Safety concerns remain one of the most significant obstacles to the large-scale deployment and continued advancement of autonomous driving (AD) systems. A major underlying cause of many safety-related incidents in AD systems is suboptimal or erroneous decision-making when the vehicle encounters corner cases (CCs)—rare, unexpected, or extreme situations that fall outside typical operating conditions. Although recent advances in artificial intelligence have driven substantial progress in both autonomous driving and corner-case research, the field still lacks a coherent conceptual foundation and a systematic, widely accepted categorization of CCs. In this survey, we address this gap by offering a structured review of the existing corner-case literature along three key dimensions: understanding, detection, and generation. The key contribution is a three-level classification of corner cases—spanning data-level, model-level, and semantic-level CCs—that clarifies and disambiguates competing definitions and perspectives on AD corner cases. Building on this framework, we examine simulator-based methods for corner-case selection and data generation, and we identify open challenges, promising directions, and potential solutions to more effectively handle corner cases in autonomous driving systems.

Out-of-distribution (OOD) detection is crucial for deploying models in open-world environments. This process aims to mitigate the issue of overconfident predictions, which is a common problem for models designed for closed-domain tasks when they encounter OOD data. Recent progress in OOD detection has shown that integrating auxiliary datasets during model training can greatly enhance OOD detection performance. However, existing methods tend to heavily depend on these auxiliary datasets to establish the decision boundary for in-distribution (ID) data, while not adequately addressing OOD object detection in safety–critical applications. In this article, we propose object-level OOD detection and introduce a hierarchical hard negative sampling (HNS) strategy that does not necessitate auxiliary data. Specifically, we offer a new metric that strategically considers difficult negatives near the decision boundary between inter-class and intra-class instances. Inspired by adversarial thinking, we sample outliers for each class, ensuring that the negative samples capture both diversity and informative traits. We conducted comprehensive experiments on three public datasets. The results demonstrate that HNS performs superiorly in object-level OOD detection, even without auxiliary datasets. The source code can be accessed at: https://github.com/Aurevior1/HNS .

The development of intelligent driving systems has drawn significant attention to enhancing the safety of autonomous vehicles and their intended functionality. Despite this, current accelerated testing approaches remain inadequate in assessing system reliability, as they fail to simulate scenarios involving collisions between vehicles and pedestrians and identify unknown risks. To address these limitations, scenario-based testing methods have been proposed, which seek to identify critical scenarios with a high frequency of exposure to safety risks. A comprehensive review of these methods is thus of paramount significance. In this article, we provide a timely and systematic literature review of existing accelerated testing for autonomous vehicles. We propose a taxonomy of these methods, discuss each subfield, and highlight open problems and future directions. Our objective is to provide a clear and concise overview of the state of the art in this field and to offer insights into the effectiveness of scenario-based testing approaches. By doing so, we aim to facilitate the identification of critical scenarios and the assessment of risk exposure frequencies, which are essential for enhancing the safety and reliability of autonomous vehicles.

Ride-pooling significantly reduces traffic congestion by enhancing fleet utilization through effective ride-matching. Real-world ride-pooling systems are dynamic, with fluctuations in driver availability and demand throughout the day. This necessitates adaptive ride-matching strategies that can quickly adjust to changing proximities and identify new carpooling opportunities by recalculating driver-rider correlations. However, most current methods primarily focus on static demand-supply scenarios and short-term accessibility, falling short in dynamic environment. In this study, we introduce a dynamic heterogeneous network model that captures the evolving nature of ride-pooling systems, where new requests and carpooling arrangements continuously emerge. We propose an embedding model-based matching decision process that operates online, adjusting to changes in the network’s structure. This process involves constructing a dynamic heterogeneous ride-pooling network that encompasses diverse node attributes and driver-rider connections, updating these representations to reflect the network’s evolution, and quickly identifying and ranking candidate riders for efficient online matching. Our approach demonstrates improved performance in offline evaluations using datasets from Austin, TX (RideAustin) and Chengdu, China (DiDi Chuxing). We observe a reduction in the necessary fleet size as new orders are placed, and an improvement in drivers’ matching probability compared to existing methods (e.g., an increase of 5.4–31.1% in the assignment rate on DiDi dataset), showcasing the advantage of employing dynamic network embedding to cut down on matching time (e.g., a decrease of 3.7–228.8 seconds in running time on DiDi dataset). Furthermore, we develop a simulated ride-pooling system (SRPool) that mimics dynamic demand-supply fluctuations and supports vehicle routing, providing a robust platform for evaluating ride-matching strategies. Our strategy not only excels in the SRPool environment but also effectively minimizes the total trip distance and rider waiting times.

