AutoSplat is a framework for autonomous driving scene reconstruction using constrained Gaussian splatting. The method addresses challenges in reconstructing dynamic driving scenarios by imposing geometric constraints on Gaussians representing road and sky regions, enabling multi-view consistent simulation of complex scenarios like lane changes. It leverages 3D templates for foreground object initialization and introduces a reflected Gaussian consistency constraint to supervise both visible and unseen sides of foreground objects. Additionally, residual spherical harmonics are estimated for each Gaussian to model dynamic appearances. Extensive experiments on Pandaset and KITTI datasets show that AutoSplat outperforms state-of-the-art methods in scene reconstruction and novel view synthesis. The method also incorporates background geometry constraints to ensure consistent rasterization and handles dynamic foreground objects through temporal-dependent spherical harmonics. The framework is evaluated on various autonomous driving scenarios, demonstrating its effectiveness in generating realistic and accurate reconstructions. AutoSplat's key contributions include geometric constraints for background regions, use of 3D templates for foreground initialization, dynamic appearance modeling, and comprehensive evaluation against SOTA methods. The method is efficient and robust, making it suitable for real-world applications in autonomous driving.AutoSplat is a framework for autonomous driving scene reconstruction using constrained Gaussian splatting. The method addresses challenges in reconstructing dynamic driving scenarios by imposing geometric constraints on Gaussians representing road and sky regions, enabling multi-view consistent simulation of complex scenarios like lane changes. It leverages 3D templates for foreground object initialization and introduces a reflected Gaussian consistency constraint to supervise both visible and unseen sides of foreground objects. Additionally, residual spherical harmonics are estimated for each Gaussian to model dynamic appearances. Extensive experiments on Pandaset and KITTI datasets show that AutoSplat outperforms state-of-the-art methods in scene reconstruction and novel view synthesis. The method also incorporates background geometry constraints to ensure consistent rasterization and handles dynamic foreground objects through temporal-dependent spherical harmonics. The framework is evaluated on various autonomous driving scenarios, demonstrating its effectiveness in generating realistic and accurate reconstructions. AutoSplat's key contributions include geometric constraints for background regions, use of 3D templates for foreground initialization, dynamic appearance modeling, and comprehensive evaluation against SOTA methods. The method is efficient and robust, making it suitable for real-world applications in autonomous driving.