GAN-based Snowfall Point-cloud AugmentOR (GAN spaor ) to generate high-quality synthetic snowfall point clouds as augmentations. The basic idea of GAN spaor is to transfer annotated point clouds to snowfall versions by simultaneously learning the global style of real snowfall point clouds and the local details of physics-induced ones. Our framework fuses data-driven and physical modeling methods for rapidly generating data in snowy weather. To evaluate the effectiveness of GAN spaor , we employ a number of recent 3D object detection methods and train them by using the synthetic samples of GAN spaor as auxiliary augmentations. Moreover, we conduct a comparative analysis of the characteristics of the data distributions of the snowy point clouds synthesized by GAN spaor . Experimental results demonstrate that GAN spaor can improve the performance of 3D object detection methods compared with other existing snowfall point cloud simulators.

Complex traffic scenarios at uncontrolled intersections are crucial for the test validation of autonomous driving systems. The core of the test scenario construction lies in the accurate modeling of the complex interaction behaviors between vehicles in dynamic traffic. Data-driven models are difficult to support long-term simulation due to the existence of cumulative errors. In addition, existing mechanistic models usually assume rational driver behavior and focus mainly on improving efficiency and safety, thus simplifying vehicle interactions. To overcome the limitations of existing studies, we construct a complex traffic interaction model based on social force theory. This model captures the intricate interactions among vehicles at uncontrolled intersections by introducing the concepts of driving and repulsive forces. In particular, we propose a novel concept of segmented conflicting repulsion, an approach that can accurately model high-risk scrambling interactions between vehicles at intersections. Validation of the model using real data sets demonstrates its ability to accurately reproduce complex interaction behaviors at real-world intersections. Further, simulation analysis and application results reveal that our model-generated scenarios significantly outperform those created by SUMO in terms of complexity, thereby effectively enhancing the safety assessment of the autonomous driving system, Apollo.

As autonomous driving technology scales up, complex urban intersections pose significant safety challenges. Current testing methods struggle to simulate these complex scenarios at a manageable cost, making simulation testing essential. For effective evaluation, establishing comprehensive and objective complexity metrics is crucial. However, existing complexity evaluation methods often depend on the performance of the primary vehicle and are based on local interaction relationships, which lack a global perspective and objectivity and have yet to be validated by autonomous driving systems. To address this issue, this paper proposes a multidimensional complexity assessment framework that introduces system-level indicators such as vehicle count, interaction density, disorder, and risk. This framework quantifies the complex interactions at intersections from a global perspective, independent of primary vehicle performance. Experimental results demonstrate that the complexity evaluation results are highly consistent with the performance of a high-level autonomous driving system (Apollo). The framework has been successfully applied to test scenario generation on the Apollo platform, achieving twice the scenario generation efficiency of traditional methods, thus showcasing substantial engineering value.

Merge scenarios on highway are often challenging for autonomous driving, due to its lack of sufficient tacit understanding on and subtle interaction with human drivers in the traffic flow. This, as a result, may impose serious safety risks, and often cause traffic jam with autonomous driving. Therefore, human-like autonomous driving becomes important, yet imperative. This paper presents an interaction-aware decision-making and planning method for human-like autonomous driving in merge scenarios. Rather than directly mimicking human behavior, deep inverse reinforcement learning is employed to learn the human-used reward function for decision-making and planning from naturalistic driving data to enhance interpretability and generalizability. To consider the interaction factor, the reward function for planning is utilized to evaluate the joint trajectories of the autonomous driving vehicle (ADV) and traffic vehicles. In contrast to predicting trajectories of traffic vehicles with the fixed trajectory of ADV given by the upstream prediction model, the trajectories of traffic vehicles are predicted by responding to the ADV's behavior in this paper. Additionally, the decision-making module is employed to reduce the solution space of planning via the selection of a proper gap for merging. Both the decision-making and planning algorithms follow a “sampling, evaluation, and selection” framework. After being verified through experiments, the results indicate that the planned trajectories with the presented method are highly similar to those of human drivers. Moreover, compared to the interaction-unaware planning method, the interaction-aware planning method behaves closer to human drivers.

Human-like autonomous driving can enhance user acceptance and integration within traffic. In light of this, this paper presents a planning method for the human-like longitudinal trajectory in car-following scenarios with offline sample-based maximum entropy deep inverse reinforcement learning (DIRL). Specifically, the proposed method doesn’t mimic human driving behavior directly. Instead, it uses naturalistic driving data to learn the internal reward function that leads to these driving behaviors. To enhance the capacity for fitting the human reward function, DIRL leverages deep neural networks to replace linear functions used by traditional IRL. However, the long-tail effect of naturalistic driving data makes it challenging for DIRL to capture the reward function in edge scenarios. A simulated dataset covering edge scenarios is collected by employing feature-based inverse reinforcement learning to deal with this challenge. Furthermore, this paper trains a long-term actor network guided by DIRL’s reward network. The long-term actor network significantly reduces the computation cost by three orders of magnitude compared to the reward network-based method, while also avoiding system oscillation in contrast to the traditional one-step actor network. The simulation experiments confirm that the planning results from the proposed method are closer to human drivers’ behavior than the baseline. And, the hardware-in-the-loop experiment results affirm the proposed method’s effectiveness and good real-time performance. Note to Practitioners —Human-like autonomous driving can enhance user acceptance and integration within traffic. This paper proposes a planning method for human-like car-following driving behaviors using offline sample-based deep inverse reinforcement learning (DIRL). Instead of mimicking human trajectories directly, the method learns the internal reward function leading to these driving behaviors. To address challenges posed by naturalistic driving data’s long-tail effect, a simulated dataset covering edge scenarios is collected by employing feature-based inverse reinforcement learning. Additionally, this paper trains a long-term actor network guided by DIRL’s reward network to reduce the computation cost. The simulation experiments confirm that the planning results from the proposed method are closer to human drivers’ behavior than the baseline. And, the hardware-in-the-loop experiment results affirm the proposed method’s effectiveness and good real-time performance.

