Fusion-based 3D object detection is critical for autonomous driving. However, existing multimodal fusion methods often suffer from an insufficient receptive field, inefficient cross-modal interaction, and weak generalization to rare or irregular objects. We present BEV-SA, a novel bird's eye view-based fusion framework that overcomes these limitations through two key modules. The state-space BEV interaction (SSB) module serializes BEV features along a Hilbert curve and models them using a structured state space duality (SSD) mechanism, enabling global context modeling, precise modality alignment, and linear computational complexity. The accelerated separable fusion (ASF) module employs depthwise separable convolutions to efficiently merge aligned BEV features, reducing inference time by 12.5% compared to conventional fusion without sacrificing accuracy. We also introduce the anomalous-shaped commercial vehicle (ASC) dataset, a benchmark focusing on large and irregular commercial vehicles in challenging environments. Experiments on the nuScenes benchmark show that BEV-SA achieves state-of-the-art performance, improving NDS by 0.24% and mAP by 2.73%, while maintaining high efficiency. On ASC, BEV-SA demonstrates remarkable generalization to rare and irregular objects.

Eco-driving at signalized intersections must balance energy efficiency, traffic efficiency, and safety under dynamic signal and traffic constraints. Although deep reinforcement learning (DRL) has shown promise for this task, policies trained in nominal environments may lose reliability when unseen signal timings or abrupt preceding-vehicle maneuvers expose experience-sparse boundary states. Under the conventional DRL train-and-test paradigm, such evaluation failures are usually recorded as terminal outcomes rather than reused for policy improvement. To address this missing data loop, this paper proposes a closed-loop offline policy improvement framework that recycles evaluation trajectories for successor-policy refinement. Specifically, safe and failure trajectories from a predecessor policy are collected into a mixed-quality offline dataset and reused through offline reinforcement learning, enabling value-guided improvement while regularizing policy updates toward data-supported actions. Results are evaluated under randomized signal-timing scenarios and an emergency-braking case. The proposed framework improves safety robustness over conventional DRL baselines and matches or exceeds the safety performance of a safety-constrained DRL baseline, with less conservative and less reactive behavior. The refined policy attains 6.65 ± 0.71 kWh/100 km, the lowest among all deployable methods, corresponding to 92.7% of the Dynamic Programming based theoretical reference computed under ideal information conditions. Vehicle-in-the-loop tests further indicate that smoother longitudinal behavior remains observable after real execution dynamics are introduced.

Driven by low-carbon policies, the fuel cell (FC) vehicle industry is embracing rapid development opportunities. This study focuses on commercial vehicle FC systems to conduct power optimization research under variable altitude environments, aiming to resolve the significant net power attenuation issue in altitude-changing scenarios for commercial vehicles. Models for the FC air supply system and atmospheric environment were established, revealing power attenuation with altitude. Analysis characterized compressed air temperature and intercooler thermal loss across altitudes, constructing a four-dimensional mapping of load, air supply, stack performance, and parasitic power. Targeting maximum net system power, an optimal offline algorithm determined ideal air supply parameters under varying altitudes. A dual-degree-of-freedom hierarchical online optimization architecture was designed, proposing an altitude-adaptive strategy integrating offline calibration with online correction. The results show that compared with conventional strategies, 2.02% to 9.07% power enhancement at 4000m altitude. Under custom variable-altitude conditions, the proposed strategy achieves 1.53% to 3.29% power optimization superiority over comparative methods. Finally, the real-time performance of the optimization strategy was validated through hardware-in-the-loop (HIL) testing platform experiments.

Accurate development, optimization, and validation of vehicle control systems — including energy management systems (EMS) — rely on driving cycles that faithfully reproduce real-world operating conditions. Multidimensional cycles — incorporating variables such as yaw rate and road grade — capture critical dynamics beyond conventional speed profiles but lack both rigorous definitions of representativeness and reliable generation methods. This absence of structurally representative cycles can reduce the effectiveness of EMS optimization and result in significant gaps between simulation-based evaluations and actual in-vehicle performance.To address this issue, we propose CyGen-SAC, a data-driven framework for generating multidimensional driving cycles via deep reinforcement learning (DRL). At its core, a representativeness metric is constructed from a selected set of bivariate joint distributions, which preserve the key dependencies of the six-dimensional data and capture the joint relationships observed in real-world driving. Embedded as a structured reward, this metric provides dense feedback to a Soft Actor-Critic (SAC) agent, enabling the generation of driving cycles that are both distributionally aligned and highly representative.Evaluated on real-world driving data, CyGen-SAC demonstrates significant gains in distributional alignment, achieving a 6.6% improvement for the best single cycle and 23.4% across the generated set relative to a baseline method. For power-demand characteristics on a representative cycle, the average error relative to aggregate statistics from real-world measurements decreases from 43.2% to 5.8%. This improvement enhances the credibility of EMS evaluation and can be readily transferred to other control-oriented time-series tasks under structural constraints.

