End-to-End Autonomous Guidance Method Integrated With Mixture-of-Experts for Intelligent Vehicles

The end-to-end autonomous guidance method requires massive high-quality training data to achieve better performance. In reality, new and uncollectable environments– - characterized by spatial variations in geographical scenes (urban/rural/mountainous terrains with diverse road rules like left/right driving), road network topologies (complex intersections, varying lane curvatures), and terrain features (steep slopes, occlusions from tunnels or buildings)—inevitably cause generalization decline. Consequently, this paper proposes an Autonomous Guidance method based on a novel Mixture-of-Experts framework (AG-MoE). AG-MoE maps environment information to guidance by fusing knowledge from training data and pre-trained models, explicitly addressing spatial challenges: its feature extraction structure processes LiDAR point clouds with geometric priors (e.g., lane curvature, slope gradients), while the MoE structure dynamically selects expert branches for diverse scenarios (e.g., urban intersections vs. mountainous curves). This enables effective fusion of multi-scene knowledge—such as adapting to cross-regional infrastructure (roundabouts vs. auxiliary roads) or terrain-induced sensor distortions—thereby improving generalization in unknown environments. Experiments on CARLA (with varied town maps) and real-world datasets validate that AG-MoE outperforms baselines, demonstrating enhanced adaptability to unseen spatial configurations.