Autonomous driving (AD) endows vehicles with the capability to drive partly or entirely without human intervention. AD agents generate driving policies based on online perception results, which are crucial to the realization of safe, efficient, and comfortable driving behaviors, particularly in high-dimensional and stochastic traffic scenarios. Currently, deep reinforcement learning (DRL) techniques to derive and validate AD policies have witnessed vast research efforts and have shown rapid development in recent years. However, a comprehensive interpretation and evaluation of their strengths and limitations concerning the full-stack AD tasks remain uncharted. This paper presents a survey of this body of work, which is conducted at three levels. First, it analyzes the multi-level AD task characteristics and delves deeply into the current DRL methodologies primarily employed in AD. Second, a taxonomy of the literature studies is constructed from the system perspective, identifying six modes of DRL model integration into an AD architecture that span the entire spectrum of AD policy processes, from perception understanding and decision-making to motion control, as well as verification and validation. Each literature review comprehensively encompasses the main elements of designing such a system, including modeling partially observable environments, state and action spaces, reward structuring, and the design and training methodologies of neural network models. Finally, an in-depth foresight is conducted on how the eight critical issues of AD application development are addressed by the DRL models tailored for real-world AD challenges.

Highly integrated information sharing among people, vehicles, roads, and cloud systems, along with the rapid development of autonomous driving technologies, has spurred the evolution of automobiles from simple “transportation tools” to interconnected “intelligent systems”. The intelligent cockpit is a comprehensive application space for various new technologies in intelligent vehicles, encompassing the domains of driving control, riding comfort, and infotainment. It provides drivers and passengers with safety, comfort, and pleasant driving experiences, serving as the gateway for traditional automobile manufacturing to upgrade towards an intelligent automotive industry ecosystem. This is the optimal convergence point for the intelligence, connectivity, electrification, and sharing of automobiles. Currently, the form, functions, and interaction methods of the intelligent cockpit are gradually changing, transitioning from the traditional “human adapts to the vehicle” viewpoint to the “vehicle adapts to human”, and evolving towards a future of natural interactive services where “humans and vehicles mutually adapt”. This article reviews the definitions, intelligence levels, functional domains, and technical frameworks of intelligent automotive cockpits. Additionally, combining the core mechanisms of human–machine interactions in intelligent cockpits, this article proposes an intelligent-cockpit human–machine interaction process and summarizes the current state of key technologies in intelligent-cockpit human–machine interactions. Lastly, this article analyzes the current challenges faced in the field of intelligent cockpits and forecasts future trends in intelligent cockpit technologies.

Autonomous Intersection Management (AIM) systems present a new paradigm for conflict-free cooperation of connected autonomous vehicles (CAVs) at road intersections, the aim of which is to eliminate collisions and improve the traffic efficiency and ride comfort. Given the challenges of current centralized coordination methods in balancing high computational efficiency and robust safety assurance, this paper proposes an innovative conflict-free management scheme for CAVs at unsignalized intersections, leveraging safe multi-agent deep reinforcement learning (MADRL). Firstly, we formulate the safe MADRL problem as a constrained Markov game (CMG) and then transform the AIM problem into a CMG by carefully designing state, action, reward, and cost functions. Subsequently, we propose the Multi-Agent Constrained Policy Optimization (MACPO), specifically tailored to solve the CMG problem. MACPO incorporates safety constraints that further restrict the trust region formed by the Kullback-Leibler (KL) divergence, facilitating reinforcement learning policy updates that maximize performance while keeping constraint costs within their limit bounds. This leads us to introduce the MACPO-based AIM Algorithm. Finally, we train an AIM policy and compare its computation time, ride comfort, traffic efficiency, and safety with management schemes based on Model Predictive Control (MPC), Mixed Integer Programming (MIP), and non-safety-aware reinforcement learning. According to the results, compared with the MPC and MIP methods, our method has increased computational efficiency by 65.22 times and 731.52 times respectively, and has improved traffic efficiency by 2.41 times and 1.80 times respectively. In contrast to the non-safety awareness RL methods, our method achieves a zero collision rate for the first time, while also enhancing ride comfort, highlighting the advantages of using MACPO.