In the Vehicle Routing Problem with Time Windows (VRPTW), dealing with demand and travel time uncertainties poses significant challenges in terms of model efficacy, service expenditures, and customer satisfaction, notably in extensive operational scenarios. In this paper, we introduce VRPU as a solution to the robust VRPTW encompassing uncertain demands and travel time. VRPU adopts a robust optimization approach and, drawing on budget uncertainty theory, formulates uncertainty variables and reconceptualizes the arrival times at customers as a complex recursive function. The model is addressed through a two-stage Adaptive Large Neighborhood Search heuristic (ALNS), with the primary stage focused on minimizing the total number of utilized vehicles and the subsequent stage dedicated to minimizing total travel costs.Compared to the current state-of-the-art methods, the results of our two-stage Adaptive Large Neighborhood Search (ALNS) algorithm, tested on the Solomon and Gehring & Homberger benchmark datasets, demonstrate that, while incurring a certain computational overhead, it achieves more robust solutions. Additionally, when applied to the deterministic Vehicle Routing Problem with Time Windows (VRPTW), our algorithm successfully identifies the known optimal solutions within the current benchmark datasets.

As one of the fundamental tasks of autonomous driving, depth perception aims to perceive physical objects in three dimensions and to judge their distances away from the ego vehicle. Although great efforts have been made for depth perception, LiDAR-based and camera-based solutions have limitations with low accuracy and poor robustness for noise input. With the integration of monocular cameras and LiDAR sensors in autonomous vehicles, in this article, we introduce a two-stream architecture to learn the modality interaction representation under the guidance of an image reconstruction task to compensate for the deficiencies of each modality in a parallel manner. Specifically, in the two-stream architecture, the multi-scale cross-modality interactions are preserved via a cascading interaction network under the guidance of the reconstruction task. Next, the shared representation of modality interaction is integrated to infer the dense depth map due to the complementarity and heterogeneity of the two modalities. We evaluated the proposed solution on the KITTI dataset and CALAR synthetic dataset. Our experimental results show that learning the coupled interaction of modalities under the guidance of an auxiliary task can lead to significant performance improvements. Furthermore, our approach is competitive against the state-of-the-art models and robust against the noisy input. The source code is available at https://github.com/tonyFengye/Code/tree/master .

The trend of progressive technology scaling makes the computing system more susceptible to soft errors. The most critical issue that soft error incurs is silent data corruption (SDC) since SDC occurs silently without any warnings to users. Estimating SDC probability of a program is the first and essential step towards designing protection mechanism. Prior work suffers from prediction inaccuracy since the proposed heuristic-based models fail to describe the semantic of fault propagation. We propose a novel approach SLOGAN which transfers the prediction of SDC probability into a graph regression task. A program is represented in the form of dynamic dependence graph. To capture the rich semantic of fault propagation, we apply structured graph attention network, which includes node-level, graph-level and layer-level self-attention. With the learned attention coefficients from node-level, graph-level, and layer-level self-attention, the importance of edges, nodes, and layers to the fault propagation can be fully considered. We generate the graph embedding by weighted aggregation of the embeddings of nodes and compute the SDC probability by the regression model. The experiment shows that SLOGAN achieves higher SDC accuracy than state-of-the-art methods with a low time cost.

In ride matching, the search results can be personalized for a particular driver. Given a query with trip plans, it is advantageous to rank potential riders in terms of who are most appealing to the driver for increasing occupancy rates. While personalized ranking approaches such as collaborative filtering and factorization are available, they are not suitable for pooling because candidate riders are associated with different preferences, and their travel is sparsely distributed with a long tail of users for a few popular destinations. The user embedding method is a good candidate in terms of alleviating data sparsity, but it has issues such as difficulty encoding user preferences from rich information. In this study, we explore user embedding techniques for the purposes of short-term personalized rider ranking, where the aim is to present to drivers a set of potential riders who share similar itineraries with them and can be picked up on their current route. Considering trip requests, along with the preferences issued in advance, this study uses attribute representations to rank the riders based on the higher-order similarities in the participants’ itineraries in a three-step manner: (i) start with a distributed representation of the riders’ preference regarding the cost of extra distance, (ii) generate user embeddings in a heterogeneous network with the meeting points and associated waiting times, and (iii) match and rank riders for drivers depending on an attribute fusion operation by adopting a personal route and schedule. Our proposed method performs well in an offline estimation on a huge dataset from DiDi in Chengdu, China. Experimental results indicate that with the learned embeddings, we can obtain statistically significant advancements (e.g., 4.6–29.5% increase in mean reciprocal rank (MRR); 2.8–17.4% in normalized discounted cumulative gain (nDCG)) over current methods for pooling ranking. Furthermore, we implement the proposed method on our simulated pooling system. These results validate that personalized ranking can undoubtedly boost the number of trips served, and reduce the total trip distance and waiting time.

