Xuan Cai

Xuan Cai

Auto-generated profile, based on publicly available information

Xuan Cai

School of Transportation Science and Engineering, Beihang University, Beijing, China

Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
6 November 2024 in IEEE Transactions on Intelligent Vehicles
Verified
Journal Article
Closed access
KJKemou Jiang
Kemou Jiang
XCXuan Cai
Xuan Cai
Zhiyong CuiZhiyong Cui
Zhiyong Cui
ALAoyong Li
Aoyong Li
YRYilong Ren
Yilong Ren
PCPinlong Cai
Pinlong Cai
Large language models (LLMs) as autonomous agents offer a novel avenue for tackling real-world challenges through a knowledge-driven manner. These LLM-enhanced methodologies excel in generalization and interpretability. However, the complexity of driving tasks often necessitates the collaboration of multiple, heterogeneous agents, underscoring the need for such LLM-driven agents to engage in cooperative knowledge sharing and cognitive synergy. Despite the promise of LLMs, current applications predominantly center around single-agent scenarios, which limits their scope in the face of intricate, interconnected tasks. To broaden the horizons of knowledge-driven strategies and bolster the generalization capabilities of autonomous agents, we propose the $\mathsf {KoMA}$ framework consisting of the multi-agent interaction, the multi-step planning, the shared-memory, and the ranking-based reflection modules to enhance multi-agents' decision-making in complex driving scenarios. Based on the framework's generated text descriptions of driving scenarios, the multi-agent interaction module enables LLM agents to analyze and infer the intentions of surrounding vehicles based on scene information, akin to human cognition. The multi-step planning module enables LLM agents to analyze and obtain final action decisions layer by layer to ensure consistent goals for short-term action decisions. The shared memory module can accumulate collective experience to make superior decisions, and the ranking-based reflection module can evaluate and improve agent behavior with the aim of enhancing driving safety and efficiency. The $\mathsf {KoMA}$ framework not only enhances the robustness and adaptability of autonomous driving agents but also significantly elevates their generalization capabilities across diverse scenarios. Empirical results demonstrate the superiority of our approach over traditional methods, particularly in its ability to handle complex, unpredictable driving environments without extensive retraining. Project Page: https://jkmhhh.github.io/KoMA/ .

25 June 2024 in IEEE Transactions on Intelligent Vehicles
Verified
Journal Article
Closed access
XCXuan Cai
Xuan Cai
XBXuesong Bai
Xuesong Bai
Zhiyong CuiZhiyong Cui
Zhiyong Cui
Peng HangPeng Hang
Peng Hang
HYHaiyang Yu
Haiyang Yu
YRYilong Ren
Yilong Ren
Testing is a pivotal phase for uncovering potential vulnerabilities in autonomous vehicles (AVs) to develop a secure autonomy system. However, existing methods often lack consideration for efficiently exploring multiple vulnerability-revealing cases, particularly under adversarial game scenarios. We introduce an evolving series reinforcement learning (RL) framework for adversarial policy training, integrating Responsibility Sensitive Safety (RSS) and Dynamic Time Warping (DTW) theories to shape the reward function to steer the evolving direction of the subsequent series agents for exploring vulnerability-revealing attack scenarios uncharted in the refined buffered repository. Our method undertakes adversarial stress tests for both black-box and white-box AV systems under test in driving tasks that engage in games with traffic vehicles and pedestrians. The results indicate that our approach expedites the exploration of additional scenarios blamed for the AV, outperforming the baselines in the vulnerability-revealing accident and scenario diversity. Furthermore, the causality of the collisions is qualitatively analyzed to provide insights for AV system vulnerability repair. Code is available at https://github.com/caixxuan/AST-SRL .

