Energy supervisory control (ESC) for electric flying vehicles is challenged by strong air–ground mode coupling and stringent safety constraints. Conventional reinforcement learning–based ESC typically requires extensive training and are difficult to bootstrap from scratch, which often leads to inefficient exploration and delayed satisfaction of safety constraints. To address this issue, we propose a large language model–enhanced curriculum learning warm start for soft actor–critic (CL-LSAC), which incorporates expert prior knowledge from large language models. The proposed curriculum prioritizes safety constraints before fuel efficiency, guiding policy learning through a highly non-convex reward landscape. Evaluated on a high-fidelity urban air mobility mission, CL-LSAC significantly suppresses the control saturation observed in unguided baselines. Compared with static reward weighting, our method improves state-of-charge regulation accuracy by 49% and accelerates convergence by more than 15% via LLM-guided initialization, yielding data-efficient and physically consistent policies that demonstrate high real-time execution feasibility on hardware-in-the-loop platforms.
As a key technology in intelligent transportation systems, speed guidance improves vehicle eco-driving in urban areas, yet its potential failures can cause suboptimal performance. As a macro-level evaluation metric for evaluating the effectiveness of speed guidance, the Guidance Effectiveness Rate exerts a crucial impact on the optimization performance of the speed guidance. To address this, the study establishes a research pipeline encompassing large-scale data acquisition, algorithm development, mechanism analysis, and strategy design. An ensemble machine learning model (MVE) was developed to classify individual trajectory points as effective/ineffective guidance points, and then accurately estimate the trip-level Guidance Effectiveness Rate (GER) from vehicle trajectory data. Trained and validated on experimental data, the MVE model achieved a recall of 0.89. This model was then applied to classify over 1.1 million trajectory data points from a real-world speed guidance system, and the subsequent trip-level statistical analysis revealed an average GER of 0.79. The analysis identified two primary scenarios for low GER: on good roads (speeds of 60-80 km/h, acceleration <0.3 m/s2) and in low-speed congestion (speeds of 15-40 km/h, acceleration >0.5 m/s2). Furthermore, the proportion of aggressive driving peaked at a GER of approximately 0.75. Most critically, vehicle energy consumption and CO2 emissions were found to be highest at low speeds (<20 km/h) and increased significantly within the GER band of 0.7 to 0.9. Consequently, it is recommended to restrict speed guidance applications in both free-flow and severely congested conditions to mitigate negative impacts. The research supports the practical application of speed guidance-type intelligent transportation systems.
Thermal management is critical for stationary battery energy storage systems (BESSs). This study proposes a prediction-based thermal management strategy for stationary BESSs, leveraging the operational predictability of BESSs to enable preemptive cooling regulation. Numerical investigations across four seasons demonstrate that, compared with the 60 s ON/60 s OFF intermittent control, the proposed prediction-based thermal management strategy reduces the overtemperature duration by approximately 80% under extreme summer conditions, lowers the peak battery temperature by 4.07 K, and eliminates temperature-difference exceedance. Continuous airflow modulation reduces the average of the seasonal maximum rates of temperature rise and drop by 38.6% and 75.1%, respectively, and decreases the standard deviation of temperature-difference change rate by 54.3%. Moreover, the proposed strategy achieves 67.1%-99.9% fan energy savings across all conditions. By preempting thermal responses and avoiding excessive cooling, the proposed strategy simultaneously enhances battery thermal safety and longevity without consuming excessive energy.
Silicon-carbon (Si/C) composite anodes are considered ideal replacements for traditional graphite anodes due to their high specific capacity. However, their significant volume changes during charge-discharge cycles readily lead to rapid capacity decay and mechanical failure of the electrodes, severely limiting their practical application. As a key component in electrodes, the binder is essential for preserving the mechanical integrity of the composite electrode and for constraining the expansion of active materials. This study ,an in-situ curvature measurement system was utilized to evaluate the impact of binders, i.e. polyacrylic acid (PAA), carboxymethyl cellulose (CMC), and polyvinylidene fluoride (PVDF), on the bending deformation, modulus, partial molar volume and stress-strain behavior of Si/C composite electrodes during lithiation/delithiation cycles. Then, scanning electron microscopy was used for systematic characterization the changes of electrode surface and cross-sectional morphologies. The results indicated that the Si/C composite electrodes with PAA could maintain relatively best electrochemical performance and structural integrity, owing to its large Young's modulus and strong confinement capability and which can effectively suppress its volume changes. While, the electrode with PVDF binder showed significant cracking evolution and interfacial delamination due to its lower modulus and poor bonding strength. This study reveals the working mechanism of the binder’s influence on the electrode stability from a mechanical-electrochemical coupling perspective, which can provide crucial theoretical and experimental basis for the rational design and optimization of superior Si/C anode binders.