Accurate power demand estimation is essential for improving vehicle control precision, optimizing power distribution, and enhancing energy management, with torque estimation being a critical component. However, the ideal longitudinal dynamics model ignores the actual running state of the vehicle, resulting in a large error between the calculated and actual values. To address this issue, this study proposes a data-driven real-time torque estimation framework that balances accuracy and computational efficiency for new energy buses. Key features include an online wheel radius identification approach using the least squares method, a speed-slip compensation strategy based on the extreme learning machine, and a dynamic road slope correction technique. Validation using real-world driving data demonstrates that the proposed method reduces vehicle speed RMSE by 58.45

Energy management is crucial for improving fuel cell bus efficiency, yet learning-based strategies may become unstable when passenger-load-induced mass variations change traction power demand. This paper proposes a mass-adaptive intelligent energy management strategy without explicit online mass estimation. A joint state-augmentation and reward-shaping mechanism parameterizes the deviation between current traction-power demand and the empty-mass baseline, guiding the agent to adjust the fuel-cell/battery power split to compensate for this mass-induced power gap, so that a single policy remains adaptive across different vehicle masses. An embedded attention module identifies key state variables and visualizes decision priorities under different driving conditions. Simulation results show that the proposed method achieves energy savings of 6.85% and 6.47% compared with the baseline strategy while maintaining stable SoC. Furthermore, the approach shows strong algorithmic generalization within both DDPG and SAC frameworks, achieving adaptive response to mass variations and generalized transfer of energy management strategies.

Intelligent eco-driving in the Internet of Vehicles (IoV) environment, facilitated by the integration of IoV information and deep learning, has expanded the potential for optimizing the transportation system. Studying how intelligent eco-driving vehicles (IEDVs) interact with traditional vehicles in mixed traffic flow remains a challenging research topic. This paper aims to analyze the impact of IEDV penetration rate on traffic system operation and explore the behavioral interaction mechanisms between heterogeneous vehicles. Accordingly, a reinforcement learning-based eco-driving strategy combining behavior cloning and invalid action masking is proposed. By analyzing the behavioral disturbances exerted by IEDVs on surrounding vehicles, this paper investigates the macroscopic energy consumption and travel efficiency of traffic flow under different penetration levels. Notably, further upgrading and training of intelligent agents become extremely difficult in scenarios fully occupied by highly competitive IEDVs, so the research scope of penetration rate in this work is also restricted by the feasibility of subsequent agent optimization. The results demonstrate that the developed strategy enables IEDVs to achieve 47.61% energy savings with only 4.74% loss in travel timeliness compared with conventional vehicles. As IEDV penetration rises, the overall traffic flow energy consumption gradually declines. Nevertheless, excessively high penetration will trigger traffic system hysteresis and further degrade the comprehensive operational efficiency of the entire traffic flow.

Deep reinforcement learning has shown strong potential for the development of energy management strategies (EMSs) for fuel cell vehicles (FCVs). However, an EMS trained on fixed driving-cycle data may lose fuel economy competitiveness under evolving driving conditions. This suggests that EMS optimization should be revisited from a life-cycle perspective, particularly when the deployed EMS encounters driving cycles that cause severe performance degradation. A direct re-training strategy, however, may damage previously learned control knowledge and compromise performance under other conditions. To address this issue, this study proposes a generative model-enabled lifelong reinforcement learning (GM-LRL) framework for FCV energy management. A baseline soft actor-critic-based EMS is first developed on the standard cycle. The deployed EMS is then periodically diagnosed using a dynamic-programming-referenced relative optimality-gap criterion to identify the cycle on which the deployed EMS exhibits severe performance degradation. Based on this cycle, a physics-informed diffusion model is introduced to generate additional driving cycles, and an elastic-weight-consolidation-based lifelong re-training strategy is developed to improve adaptability while preserving previously acquired control knowledge. Experiments on a fuel cell heavy-duty vehicle show that the proposed method reduces the optimality-gap of origin EMS on the identified cycle by 6.19 percentage points. Moreover, it achieves an average fuel economy of 95.20% over 100 collected driving cycles, exceeding origin EMS and direct re-training by 3.14 and 3.21 percentage points, respectively. These results demonstrate that the proposed life-cycle optimization framework can both repair weak-performance regions and sustain robust EMS performance during long-term operation.

