The defects in the mirror neuron system (MNS) make it challenging for children with autism spectrum disorder (ASD) to learn social skills through third-person digital agents. To help children with ASD circumvent deficits in MNS, we propose an avatar-based intervention system that enables them to learn social skills through first-person avatars (i.e., their digital twins) in this paper. We first reconstruct a 3D face from a single image of a child who participates in the intervention, then stitch the reconstructed 3D face onto the preset 3D character model to create a 3D avatar for the child, and finally embed the 3D avatar into social games. Experimental results indicate that our methods can reconstruct realistic 3D child faces and seamlessly stitch them onto 3D character models. We have used our system to intervene with 11 children with ASD, and these children can recognize the avatars in the social games as “themselves” in the vast majority of scenarios.