Modeling driving behavior plays a pivotal role in advancing the development of human-like autonomous driving. In light of this, this paper proposes a car-following behavior modeling method with sample-based deep inverse reinforcement learning (DIRL). Due to the challenges associated with feature extraction and the limited fitting capacity of linear functions, traditional IRL, which employs feature-based linear functions to represent reward functions, exhibits low modeling accuracy. Accordingly, DIRL leverages deep neural networks to represent reward functions. However, the requirement for reinforcement learning to determine the optimal policy for DIRL's reward function makes the training and inference processes computationally resource-intensive and inefficient. To address this issue, this paper proposes the sample-based DIRL. Through solution space discretization, sample-based DIRL streamlines the integration calculation of the partition function into a summation, resulting in improved computational efficiency. Specifically, it is a three-stage framework: sampling candidate trajectories, evaluating candidate trajectories, and selecting the trajectory with the highest reward. In order to evaluate DIRL at both the level of driving behavior and the reward function, the MPC-based virtual driver with the explicit reward function is utilized to collect driving data for training and assessing the convergence of the learned reward function. The experimental results confirm that the proposed method can accurately model the car-following behavior, and acquire the driver's reward function from the driving data.

Autonomous driving represents the future of transportation, and simulation testing is a critical technology for ensuring the safety and reliability of autonomous driving. In simulation systems, a real-time and accurate radar model is crucial for enhancing the confidence of autonomous driving simulation systems. Existing radar models are typically simplistic and cannot accurately reflect the detection results of real radars. This paper introduces a data-driven radar modeling method that analyzes radar detection mechanisms to identify the model’s input parameters. Furthermore, this method decouples the model’s output parameters based on the differing mechanisms of radar output parameters and tailors model structures to align with the mechanistic characteristics of each parameter. Subsequently, considering the varied influences among input parameters, a segmented model structure is proposed that reduces input dimensionality while retaining critical information. Lastly, the model outputs are optimized using Kalman filtering. This study collects radar data from real vehicles to train the radar model. Testing has demonstrated that the radar model developed by this paper’s method yields outputs very close to actual radar predictions, with a significant improvement in prediction accuracy compared to traditional radar models. The model operates in approximately 0.8 ms, markedly faster than the real radar’s 72-millisecond detection cycle. This demonstrates that the radar model developed using this modeling method can predict radar detection results accurately and in real-time.

The development of autonomous driving technology has made simulation testing one of the most important tools for evaluating system performance. However, there is a lack of systematic methods for analyzing and assessing naturalistic driving trajectory datasets. Specifically, there is a lack of comprehensive analyses on data diversity and balance in machine learning-oriented research. This study presents a comprehensive assessment of existing highway scenario datasets in the context of traffic modeling in autonomous driving simulation tests. In order to clarify the level of traffic risk, we design a systematic risk index and propose an index describing the degree of data scatter based on the principle of Euclidean distance quantization. By comparing several datasets, including NGSIM, highD, INTERACTION, CitySim, and our self-collected Highway dataset, we find that the proposed metrics can effectively quantify the risk level of the dataset while helping to gain insight into the diversity and balance differences of the dataset.

Steering feel holds a main role in driving, making steering feedback torque (SFT) generation an important research of steer-by-wire (SBW) system. Authentic steering feel provides sensations comparable to conventional steering systems, while the SFT model aligns with driver preferences, reducing cognitive load and fatigue, thereby improving vehicle safety. Current research on driver-preferred SFT is limited to real-world trials, a method often time-intensive and suboptimal. This article presents an innovative approach for obtaining optimal SFT with reinforcement learning (RL) guided by the human reward model. The human reward model, trained on human feedback data, is introduced to guide the SFT model's optimization. Given complexity of assessing human evaluations, the reward model is validated with virtual data. Building upon previous data-driven SFT modeling, a fine-tuning method leveraging the RL method is presented, using human reward model's outputs. This approach retains the nonlinear predictive capabilities of the pretrained model while aligning with human evaluations. Results confirm the human reward model's ability to replicate human SFT standards. Models adapted using this network retain the nonlinear predictive capabilities of the initial network, even under high-frequency inputs. Overall, this work pioneers the integration of human feedback into SFT optimization, providing novel insights for the advancement of SBW systems.

Simulation is crucial for autonomous driving technology evolution. Radar, as an essential sensor in this field, significantly influences decision-making with its outputs. High-fidelity autonomous driving simulations require radar models that replicate radar outputs, including false alarms, missed alarms, and measurement errors, both in real-time and with high fidelity. The radar detection process is highly complex, and false and miss alarms add significant uncertainty to the detection results. Current radar models cannot accurately predict radar outputs. To address these issues, this study introduces a data-driven radar modeling approach. Initially, an analysis of factors influencing radar detection outcomes was conducted. Then proposes a labeling method for radar output objects, identify the corresponding scene targets, and distinguish between ghost and real objects. Following this, it introduces a modeling technique that separates radar output status and parameters, aiming to accurately predict radar outputs in the presence of false and missed alarms. It further decouples output parameters to boost prediction accuracy. Radar data is then collected to create a dataset. The radar model is developed and validated against conventional models. The model achieves a 96.5% accuracy in predicting false and missed alarms, with its predictions for radar output parameters closely approximating actual values. Compared to traditional models, there are improvements exceeding 70.60% and 93.68% respectively. Its 5-millisecond processing speed is substantially faster than actual radar speeds. This demonstrates the method's ability to create high-fidelity, real-time models.