Deep learning (DL) has shown state-of-the-art performance in trajectory prediction, which is critical to safe navigation in autonomous driving (AD). However, most DL-based methods suffer from catastrophic forgetting, where adapting to a new distribution may cause significant performance degradation in previously learned ones. Such inability to retain learned knowledge limits their applicability in the real world, where AD systems need to operate across varying scenarios with dynamic distributions. As revealed by neuroscience, the hippocampal circuit plays a crucial role in memory replay, effectively reconstructing learned knowledge based on limited resources. Inspired by this, we propose a hippocampal circuit-inspired continual learning method (H2C) for trajectory prediction across varying scenarios. H2C retains prior knowledge by selectively recalling a small subset of learned samples. First, two complementary strategies are developed to select the subset to represent learned knowledge. Specifically, one strategy maximizes inter-sample diversity to represent the distinctive knowledge, and the other estimates the overall knowledge by equiprobable sampling. Then, H2C updates via a memory replay loss function calculated by these selected samples to retain knowledge while learning new data. Experiments based on various scenarios from the INTERACTION dataset are designed to evaluate H2C. Experimental results show that H2C reduces catastrophic forgetting of DL baselines by 22.71% on average in a task-free manner, without relying on manually informed distributional shifts. The implementation is available at https://github.com/BIT-Jack/H2C-lifelong.

With the rapid development of intelligent connected vehicles, the domain-based electronic and electrical (E/E) architecture is providing the potential upgrade for the autonomous vehicle. As a representative, the domain-centralized E/E architecture can be installed in the autonomous vehicle which performs its powerful software updates, cabling reduction, and functional integration. For the efficient and stable motion control of autonomous vehicles equipped with domain-centralized E/E architecture, this paper proposes an improved deep reinforcement learning framework based on multi-hops loop delay and accelerated gradient optimization. Firstly, the domain-centralized E/E architecture and motion control problem of autonomous vehicle are modeled, respectively. Then, a multi-hops loop delay analysis is carried out for the E/E architecture to estimate the theoretical boundary value of heterogeneous topology loop delay. Subsequently, the deep reinforcement learning algorithm using modified heuristic experience replay is developed for the motion control of autonomous vehicle equipped with domain-centralized E/E architecture. In the implementation of deep reinforcement learning system, the estimated loop delay value is integrated into the motion controller optimization, and the Nesterov accelerated gradient is introduced and combined with the adaptive moment estimation to improve the optimization effect. Finally, the real-world scenarios and virtual driving environment simulation are applied to evaluate the performance of the improved deep reinforcement learning framework. The results show that the proposed motion control framework achieves better performance and guarantees the stability to the loop delay caused by domain-centralized E/E architecture.

With the increasing computational demands of intelligent connected vehicles (ICVs), traditional roadside mobile edge computing (MEC) faces resource and coverage limitations. As a mobile computing entity, a vehicle platoon enables coordinated resource scheduling and low-latency communication, serving as an effective supplement to edge computing. However, limited onboard resources make vehicle platoons insufficient for handling heavy task loads. This study focuses on the joint task offloading and resource allocation between vehicle platoons and MEC servers. A MEC-assisted vehicle platoon computing network (MVPCN) is constructed, where both platoon vehicles (PVs) and MEC servers serve as computing nodes. We formulate a joint optimization problem of task offloading and resource allocation in the vehicle platoon, aiming to minimize the long-term weighted cost of time and energy consumption. Given the complexity of the problem and the limitations of conventional solution methods, we design a Discrete Two-Stage Stackelberg Game (DTSG) to decompose the problem into two subproblems: leader (platoon leader) level and follower (computing nodes) level, and prove the existence of a Stackelberg equilibrium (SE). Based on the game model, we propose a Nested Platoon Task Offloading and Resource Allocation (NPTORA) algorithm. The leader employs the multi-agent proximal policy optimization (MAPPO) to generate offloading decisions for the platoon, while the followers allocate computational resources using a Dynamic Priority-based Hybrid Optimization (DPHO) algorithm based on the received offloading decisions. Simulation results show that NPTORA outperforms baseline methods in task success, cost efficiency, and adaptability, demonstrating strong performance and practical potential.