Reinforcement learning (RL) is a promising approach for end-to-end autonomous driving, but its practical deployment remains challenging due to low sample efficiency and sensitivity to reward design. To address these challenges, this study presents a novel Q-advantage integrated human-guided reinforcement learning (QIHG-RL) framework that effectively combines the strengths of machine learning and human expertise. The QIHG-RL framework features: 1) an ensemble Q-advantage function that aggregates multiple value networks to enhance value estimation, and 2) an integration mechanism that embeds the Q-advantage into both the actor-critic network and the prioritized experience replay. This design allows the agent to leverage sparse and sub-optimal human demonstrations, accelerating policy learning in the early training phase while gradually enhancing exploration as training progresses. The framework is evaluated across three safety-critical driving tasks. Experimental results show a 167% improvement in sample efficiency compared to standard RL methods and a 14% performance gain over a state-of-the-art human-guided RL baseline. Furthermore, a Sim2Real pipeline combining domain randomization and semantic denoised remapping facilitates successful deployment on a real-world autonomous vehicle.

The development of hydrogen fuel cell (FC) commercial vehicles is crucial for advancing FC industrialization. However, the reliability limitations of FC systems under variable altitude conditions have hindered widespread use in plateau transportation. This study proposes a coordinated control method combining adaptive cathode surge suppression and optimized anode purging aimed at enhancing the reliability of the FC gas supply system under variable altitude conditions. A model-adaptive updating strategy is adopted to design a model predictive control (MPC) approach for cathode gas flow and pressure regulation, integrating a dual-layer surge discrimination mechanism to prevent surge in altitude-gradient air compressors under extreme conditions. For anode pressure and flow regulation, a super-twisting algorithm-based control method is implemented. An entropy-weighted TOPSIS framework is employed to optimize purge strategies within a multiindex evaluation system. Under the China Heavy-duty commercial vehicle test cycle (CHTC) condition, the maximum control errors for oxygen excess ratio (OER) and cathode pressure demonstrate 19.88% and 10.72% reductions, respectively, compared with sliding mode control (SMC). The absolute error of anode pressure at 4000 m altitude decreases by 4.21% per cycle relative to plain-area strategies. Finally, hardware-in-the-loop experiments validate the method’s feasibility.

The energy-efficient driving of Distributed Drive Electric Vehicles (DDEVs) presents a complex optimization challenge: exploiting over-actuated torque distribution capabilities to minimize energy consumption while strictly adhering to high-dimensional safety constraints, specifically individual tire slip ratios. Existing Deep Reinforcement Learning (DRL) methods often struggle to balance these conflicting objectives, typically relying on scalar penalties that fail to guarantee safety in dynamic scenarios. This paper proposes a contribution-oriented framework, LagCriticSAC, which introduces a novel dual-critic architecture to decouple safety assessment from reward maximization. Unlike standard Lagrangian approaches that rely on global cost signals, our method employs dedicated cost-critics to provide state-aware predictions of long-term constraint violations for each wheel. This mechanism dynamically adjusts learnable Lagrange multipliers, creating a closed-loop feedback system that autonomously shifts between risk-averse and efficiency-priority modes. Experiments demonstrate that this approach not only achieves superior energy efficiency, outperforming Soft Actor-Critic (SAC) and Model Predictive Control (MPC) benchmarks by 6.2% and 12.4% respectively, but also strictly enforces per-wheel longitudinal slip constraints within ±0.1, ensuring operational safety even in high-density traffic environments.

In high-temperature environments, thermal management systems are essential to improving comfort, ensuring the safe operation of electric vehicles, and reducing energy consumption. This study introduces a multi-agent control framework that leverages the Deep Deterministic Policy Gradient (MADDPG) algorithm to optimize temperature regulation and energy efficiency in an integrated thermal management system (ITMS). In recognition of the variability of air conditioning (AC) system energy efficiency under different vehicle operating conditions, the coordination agent dynamically plans target temperature trajectories for the cabin and battery, whereas the execution agent follows these trajectories and regulates motor system temperatures, ensuring coordinated energy optimization. To enhance the robustness and mitigate the risks of local optima caused by hyperparameter sensitivity, the Cross-Entropy Method (CEM) is integrated into MADDPG, forming the CEM-MADDPG control strategy. This hybrid method combines the global search capability of evolutionary algorithms with the sample efficiency of deep reinforcement learning, achieving a balance between stability and efficiency. Experimental results show that the proposed method effectively controls the temperatures of the battery, cabin, and motor, reducing the overall energy consumption of the ITMS by 19.8% and 7.2% respectively, compared to rule-based and model predictive control (MPC) methods.