Contemporary battery systems demand extended durability and high safety. Inspired by biological self-healing, self-healing strategies have shown promise in mitigating chemical degradation and physical damage. This review summarizes self-healing materials and mechanisms reported for battery applications, comparing their effectiveness. Importantly, we find that current research focuses on individual components, while device-level holistic self-healing remains largely unexplored. Distinct from prior reviews that concentrate on advancing self-healing technologies, this article maps self-healing needs to representative application scenarios and outlines a vision for intelligent battery systems. This review provides guidance for material selection and proposes a paradigm shift: transforming batteries from static energy containers into active energy systems.
An effective thermal management system is essential for ensuring the battery operating with optimal temperature. Poor heat transfer efficiency between the battery and the cold plate results in refrigerant-based thermal management systems (RTMS) being inadequate. Herein, an L-shaped ultra-thin vapor chamber (UTVC) strategy is proposed to enhance the performance of RTMS. The L-shaped UTVCs facilitate the construction of high thermal conductivity channels between the battery and the cold plate, enabling rapid and efficient heat transfer. Although this strategy results in a 1.9% reduction in battery energy density, the heat transfer efficiency of system is improved by more than 26.36% under test conditions. This improvement aids in better battery preheating and cooling under low ambient temperature conditions, subsequently affecting the thermal and electrical characteristics of battery. In this study, the presented strategy reduces the temperature response time by larger than 93.3% and decreases the preheating time required for the battery to achieve 2.28C charge by 50.4%. During 1C discharge, the average temperature rise and temperature difference are consistently kept below 5 °C and 2 °C, respectively. This work demonstrates the potential of the proposed approach to improve RTMS performance.
Conventional electrolytes severely limit lithium battery performance at low temperatures due to sluggish ion transport and high desolvation energy barriers. Weakly solvating electrolytes facilitate ion desolvation but suffer from limited salt dissociation, leading to the formation of extensive anion-rich aggregates (AGGs). These AGGs increase viscosity and reduce ion mobility, ultimately causing cell failure at low temperatures (≤ −20 °C). Here, we propose a moderate solvation design strategy by tuning weak lithium–solvent interactions to balance salt dissociation and ion mobility. By enhancing the Li+ transference number and reducing the desolvation energy barrier, this strategy facilitates the formation of a thin, robust, and uniform cathode electrolyte interface and a LiF-rich solid electrolyte interphase. Together, these features suppress parasitic reactions at the cathode surface and inhibit lithium dendrite growth, ensuring stable electrode-electrolyte interfaces even under subzero conditions. Furthermore, the designed electrolyte demonstrates flame retardancy and high-voltage tolerance. As a result, the Li||LiNi0.8Mn0.1Co0.1O2 full-cells demonstrate an impressive capacity retention of 95.7% at 4.5 V under −30 °C over 45 cycles (0.1 C). This design approach provides a promising pathway for the development of safe, high-voltage lithium(-Ion) batteries that can operate reliably in low-temperature conditions.
Hybrid proton exchange membrane fuel cell (PEMFC) powertrains offer a promising solution for long-endurance and zero-emission unmanned aerial vehicles (UAVs). However, balancing hydrogen economy, PEMFC durability, and energy-storage safety remains challenging under dynamic load demands. To address this issue, this paper proposes a physics-guided adaptive high-low-pass soft actor-critic (PG-AHL-SAC) energy management strategy (EMS). In the proposed framework, propulsion power demand is decomposed into frequency-specific components. The low-frequency component is primarily assigned to the PEMFC, while the battery (BAT) and supercapacitor (SC) buffer medium- and high-frequency transients, thereby reducing dynamic stress on the fuel cell. A continuous-action SAC controller then optimizes real-time power distribution using a multi-objective reward function accounting for hydrogen consumption, PEMFC degradation, and energy-storage constraints. The framework is evaluated under a representative single-mission flight profile, incorporating both noise-free and noisy mission profile. The results demonstrate that PG-AHL-SAC achieves a competitive hydrogen economy while strictly maintaining the BAT state of charge and SC energy limits. Compared with benchmark EMSs, the proposed strategy provides smoother PEMFC power trajectories and a more favorable trade-off among hydrogen economy, PEMFC durability, and energy-storage safety, performing closely to the offline dynamic programming optimum without requiring prior load knowledge. Furthermore, sensitivity analyses indicate robust performance under filtering parameter and cost-weight perturbations. Ultimately, these findings reveal that PG-AHL-SAC provides an effective, health-aware EMS for multi-source UAV powertrains operating under the specific mission profile evaluated. While robustness against noisy mission profile and parameter perturbations is confirmed within this tested scope, validation across broader operating conditions remains for future work.