As the Internet of Things proliferates, cloud-assisted mobile-edge computing (MEC) enables intelligent connected vehicles (ICVs) to offload their computationally intensive tasks to servers within the Internet of Vehicles, thereby reducing delay and energy consumption. However, most existing research on edge computing offloading overlooks the dependency relationships between subtasks. These dependencies significantly increase the complexity of task offloading, making it difficult to devise general solutions for scenarios of varying scales a challenging endeavor. To tackle this challenge, we present a heterogeneous graph attention network (HGAT) augmented deep reinforcement learning dependency-aware task offloading framework, aiming to achieve minimal task completion time and energy consumption. The dynamic system of vehicles and servers is modeled as an undirected graph, with nodes corresponding to servers/vehicles and edges capturing the intensity of task competition. Tasks are modeled as directed acyclic graphs, where nodes denote subtasks and directed edges define their dependencies. An HGAT-based encoder is then introduced to effectively capture the intricate relationships between subtasks and each servercores. Subtask selection and servercores assignment are formulated as a Markov decision process and solved using the proximal policy optimization method. Simulation results demonstrate that the proposed algorithm outperforms existing ones across various scenarios, showcasing superior adaptability and performance benefits.

Recently, graph reinforcement learning (GRL)-based methods have demonstrated superior performance in solving decision-making issues. However, existing GRL-based methods encounter challenges in adequately integrating driving environment information and accurately capturing vehicular interactions. Resolving these issues is crucial for enhancing the safety and efficiency of intelligent transportation systems. To address these challenges, this paper presents a novel multi-modal failure-aware graph representation (MM-FA-GR) approach. The objective is to enhance the completeness and stability of graph representation techniques in the GRL-based setting, thereby improving the decision-making performance of CAVs operating in mixed autonomy traffic. Initially, a risk assessment invariant feature extractor (RA-IFE) is introduced to efficiently and selectively aggregate vehicle driving features into the node feature matrix. Subsequently, a multi-modal interaction model (MIM) is developed to comprehensively represent the mutual effects among vehicles and construct a multi-dimensional adjacency matrix. Moreover, a dynamic failure model (DFM) is incorporated to assess the sensing and communication failures of CAVs, enhancing the model’s robustness in non-ideal driving environments. Finally, a GRL model is established to solve the optimized driving strategies for CAVs. The proposed MM-FA-GR method has large potential to advance the graph representation technology, thereby enhancing decision-making performance in mixed autonomy traffic and improving robustness in non-ideal driving conditions. Comprehensive experiments are conducted with three typical traffic scenarios. Results show that our proposed MM-FA-GR method outperforms several baselines regarding safety, efficiency, and stability, highlighting the effectiveness of the core components within the proposed method. Moreover, the quantitative experiment validates the generalization capability of the proposed method in addressing the decision-making challenges of CAVs across diverse scenarios and traffic densities.

High-level autonomous vehicles and domain-based electronic and electrical (E/E) architectures are important development directions of the intelligent automobile industry. The domain-centralized E/E architecture has become the potential upgrade to the autonomous vehicle benefitting from its powerful software updates, cabling reduction, and functional integration. Aiming at the efficient motion control of the autonomous vehicle equipped with domain-centralized E/E architecture, a novel control framework with algorithms improvement is proposed in this paper, which contains the multi-hops loop delay analysis to solve the control stability problem caused by the heterogeneous topology loop delay of domain-centralized E/E architecture. In this framework, the motion controller is generated through the combination of modified double reinforcement learning algorithm and multi-steps predictive control method, and the loop delay is integrated into the controller optimization. Through the virtual driving environment simulation and real world scenario, the results show that the proposed framework achieves better performance in terms of path tracking and obstacles avoidance, and the stability of control strategies to loop delay is also guaranteed.

