Recent advances in Large Vision–Language Models (LVLMs) have demonstrated strong cross-modal reasoning capabilities, offering new opportunities for decision-making in autonomous driving. However, existing end-to-end approaches still suffer from limited semantic consistency, weak task controllability, and insufficient interpretability. To address these challenges, we propose SemAlign-E2E (Semantic-Aligned End-to-End), a semantic-aligned multimodal LVLM framework that unifies visual, LiDAR, and task-oriented textual inputs through cross-modal attention. This design enables end-to-end reasoning from scene understanding to high-level driving command generation. Beyond producing structured control instructions, the framework also provides natural-language explanations to enhance interpretability. We conduct extensive evaluations on the nuScenes dataset and CARLA simulation platform. Experimental results show that SemAlign-E2E achieves substantial improvements in driving stability, safety, multi-task generalization, and semantic comprehension, consistently outperforming state-of-the-art baselines. Notably, the framework exhibits superior behavioral consistency and risk-aware decision-making in complex traffic scenarios. These findings highlight the potential of LVLM-driven semantic reasoning for autonomous driving and provide a scalable pathway toward future semantic-enhanced end-to-end driving systems.

To address the challenge of directly translating predicted distributions of dynamic obstacles into three-dimensional (3D) planning constraints during unmanned aerial vehicle (UAV) flight, we propose Torch GIP-RB-SI-3DA*, a graph-interaction-prediction-based method for planning safe flight intervals and dynamic 3D paths. The model predicts trajectory distributions over a 15 s horizon from the 3D positions and velocities observed in six consecutive frames and maps the calibrated distributions to instantaneous voxel-level risks. The front end enforces both instantaneous- and cumulative-risk constraints through multi-label search. The back end performs finite-difference smoothing and joint retiming, followed by dense-sampling verification of collision freedom and the risk, velocity, acceleration, and jerk constraints. Across 10 independent training seeds, Torch GIP reduces the average displacement error (ADE) and final displacement error (FDE) by 5.07–15.25% and 5.93–9.29%, respectively, relative to NoGraph, while also achieving lower negative log-likelihood (NLL). All these improvements remain statistically significant after Holm correction. In 100 pre-screened in-distribution (ID) scenarios, Torch GIP generates 88 candidate trajectories, of which 82 are collision-free under ground truth, 76 satisfy the ground-truth instantaneous-risk constraint, and 82 satisfy the ground-truth risk budget in end-to-end evaluation. Although its candidate yield is lower than that of the constant-velocity (CV) baseline, its yields under all three ground-truth evaluation criteria are higher than those of both CV and rule-based GIP. Paired ablation studies identify joint retiming as a critical component for maintaining end-to-end yield. Across 50 out-of-distribution (OOD) scenarios, the method generates 48 planned trajectories, of which 41 are collision-free under ground truth, and 33 satisfy the ground-truth risk budget. All 16 replays of UZH-FPV trajectories with known future motion pass both offline and sampling-based verification. These results demonstrate that the proposed method provides a systematic and verifiable transformation from predicted distributions to 3D planning constraints.