With the rapid expansion of electric vehicles and electrochemical energy storage systems, retired batteries are entering a phase of large-scale recycling. Developing efficient and environmentally friendly recycling processes urgently requires a clear understanding of the electrochemical instability and thermal diffusion behaviour of different battery chemistries under low-temperature pretreatment conditions. In this study, the open-circuit voltage (OCV) evolution, mechanical failure behaviour under compression, and post-failure thermal diffusion characteristics of lithium iron phosphate (LFP), nickel cobalt manganese (NCM), and sodium-ion batteries (SIB) are systematically investigated at 0 °C, -10 °C, and -20 °C under varying states of charge (SOC). The results show that the displacement range over which OCV evolves from stability to fluctuation and eventual collapse depends strongly on SOC, revealing a coupling between stored energy and structural integrity. At 50% SOC, LFP and NCM batteries exhibit greater ductility and delayed load-bearing limits, whereas SIB batteries display pronounced brittle instability, characterized by severe post-peak load fluctuations associated with staged fragmentation, which are further intensified at high SOC. Thermal analysis indicates that decreasing temperature significantly lowers peak failure temperatures. Under identical conditions, NCM batteries exhibit higher peak temperatures and recurrent minor explosion-like sound, indicating elevated thermal risks during crushing operations. In contrast, SIB batteries show relatively stable thermal responses with low temperature sensitivity. These findings provide quantitative guidance for low-temperature pretreatment strategies, crushing process optimization, and safety window definition in end-of-life battery recycling.
Driving on slippery extreme condition poses an huge challenge for automated electric vehicles (AEVs) due to the significant degradation of tire–road friction and the strong coupling between longitudinal and lateral vehicle dynamics. The performance of traditional model predictive control (MPC) schemes relies heavily on accurate vehicle dynamics models and carefully tuned cost functions, both of which are difficult to achieve under slippery curved road conditions.To address these challenges, a hybrid vehicle motion model is first constructed by integrating a physics-based vehicle dynamics model with a 1-Lipschitz residual neural network, which improves prediction accuracy and adaptability under uncertain road conditions. Based on the hybrid model, an MPPI controller is developed in which actor–critic networks are employed to adaptively tune the stage cost weight factors, while a temporal-difference (TD) learning algorithm is used to learn the terminal cost from selected MPPI rollout trajectories. In this manner, short-horizon control performance and long-term cost optimization are jointly addressed within a unified MPPI framework.Compared with model-free reinforcement learning approaches, the proposed controller preserves high physical interpretability and enhanced safety awareness by embedding learning mechanisms into predictive control rollouts, rather than directly applying learned actions to the vehicle. Extensive simulation results demonstrate the effectiveness and superiority of the proposed RL-HMPPI controller under slippery extreme road conditions.
Electrochemical impedance spectroscopy (EIS) serves as a powerful analytical tool for characterizing electrochemical energy storage components, including lithium-ion batteries (LIBs) and lithium-ion capacitors (LICs). Although equivalent circuit models (ECMs) can comprehensively analyze EIS, the complexity of model structure selection and instability of parameter identification significantly limit their practical application. To address these limitations, this study proposes a flexible circuit modeling method based on a multi-objective optimization framework that automatically generates multiple optimal ECMs for given EIS measurements. The proposed method employs the non-dominated sorting genetic algorithm II (NSGA-II) for multi-objective optimization, utilizing non-dominated sorting and crowding distance calculations to simultaneously optimize ECM fitting accuracy and circuit complexity. Additionally, a two-stage hybrid approach combining the genetic algorithm (GA) and nonlinear least squares (NLS) is implemented to achieve fast and stable parameter identification. The proposed methodology is validated on two 25Ah LIBs under various state of charge (SOC) conditions and temperatures, as well as two 3200F LICs tested under different aging conditions and SOCs. Results demonstrate that more than eight viable ECMs are generated for each case, with 21.17% of the candidate models achieving normalized root mean square error (NRMSE) values below 1% while maintaining moderate circuit complexity (6-8 circuit elements). These findings confirm the effectiveness and practical viability of the proposed flexible modeling approach.