Autonomous driving technology and platooning driving technology are important directions for the development of intelligent and connected vehicles. Aiming at the motion control problem of autonomous vehicle platoon, this article proposes a multilayer predictive control framework (MPCF) based on heuristic learning agent and improved distributed model. First, the leading autonomous vehicle and following heterogeneous vehicles are modeled, respectively, and the motion control problem of autonomous platoon is described. Then, the multilayer motion control framework is designed, which contains highly automated tracking control optimization for the leading vehicle (LV) and high-precision formation keeping optimization for the following vehicles (FVs). In the upper layer, the heuristic Dyna algorithm-based predictive control (HDY-PC) method is proposed to improve the path tracking performance of the LV. In the lower layer, the improved distributed model-based predictive control (IDM-PC) method is developed to guarantee the motion effectiveness and stability of the vehicle platoon. Besides, the multilayer control framework can handle various communication topologies and dynamic cut-in/cut-out maneuvers. The virtual environment simulation shows that the proposed motion control framework for heterogeneous autonomous vehicle platoon achieves better performance in path tracking and platoon keeping. The adaptability of the framework is also verified using another real-world scene.

In-vehicle time-sensitive networking (TSN) delivers highly secure, ultralow latency deterministic communication for intelligent connected vehicles (ICVs). To tackle the traffic scheduling problem of in-vehicle TSN, this study establishes in-vehicle network topologies and flow models, abstracts the traffic scheduling problem as a job-shop scheduling problem (JSSP), and formulates a priority-based optimization function capable of various end-to-end (E2E) delay requirements. A joint routing and scheduling optimization strategy based on improved grey wolf optimization (IGWO) is proposed, which incorporates acrlong LF, historical experience learning, and acrlong TS operators to significantly enhance search capabilities and optimization efficiency. This strategy can rapidly solve large-scale in-vehicle network scheduling and generate scheduling results with outstanding delay performance. Dynamic routing that combines load-balanced and shortest path effectively minimizes interference between flows, further reducing E2E delay. Simulation experiments grounded in realistic ICV scenarios demonstrate the effectiveness of the proposed strategy. Furthermore, the simulation results verify the impact of flow period parameters and network topologies on E2E delay, offering guidance for in-vehicle TSN engineering design.

Time-sensitive networking (TSN) can offer deterministic low-latency communication, making it a critical solution for high-level autonomous vehicle's in-vehicle network. The deterministic transmission of TSN relies on TSN traffic scheduling. To ensure real-time transmission performance and vehicle functional safety, in-vehicle TSN scheduling aims to reduce end-to-end delay. Despite the promising potential of graph neural networks and deep reinforcement learning (DRL) in navigating complex TSN scheduling environments, its application has predominantly been limited to enhancing schedulability without a targeted focus on minimizing delays. This article introduces a DRL in-vehicle TSN scheduling framework based on the graph attention network (GAT). The scheduling problem is abstracted as a delay optimization problem and mapped to a Markov decision process (MDP), which is solved using the proximal policy optimization (PPO) algorithm. The GAT with attention mechanism is incorporated to extract critical information to enhance feature extraction and improve scheduling accuracy. This GAT-based PPO method can achieve high-precision offline scheduling through training, producing low-delay scheduling results. Simulation results demonstrate that the proposed method improves offline scheduling performance compared to other DRL-based scheduling methods. Leveraging the trained neural network, the proposed method can also deliver high robustness in online scheduling under link failure scenarios. It can produce a scheduling solution in just 3.8 s, and the scheduling results for all failure scenarios surpass those of rule-based benchmarking methods.

The autonomous vehicle is widely applied in various ground operations, in which motion planning and tracking control are becoming the key technologies to achieve autonomous driving. In order to further improve the performance of motion planning and tracking control, an efficient hierarchical framework containing motion planning and tracking control for the autonomous vehicles is constructed in this paper. Firstly, the problems of planning and control are modeled and formulated for the autonomous vehicle. Then, the logical structure of the hierarchical framework is described in detail, which contains several algorithmic improvements and logical associations. The global heuristic planning based artificial potential field method is developed to generate the real-time optimal motion sequence, and the prioritized Q-learning based forward predictive control method is proposed to further optimize the effectiveness of tracking control. The hierarchical framework is evaluated and validated by the numerical simulation, virtual driving environment simulation and real-world scenario. The results show that both the motion planning layer and the tracking control layer of the hierarchical framework perform better than other previous methods. Finally, the adaptability of the proposed framework is verified by applying another driving scenario. Furthermore, the hierarchical framework also has the ability for the real-time application.