Soft errors can lead to silent data corruption (SDC), seriously compromising the reliability of a system. To detect SDC, a profiling of SDC-causing instructions is usually needed to decide which instructions to protect. Current approaches gain SDC-causing instructions by using machine learning algorithms. Most of existing algorithms suffer from a lack of accuracy. Researchers choose certain structural features as input based on their understanding of fault propagation. Such hand-tuned features prevent models from reproducing the reasoning of fault propagation and that, in turn, limits their ability to make good prediction decisions. We propose GATPS, which is a Graph Attention neTwork to Predict SDC-causing instructions. The task of SDC prediction is converted into node classification in a heterogenous graph, which applies different types of edges to represent different instruction relations. Low dimensional embedding of each node is computed by attending over its neighbors through multiple types of edges. The hidden structural features related to SDC propagation can be captured automatically. To quantify fault effects between instructions, attention mechanism is applied to assign different importance to nodes of a same neighborhood. Experimental results show that GATPS improves F1 score of SDC prediction by 11.0 %-21.7 % over most competitive machine learning methods.

In this paper, we aim to provide an optimal passenger matching solution by recommending ridesharing groups of passengers from GPS trajectories. Existing algorithms for rider grouping usually rely on matching pre-selected origin-destination coordinates. Unfortunately, the semantics in the spatial layout (e.g., social interactions and properties of the locations) are ignored, leading to inaccuracies in discovering the ridesharing groups. Meanwhile, the destinations manually entered by users impact the accuracy of matching, as these addresses are usually not available in a road network or are not optimal for passenger pickup. This is particularly true when a passenger travels in a less familiar place. Given a set of passengers and the distribution of their destination, our approach is to compute the ridesharing matching between passengers. The raw GPS trajectories can be characterized by a combination of time constraints, traffic environments, and social activities. We first developed a PrefixSpan-prediction using a partial matching (P-PPM) destination-prediction algorithm to mine the frequent movement patterns from the trajectory data and determine the confidence of the movement rules. Our method uses the total travel time as the matching objective. Our approach is superior to the baseline methods in terms of accuracy (increased from 46% to 80%). We have also achieved significant improvements on other metrics, such as users' saved travel distance. We demonstrated that using our proposed method, a group of passengers could save over 19% of total travel miles, which shows that the ridesharing scheme could be effective.

Ridesharing has attracted increasing attention in recent years, and combines the flexibility and speed of private cars with the reduced cost of fixed-line systems to benefit alleviating traffic pressure. A major issue in ridesharing is the accurate assignment of passengers to drivers, and how to maximize the number of rides shared between people being assigned to different drivers has become an increasingly popular research topic. There are two major challenges facing ride-matching: scalability and sparsity. Here, we show that network embedding drives the optimal matches between drivers and riders. Contrary to existing approaches that merely depend on the proximity between passengers and drivers, we employ a heterogeneous network to learn the latent semantics from different choices in two types of ridesharing, and extract features in terms of user trajectories and sentiment. A novel framework for ridesharing, RShareForm, which encodes not only the objects but also a variety of semantic relationships between them, is proposed. This article extends the existing skip-gram model to incorporate meta-paths over a proposed heterogeneous network. It allows diverse features to be used to search for similar participants and then ranks them to improve the quality of ride-matching. Extensive experiments on a large-scale dataset from DiDi in Chengdu, China show that by leveraging heterogeneous network embedding with meta paths, RShareForm can significantly improve the accuracy of identifying the participants for ridesharing over existing methods, including both meta-path guided similarity search methods and variants of embedding methods.