Solid-liquid phase change is a crucial technique for energy storage and thermal management in space vehicles. However, the underlying mechanisms of thermocapillary convection, natural convection, and their interactions during the solidification remain inadequately studied. To elucidate the convective heat transfer mechanisms of phase-change materials (PCMs) in thermal management systems for space exploration vehicles, this study employs n-octadecane as the PCM and conducts numerical simulations to systematically investigate the thermocapillary effect under both microgravity and normal-gravity conditions. The influence of cavity aspect ratio on the coupled behavior between thermocapillary and natural convection is also examined. The findings indicate that under one-sided cooling, the thermocapillary effect drives high-temperature fluid to form a clockwise vortex and transport it toward the phase interface, thereby inhibiting solidification, with the inhibitory effect intensifying as the temperature difference in the liquid phase increases. Under two-sided cooling, thermocapillary flow enhances heat transfer from the center to the cold boundaries, which promotes solidification. The vortices generated by natural convection align in direction with those induced by thermocapillary flow, leading to consistent solidification effects, namely inhibition in one-sided cooling and promotion in two-sided cooling. Furthermore, increasing the cavity height (i.e., reducing the aspect ratio) strengthens natural convection while weakening the thermocapillary effect, eventually allowing natural convection to dominate the solidification process. Under coupled conditions, the two convection mechanisms constrain each other, without exhibiting significant synergistic enhancement. This work clarifies the critical regulatory role of thermal boundary conditions and geometric configuration on the convective mechanisms during PCM solidification, providing theoretical guidance for the thermal design of energy storage and thermal management systems of space vehicles.
The fuel cell temperature fluctuates with load variations, causing external characteristics such as the polarization curve and maximum power point to drift. Most existing methods treat fuel cell energy management and thermal management as independent objectives, ignoring their inherent electro-thermal coupling, which significantly degrades control accuracy and increases costs. To address this, this paper proposes a hierarchical electro-thermal control method for fuel cell hybrid vehicles. In the thermal management layer, the cooling system is regulated to track the optimal temperature reference, ensuring the stack operates at theoretical maximum output voltage and power capability. In the energy management layer, a temperature-driven dynamic power boundary is utilized, and the output power is regulated adaptively based on the real-time thermal state. Simulation results under the CWTVC driving cycle show that the proposed strategy effectively maintains the stack at its optimal thermal state and avoids damaging low-load and high-load zones. Compared to a conventional energy management strategy (EMS), the total operational cost is reduced from 55.68 USD to 55.24 USD, with a 7.5% reduction in degradation costs. These results demonstrate that the collaborative strategy significantly improves the safety, durability, and economic performance of the fuel cell system.
Accurate state-of-health (SOH) estimation is essential for safe and reliable lithium-ion battery management, yet practical deployment requires not only point accuracy but also strong generalization, explainability, interpretability, and transparency under domain shift. To address this need, we propose a Physics-Prompt Cross-Attention Network (PPCAN) that incorporates electrochemical degradation knowledge into a data-driven predictor through a cross-attention mechanism. Unlike physics-informed neural networks (PINNs), which impose physical laws as soft penalties in the loss function, PPCAN represents key electrochemical variables—including Ohmic resistance (R0), SEI resistance (Rsei), lithium-ion diffusion coefficient (Ds), and cyclable lithium inventory (Cli)—as structured physical prompts that guide feature extraction and prediction. Because the current-cycle lithium-inventory proxy Cli(t) is capacity-derived and may act as a label proxy, we distinguish this setting from leakage-safer controls rather than treating it as a strictly label-independent input. A dedicated leakage-stress matrix shows that the current-cycle setting provides a diagnostic upper-bound result (OOD RMSE 0.0056), while a one-cycle-delayed cycle-level setting using Cli(t − 1) achieves a similar OOD RMSE of 0.0065 without using the target-cycle capacity. In this setting, Cli(t − 1) is a historical state available after the previous cycle has been completed, and the model estimates cycle-t SOH from the sensor sequence observed in cycle t; it is therefore a quasi-online SOH estimation protocol rather than a pre-cycle prediction protocol. The model also transfers across battery chemistries, retaining useful performance on mixed-chemistry batteries when initialized from single-material pretraining. Overall, the proposed framework provides a transparent and physically grounded route to battery SOH estimation under changing operating conditions.