Abstract Dynamic programming (DP) serves as the optimal benchmark for the energy management of plug-in hybrid electric vehicles (PHEVs). However, its high computational cost and the complexity of the resulting state-of-charge (SOC) trajectory prevent its direct application in real-time control, creating a barrier between theoretical optimality and engineering practice. To address this challenge, this paper proposes a Trend-Aware Piecewise Simplification (TAPS) method, designed to distill the core strategic knowledge from the DP solution and translate it into an executable reference. The TAPS method first employs a change-point detection algorithm to automatically segment the SOC trajectory into several distinct phases. Subsequently, a trend-based percentile assignment approach sets a constant SOC target for each phase, thereby preserving the macroscopic intent of the original strategy. Simulation validation on a high-fidelity PHEV model demonstrates that the TAPS method achieves significantly better fuel economy than conventional rule-based strategies and the superiority of the proposed method is confirmed by statistical significance analysis. Furthermore, compared to a strategy that directly tracks the original DP trajectory, TAPS avoids the performance degradation caused by following high-frequency signals, verifying its necessity for practical engineering applications. This research provides an effective pathway for converting complex offline optimal solutions into practical online strategies.

In the development of fuel cell hybrid electric trucks, the energy management strategy (EMS) plays a key role in enhancing system efficiency and operational economy. The objectives of the EMS include not only minimizing hydrogen consumption but also extending the powertrain system's lifespan to support long-term sustainable commercial operation. This paper proposes a multi-objective energy management framework that integrates Dynamic Programming (DP) and the Soft ActorCritic (SAC) algorithm. Specifically, DP is utilized offline to generate optimal energy allocation trajectories, which are then used to initialize the SAC replay buffer and thereby warm start the training process. Furthermore, a multi-objective reward function is designed to jointly optimize hydrogen consumption and the degradation of both the fuel cell and lithium-ion battery systems. Simulation results show that the proposed strategy incurs only 2.86% higher total cost compared to the DP baseline, while achieving $\mathbf{5. 1 1 \%}$ and $\mathbf{2 3. 7 4 \%}$ lower costs than the standard SAC and rule-based strategies, respectively. In addition, the training time is reduced by 50.64% compared to the standard SAC method, confirming the effectiveness of the pretrained SAC approach in accelerating convergence and delivering highperformance energy management.

Fuel Cell Hybrid Electric Vehicles (FCHEVs) offer a promising path toward sustainable transportation, but their operational economy and component durability are highly dependent on the energy management strategy (EMS). Conventional deep reinforcement learning (DRL) approaches to EMS often suffer from training instability and are typically reactive, failing to leverage predictive information such as upcoming road topography. To overcome these limitations, this paper proposes a proactive, slope-aware EMS based on an expert-guided DRL framework. The methodology integrates a rule-based expert into a Soft Actor-Critic (SAC) algorithm via a hybrid imitation–reinforcement loss function and guided exploration, enhancing training stability. The strategy was validated on a high-fidelity FCHEV model incorporating component degradation. Results on the dynamic Worldwide Harmonized Light Vehicles Test Cycle (WLTC) show that the proposed slope-aware strategy (DRL-S) reduces the SOC-corrected overall operating cost by a substantial 14.45% compared to a conventional rule-based controller. An ablation study confirms that this gain is fundamentally attributed to the utilization of slope information. Microscopic analysis reveals that the agent learns a proactive policy, performing anticipatory energy buffering before hill climbs to mitigate powertrain stress. This study demonstrates that integrating predictive information via an expert-guided DRL framework successfully transforms the EMS from a reactive to a proactive paradigm, offering a robust pathway for developing more intelligent and economically efficient energy management systems.