End-to-end autonomous driving, which directly maps raw sensor inputs to low-level vehicle controls, is an crucial part of Embodied AI. Despite successes in applying Multimodal Large Language Models (MLLMs) for high-level traffic scene semantic understanding, it remains challenging to effectively translate these conceptual semantics understandings into low-level motion control commands and achieve cross-scene driving generalization and consensus. We propose Sce2DriveX, a human-like chain-of-thought (CoT) driving reasoning MLLM framework, designed to achieve progressive learning from multi-view scene understanding to behavior analysis, motion planning, and vehicle control driving process. Sce2DriveX utilizes multimodal joint learning of local scene videos and global Bird’s Eye View (BEV) maps to deeply understand long-range spatiotemporal relationships and road topology, enhancing its 3D dynamic/static scene perception and reasoning capabilities and achieving cross-scene generalization. Meanwhile, it reconstructs the implicit cognitive chain inherent in human driving, further enhancing the consensus between autonomous driving and human thought. To improve model performance, we construct the first comprehensive Visual Question Answering (VQA) driving instruction dataset, which tailored for 3D spatial understanding and long-axis task reasoning, and introduce a task-oriented three-stage training pipeline to support supervised fine-tuning. Extensive experiments demonstrate that Sce2DriveX achieves state-of-the-art performance across tasks from scene understanding to end-to-end driving, as well as robust generalization in handling diverse driving scenes on the CARLA Bench2Drive benchmark.

Human-level driving is the ultimate goal of autonomous driving. As the top-level decision-making aspect of autonomous driving, behavior decision establishes short-term driving behavior strategies by evaluating road structures, adhering to traffic rules, and analyzing the intentions of other traffic participants. Existing behavior decisions are primarily implemented based on rule-based methods, exhibiting insufficient generalization capabilities when faced with new and unseen driving scenarios. In this paper, we propose a novel behavior decision method that leverages the inherent generalization and commonsense reasoning abilities of visual language models (VLMs) to learn and simulate the behavior decision process in human driving. We constructed a novel instruction-following dataset containing a large number of image–text instructions paired with corresponding driving behavior labels, to support the learning of the Drive Large Language and Vision Assistant (DriveLLaVA) and enhance the transparency and interpretability of the entire decision process. DriveLLaVA is fine-tuned on this dataset using the Low-Rank Adaptation (LoRA) approach, which efficiently optimizes the model parameter count and significantly reduces training costs. We conducted extensive experiments on a large-scale instruction-following dataset, and compared with state-of-the-art methods, DriveLLaVA demonstrated excellent behavior decision performance. DriveLLaVA is capable of handling various complex driving scenarios, showing strong robustness and generalization abilities.

Adaptive cruise control (ACC) enables efficient, safe, and intelligent vehicle control by autonomously adjusting speed and ensuring a safe following distance from the vehicle in front. This paper proposes a novel adaptive cruise system, namely the Safety-First Reinforcement Learning Adaptive Cruise Control (SFRL-ACC). This system aims to leverage the model-free nature and high real-time inference efficiency of Deep Reinforcement Learning (DRL) to overcome the challenges of modeling difficulties and lower computational efficiency faced by current optimization control-based ACC methods while simultaneously maintaining safety advantages and optimizing ride comfort. Firstly, we transform the ACC problem into a safe DRL formulation Constrained Markov Decision Process (CMDP) by carefully designing state, action, reward, and cost functions. Subsequently, we propose the Projected Constrained Policy Optimization (PCPO)-based ACC Algorithm SFRL-ACC, which is specifically tailored to solve the CMDP problem. PCPO incorporates safety constraints that further restrict the trust region formed by the Kullback–Leibler (KL) divergence, facilitating DRL policy updates that maximize performance while keeping safety costs within their limit bounds. Finally, we train an SFRL-ACC policy and compare its computation time, traffic efficiency, ride comfort, and safety with state-of-the-art MPC-based ACC control methods. The experimental results prove the superiority of the proposed method in the aforementioned performance aspects.