Reliable fault diagnosis across heterogeneous battery systems is vital for the safety of large-scale electric vehicle fleets. This study addresses this challenge by developing a cross-domain diagnostic framework using a Multi-feature Convolutional Neural Network (MCNN) and long-term operational data from 1,000 electric vehicles. The methodology utilized Gramian Angular Field encoding to transform multi-dimensional time-series signals into two-dimensional feature-enhanced images, providing a robust representation of electrochemical anomalies. The core innovation lies in an intelligent parameter-filtering strategy designed to bridge the knowledge gap between nickel-cobalt-manganese and lithium iron phosphate systems. By integrating residual connections and attention modules into the MCNN architecture, the framework adaptively recalibrated salient fault features while mitigating domain-specific biases. Performance evaluation results demonstrate that the enhanced transfer network achieves high-precision identification, maintaining diagnostic accuracies between 96.8% and 98.46% across heterogeneous datasets. In addition to its high diagnostic precision, the proposed methodology demonstrates superior computational efficiency. The trained architecture is exceptionally lightweight, featuring a minimal memory footprint of 1.11 MB and an average inference time of 8.66 ms per data sample. These attributes render the framework highly suitable for real-time safety monitoring in resource-constrained onboard battery management systems. This approach is complemented by an adaptive isolation forest algorithm for micro-level pinpointing, successfully isolating specific resistance and capacity anomalies at the cell level. The integrated framework ensures reliable safety monitoring across diverse battery architectures, providing a robust solution for proactive maintenance in modern transportation environments.
Reinforcement learning has shown significant potential for autonomous racing, but it still faces challenges such as training instability, inefficient exploration, and unsafe action outputs in high-dynamic racing scenarios. This paper proposes a Trajectory guidance and Dynamics constraints Reinforcement Learning (TraD-RL) framework for autonomous racing. The proposed method incorporates expert prior knowledge into policy learning through Minimum Curvature Racing Line (MCRL) guidance, explicit vehicle dynamics constraints, and two-stage curriculum learning. MCRL provides global path and velocity references through observation augmentation and reward shaping, thereby improving exploration efficiency and racing performance. Yaw rate and sideslip angle constraints are introduced to characterize the vehicle dynamic safe operating envelope, and the corresponding stability costs are incorporated into policy optimization through Lagrangian relaxation. Moreover, the two-stage curriculum learning strategy enables a progressive transition from stable trajectory following to high-speed performance exploration. Experiments on two racetracks demonstrate that TraD-RL improves racing performance while maintaining a favorable balance between speed and dynamic stability. Further analyses of ablation, sensitivity, and robustness validate the effectiveness and stability of the proposed framework.
In real-world driving conditions, onboard state-of-health (SOH) estimation for lithium-ion batteries faces three major challenges: (1) most operational data lack accurate capacity labels, limiting their usefulness for supervised learning; (2) labels derived from ampere-hour integration are highly sensitive to operating conditions and measurement errors; and (3) vehicle-side observable features often exhibit weak mechanistic relevance and limited comparability across battery platforms with different nominal capacities. To address these issues, this study proposes a unified SOH estimation framework tailored for low-confidence fleet data. The method first calibrates capacity labels through SOC linearization, temperature/current-rate correction, and robust constraints, converting low-confidence operational segments into more reliable supervision. A multi-level feature system is then constructed by coupling dimensionless statistical descriptors with mechanism-based indicators, including incremental capacity (IC) peaks and relaxation signatures, thereby strengthening feature–SOH associations and improving robustness under capacity heterogeneity. Because IC and relaxation information cannot be extracted for every charging event, explicit binary masks and time-since-last-measurement variables are introduced so that models can down-weight stale or missing mechanistic features while still utilizing them when available.Based on the calibrated labels and unified feature representation, multiple representative models, including Informer, Autoformer, TCN, LSTM, N-BEATS, CNN, MLP, SVR, and XGBoost, are systematically benchmarked under a common evaluation protocol. The outputs of selected base models are further combined through a simple averaging ensemble, which exploits complementary error patterns to improve prediction stability without increasing model complexity. Evaluation using a three-year dataset from 300 vehicles with 155 Ah batteries and a limited 180 Ah cross-capacity validation setting provides preliminary evidence of applicability under capacity heterogeneity. On the 155 Ah fleet, the ensemble model achieves a fleet-level MAE of 1.11% and a MAPE of 1.31%, while the predicted end-of-life SOH of the held-out 180 Ah test vehicle remains close to the measured value.