End-to-end navigation strategies using reinforcement learning (RL) can improve the adaptability and autonomy of Autonomous ground vehicles (AGVs) in complex environments. However, RL still faces challenges in data efficiency and safety. Neuroscientific and psychological research shows that during exploration, the brain balances between fear and curiosity, a critical process for survival and adaptation in dangerous environments. Inspired by this scientific insight, we propose the "Feariosity" model, which integrates fear and curiosity model to simulate the complex psychological dynamics organisms experience during exploration. Based on this model, we developed an innovative policy constraint method that evaluates potential hazards and applies necessary safety constraints while encouraging exploration of unknown areas. Additionally, we designed a new experience replay mechanism that quantifies the threat and unknown level of data, optimizing their usage probability. Extensive experiments in both simulation and real-world scenarios demonstrate that the proposed method significantly improves data efficiency, asymptotic performance during training. Furthermore, it achieves higher success rates, driving efficiency, and robustness in deployment. This also highlights the key role of mimicking biological neural and psychological mechanisms in improving the safety and efficiency through RL.

Learning-based techniques hold considerable promise in achieving human-like autonomous driving. However, one deployed policy encounters difficulties in satisfying the drivers’ diverse decision-making preferences simultaneously. Meanwhile, training personalized policies for each driver from scratch is time-consuming and resource-intensive. To address these challenges, this paper proposes a human-guided continual learning framework, wherein the human drivers could real-time take over a deployed policy when it performs unsatisfactorily, and the autonomous vehicle (AV) agent would automatically acquire human demonstrations and dynamically alter itself in accordance with personalized decision-making preference. Furthermore, a priority experience memory-enabled elastic weight consolidation (PEM-EWC) mechanism is developed to prevent the AV agent from overfitting to a limited number of human demonstrations and catastrophically forgetting its acquired fundamental driving abilities. Driver-in-the-loop simulations and real-world experiments are conducted in representative autonomous driving decision-making scenarios, and experimental results demonstrate the superior equilibrium of our proposed approach in terms of driving safety, human likeness, and training efficiency, compared to other baselines, which suggests that it provides a promising solution for personalized decision-making in autonomous driving. The supplementary video is available at https://youtu.be/HKF0ayxMycc.

Exploring the full potential of energy savings for new energy vehicles in a future connected transportation system is a challenging task. To address how connected buses can leverage surrounding traffic information to improve their energy efficiency, an intelligent fuel cell bus energy management method based on traffic scenario frozen callback is proposed, which enables high real-time performance in online energy management. To tackle the issue of inconsistent data dimensions caused by random fluctuations in the number of vehicles in a fixed traffic flow, a traffic flow representation based on grid grayscale images is designed. Building upon this representation, a speed trajectory prediction model based on traffic scenario frozen callback is developed. Subsequently, offline historical global optimal data are used to construct a training dataset that links speed trajectories to optimal control sequences. An end-to-end energy management framework based on the adaptive neuro-fuzzy inference system (ANFIS) is presented and validated in scenarios that before entering bus station and after exiting bus station. Simulation results demonstrate that, the proposed energy management strategy (EMS) approaches the overall energy consumption of dynamic programming (DP), reaching 97.76 % and 98.82 % in the two kinds of scenarios of its performance, outperforms the other two comparative EMSs. In terms of timeliness, the computational time spent by the proposed EMS is only 0.2076 times and 0.1952 times that of traditional model predictive control (MPC)-based EMS in the separate scenario.

The reinforcement learning (RL) is a powerful framework with significant potential to enhance the autonomous driving (AD) performance. However, its trial-and-error nature presents significant hurdles in terms of safety, efficiency, and stability. To address these challenges, we propose an adaptive rollout model-based policy optimization (AR-MBPO) algorithm tailored for car-following motion planning in autonomous electric vehicles (AEVs). The algorithm improves the overall performance by incorporating an error-aware ensemble environmental model and leveraging branched rollouts for efficient sample collection and policy optimization. A key innovation of AR-MBPO is an adaptive rollout mechanism that dynamically adjusts based on the predictive accuracy, mitigating the impact of model inaccuracies. In addition, the energy efficiency is explicitly integrated into the optimization process to minimize the energy consumption. We evaluate AR-MBPO through AEV car-following simulations, where it demonstrates superior performance, including rapid convergence, and reduced reliance on real-world interactions. The method simultaneously optimizes safety, traffic efficiency, and energy efficiency through dynamic distance adjustments, as evidenced by a 0% collision rate in testing scenarios and an 8.2% energy consumption reduction compared with nonenergy-aware baselines. The results suggest potential applications in AD systems for improved safety and energy efficiency.