Closed-loop Performance Analysis of end-to-end Autonomous Driving Models by Rendering Fidelity: An AlpaSim 3DGS Evaluation

This study evaluates the closed-loop driving performance of three heterogeneous end-to-end autonomous driving models, representing the bird’s-eye-view (BEV)-Transformer, convolutional neural network (CNN), and vision-language-action (VLA) architectures, through deterministic experiments on a high-fidelity 3D Gaussian Splatting (3DGS) simulator in conjunction with a synthesis of results from multiple public benchmarks. Across 49 evaluation scenes, a compact 132M-parameter BEV-Transformer model achieves the highest mean valid driving distance of 141.0 m, exceeding by 7.0% the 131.8 m achieved by a VLA model approximately 76 times larger. The same UniAD architecture shows widely varying closed-loop performance across platforms: it yields a success rate (SR) of 16.36% on Bench2Drive with CARLA-retrained weights, an SR of 47.0% on DriveE2E with nuScenes weights, and the top rank on AlpaSim 3DGS with 141.0 m using nuScenes weights. These results suggest that rendering fidelity and training-evaluation domain alignment jointly influence deployment evaluation. Intersection direction errors emerge as a structural limitation common to all three models. These findings offer practical implications for simulation platform selection and model deployment decisions in military autonomous vehicles.