11 February 2025 in IEEE Transactions on Vehicular Technology
Verified
Journal Article
Closed access
Yuanhao HuangYuanhao Huang
Yuanhao Huang
HWHao Wang
Hao Wang
XBXuesong Bai
Xuesong Bai
XCXuan Cai
Xuan Cai
HYHaiyang Yu
Haiyang Yu
YRYilong Ren
Yilong Ren
Existing multi-agent algorithms struggle with spatial coordination and require extensive prior environmental information for effective large-scale operation. This study introduces an innovative multi-UAV system that employs an evolutionary algorithm, reconceptualizing collaborative search tasks as a multi-objective cooperative challenge. Inspired by bacteria's stochastic exploration for pheromones, the algorithm utilizes pheromone concentration and gradient effects to enhance multi-source exploration's efficiency and robustness. A bespoke multi-UAV framework with UAV-to-UAV (U2U) communication was developed to implement and test the algorithm. Comprehensive experiments tested various chemotaxis strategies, guidance strengths, team sizes, and starting conditions, demonstrating the algorithm's consistent effectiveness in expansive terrains without relying on prior information.

29 November 2024 in Energy
Verified
Journal Article
Closed access
XCXuan Cai
Xuan Cai
WZWei Zhou
Wei Zhou
ZCZhiyong Cui
Zhiyong Cui
XBXuesong Bai
Xuesong Bai
FLFan Liu
Fan Liu
YRYilong Ren
Yilong Ren

10 September 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
XBXuesong Bai
Xuesong Bai
HYHaiyang Yu
Haiyang Yu
CXCai Xuan
Cai Xuan
JXJiandong Xing
Jiandong Xing
YRYilong Ren
Yilong Ren
Reconstructed driving scenes provide realistic roads, backgrounds, and sensor context for autonomous-driving simulation, but vehicles recovered with the scene are usually reliable only along recorded camera trajectories. Rotating or relocating them exposes missing surfaces and unstable appearance. A complete vehicle reconstructed from a single image offers broader visibility, yet its scale, road contact, LiDAR geometry, and illumination need not match the target scene. We present LAGR, an offline framework that converts a complete 3D Gaussian Splatting (3DGS) initialization into a metrically grounded, scene-adapted vehicle. LAGR uses road-retaining LiDAR to establish contact and object-only LiDAR to correct scale and shape through confidence-weighted bounded deformation. It then constructs scene-conditioned targets across multiple backgrounds and orientations, combines their low-frequency color and illumination with vehicle detail, and distills the result into spherical-harmonic DC coefficients while fixing geometry, opacity, and view-dependent appearance. Across nine vehicles, LAGR achieves a 43.16 F-score at 5 cm against evaluation-only object LiDAR, compared with 31.82 for the strongest direct image-to-3D baseline. Across 17 placements and 204 evaluation views, the bake increases peak signal-to-noise ratio from 18.65 to 22.98 dB and reduces Learned Perceptual Image Patch Similarity from 0.1136 to 0.0717. The resulting vehicle can be moved and rotated without rerunning the appearance network. Including contact shadow, its median incremental appearance cost is 8.09 ms, versus 28.54 ms for per-frame color transformation and 103.51 ms for a one-step diffusion architecture. LAGR therefore produces one reusable vehicle 3DGS for repeated simulation.

2 September 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
XWXiao Wen
Xiao Wen
NZNanfang Zheng
Nanfang Zheng
CXCai Xuan
Cai Xuan
FZFan Zhang
Fan Zhang
SHSonghua Hu
Songhua Hu
ZPZiyuan Pu
Ziyuan Pu
Autonomous vehicles (AVs) are expected to revolutionize urban mobility by enhancing safety, efficiency, and sustainability, yet their actual deployments remain opaque. Here, we systematically and quantitatively analyze the spatial footprint of AV services across 1,408 cities worldwide to delineate the deployment disparities. We identify a dual-layer concentration pattern associated with socioeconomic, climatic, infrastructural, and functional characteristics. Globally, operations cluster in affluent and temperate regions with mature infrastructure, where AV-served cities experience 54–90% less annual snowfall than comparable AV-unserved counterparts. Locally, services concentrate in economically vibrant cores with high road density but simplified intersection geometries, and are largely absent from complex road networks and transit-poor areas. Using a Deployment Compatibility Gap (DCG) index, we demonstrate that the Global South exhibits markedly lower deployment compatibility than the Global North, rendering it an AV service desert. These findings suggest that current operational limitations may contribute to long-term mobility inequality, highlighting the need for inclusive strategies to prevent the marginalization of global cities.