Recent 3D avatar reconstructed by 3D Gaussian Splatting (3DGS) achieves high-quality results on known poses but produce artifacts for unseen poses. To address this issue, we propose DualAvatar, a robust avatar reconstruction method for reconstructing and rendering unseen poses more accurately from monocular video. Our method achieves this by simultaneously optimizing a 3DGS avatar and a mesh avatar. During training, the mesh avatar constrains the unseen regions of the 3DGS avatar, while the 3DGS avatar provides geometry information of the mesh avatar. Experimental results on the PeopleSnapshot dataset demonstrate the superior performance of DualAvatar.