The long voltage plateaus in LiFePO4 (LFP) batteries may cause the failure of existing data-driven state-of-health (SOH) estimation methods due to overly long health feature (HF) sampling intervals, rendering them almost indistinguishable from direct capacity measurement. However, most studies still rely solely on electrical signals, whereas force signal-based methods remain underdeveloped with limited practicality. To this end, we first propose a force-electrical coupled method (FECM) for feature engineering, which uses the second inflection points of charging force curves (SIPCF) as unified fixed references and deeply integrates force and electrical signals during feature extraction, rather than extracting HFs separately from each signal as model inputs. This increases the likelihood of extracting more effective HFs. These HFs reflect multidimensional aging mechanisms and comprehensively describe aging processes. We employ support vector regression (SVR) with hyperparameters optimized by particle swarm optimization (PSO) for fast LFP battery SOH estimation. Across the two in-house battery aging datasets with different initial conditions, the mean root-mean-square error (RMSE) for SOH estimation is only 0.45%. Compared with five typical methods, FECM improves estimation accuracy by up to 78% while reducing sampling time by up to 74%. Additionally, extracting specific HFs can further balance sampling time and estimation accuracy.
Autonomous driving control of heavy-duty vehicles remains highly challenging, particularly for cooperative transport systems (CTSs) carrying oversized payloads with multiple vehicle carriers. A primary difficulty arises from the frequent occurrence of stability-related accidents, such as rollover, during critical maneuvers. In this paper, a novel rollover prevention control pipeline for CTSs equipped with all-wheel-steering (AWS) distributed-electric-drive carriers is proposed. The framework is built upon an innovative iterative model predictive path integration (i-MPPI) method, which exhibits distinct advantages over common model predictive control (MPC) in dealing with the complex, nonlinear, high-dimension coupled dynamics of CTSs. To ensure effective rollover prevention performance, a control-barrier-function (CBF)-inspired safety cost is embedded into the sampling-based trajectory evaluation process, and a one-step safety modification is further applied to the i-MPPI control command. Extensive simulation results demonstrate that the proposed pipeline successfully prevents rollover while outperforming multiple baseline approaches in typical critical situations. The maximum LTRs decrease 13.33% and 18.57% compared with applying the conventional ReLU cost in two testing scenarios, respectively, and the computation speed reaches 7 – 10 times faster than common MPC. Moreover, the results reveal that the AWS distributed-drive architecture enables flexible pose regulation of both the payload and the carriers, allowing rollover mitigation to be achieved without significant degradation of trajectory tracking accuracy. The path tracking error of carriers decreases more than 50% and 80% in the two scenarios compared with using front-wheel-steering-and-driving FWSD carriers.
Autonomous Vehicles (AVs) currently face a significantly higher risk of being rear-ended by Human-Driven Vehicles (HDVs) compared to conventional vehicles—approximately 1.6 times higher. This vulnerability largely stems from the prevailing unilateral decision-making paradigm, which predicts surrounding traffic behaviors independently of the AV’s actions, thereby neglecting the reciprocal interplay between the AV and following HDVs. Consequently, AVs may execute maneuvers that are technically feasible but unexpected or hazardous for nearby human drivers, which can increase rear-end collision risk in high-interaction scenarios. To bridge this gap, this study proposes a World Model-based architecture designed to explicitly reason about these reactive behaviors. First, a World Model is established to learn environmental dynamics, enabling continuous inference of how HDVs react to specific AV actions. Second, a planner based on the Actor-Critic architecture is developed and trained entirely within this World Model, facilitating the learning of farsighted, closed-loop interaction policies. This architecture is implemented in a novel longitudinal control method, named ThinkACC, and tested in high-interaction platoon scenarios. Compared to baseline Adaptive Cruise Control (ACC) system, ThinkACC achieves an 86.8% reduction in rear-end collisions while maintaining high levels of comfort and efficiency. In particular, the agent emerges with human-like defensive driving behaviors, such as proactively accelerating to mitigate risks of rearward collision. These findings demonstrate the capability of World Models to reason about bidirectional interactions, offering a viable path for enhancing AVs safety and generalization in mixed traffic environments.
