比亚迪股份有限公司

With the rapid development of vehicle-road-cloud integrated systems and artificial intelligence (AI) technologies, intelligent connected vehicles (ICVs) can be empowered with enhanced capabilities to address the inherent limitations of single-vehicle driving, such as restricted perception range and insufficient computational resources, making them an ideal platform for realizing future autonomous driving. However, there exists a notable research gap in applying AI to energy-efficient driving strategies for hybrid electric vehicles (HEVs), while simultaneously ensuring improved traffic efficiency. Therefore, this study presents a novel cloud-based sequential optimization deep deterministic policy gradient (SO-DDPG) framework addressing the energy-efficient and time-efficient control strategy for hybrid electric vehicles in urban environments. Firstly, a hierarchical “cloud decision-making, vehicle control” architecture is implemented by featuring an innovative integration of real-time energy management to achieve coordinated optimization of speed planning and power distribution. Secondly, a double-layer DDPG architecture is employed for control strategy learning. The first layer trains the energy management model in the cloud, taking vehicle motion states and powertrain system states as input variables to generate engine power control signals. The second layer integrates various information including traffic signal phases and preceding vehicle states to plan optimal velocity sequences. Finally, comprehensive validation experiments conducted on urban roads demonstrate the algorithm’s robust performance: Comparative analyses show that SO-DDPG achieves 6.05

The battery thermal management system (BTMS) is critical for ensuring the safety and extending the cycle life of lithium-ion batteries. An effective BTMS maintains batteries within an optimal temperature range while improving overall system performance. To address the temperature non-uniformity of single bottom-mounted liquid cold plates, this study proposes a hybrid heat pipe-liquid cold plate BTMS to enhance local heat dissipation and overall thermal performance. A compact electric vehicle is selected as the research object, and the battery pack configuration is determined based on driving power and range requirements. An 80S2P configuration satisfies both the energy and power requirements. At the module level, the hybrid BTMS is compared with a conventional liquid cold plate system, demonstrating superior temperature uniformity and lower maximum temperature. A single-factor analysis is conducted to investigate the effects of channel width, cold plate thickness, and channel depth on maximum temperature, temperature difference, and pressure drop. Both discretevariable optimization and continuous-variable optimization strategies are employed to identify the optimal cold plate structure. After optimization, the hybrid BTMS decreased the maximum temperature, maximum temperature difference, and flow-channel pressure drop of the battery module by 0.94 degrees C, 1.9 degrees C, and 22 Pa, respectively. Validation under multiple discharge rates, ambient temperatures, and real driving cycles confirms that the optimized BTMS maintains robust thermal performance across diverse operating conditions. These results provide a practical engineering reference and a systematic methodology for the thermal management design of high-energy-density battery systems.

Vision-Language-Action (VLA) models augmented with world modeling represent a promising paradigm for end-to-end autonomous driving. While pixel-level future prediction enables fine-grained spatiotemporal reasoning, it compromises robustness in noisy driving scenarios. Conversely, latent-based world models alleviate this sensitivity but often incur limited interpretability and representational degradation due to absent pixel-level grounding. To reconcile this trade-off, we propose HyWorldVLA, a hybrid world-VLA framework that unifies pixel-level supervision and latent representation learning. In the pre-training stage, HyWorldVLA predicts video latents encoded by a pre-trained video VAE, while simultaneously reconstructing video frames to provide precise pixel-level grounding. During the subsequent co-fine-tuning phase, the model exclusively predicts latent features, which are fed into an action expert to generate trajectories. Extensive experiments on NAVSIM v1 and v2 benchmarks demonstrate that HyWorldVLA significantly outperforms both pixel-based and latent-based world model baselines. Notably, we present the first comprehensive qualitative and quantitative analysis of world model noise robustness in autonomous driving, establishing a new benchmark for evaluating future architectures.

Generating safety-critical scenarios that can expose system limits is crucial for the safety verification of autonomous driving, yet such scenarios are scarce and costly to obtain in the real world. Data-driven and knowledge-based methods have been widely applied in scenario generation for simulation testing, while the proposal of LLM provides a new paradigm. With the powerful semantic understanding and generation capabilities, LLM mitigate the blindness of data-driven approaches and eliminate the extensive manual effort required by purely rulebased methods. We proposes a safety-critical scenario generation method based on LLM guided reinforcement learning. Using LLM to transform natural language described scenarios and reward function codes guiding intelligent agents, thereby generating diverse scenarios that are consistent with the description. In the training process of RL, intrinsic hindsight rewards are given to key state chains with adversarial behavior, enhancing the key states that can transition to potential adversarial behavior, achieving task completion while improving scenario adversarial performance. The experimental results show that the scenario generated by the proposed method is highly consistent with the code generated by LLM and language description intention, and shows adversary on the tested vehicle, creating conditions for active detection of the safety boundary.

Federated learning (FL) is increasingly vulnerable to backdoor attacks, where malicious participants inject poisoned updates to implant hidden behaviors within the global model. Existing defenses often rely on ad-hoc aggregation rules, heuristic parameter modifications, or external trusted datasets, severely limiting their robustness, interpretability, and transferability across diverse architectures and attack modalities. To address these fundamental limitations, we propose a Guided Unlearning for Adversarial Robustness and Defense (i.e., GUARD). It is a unified, data-centric framework, which leverages the principles of machine unlearning to systematically secure FL systems. Specifically, GUARD establishes a rigorous theoretical link between data influence estimation and backdoor mitigation. It utilizes fast influence-function approximation to efficiently isolate poisoned training contributions without requiring auxiliary clean data. Building on this precise detection, the framework executes a sequential, closed-loop defense pipeline: an influence-guided unlearning stage that explicitly neutralizes malicious parameter biases through reverse optimization, followed by a stability-preserving lightweight fine-tuning stage on internally verified clean data to seamlessly restore standard predictive performance. Comprehensive experiments across diverse datasets, advanced backdoor attacks, and model architectures demonstrate that GUARD achieves superior threat eradication and clean accuracy preservation compared to state-of-the-art baselines. By transforming backdoor removal into a causally interpretable unlearning process, this work offers a highly robust and principled pathway toward trustworthy FL.