AI maps 200,000 3D face datasets to sharpen humanoid robots

A research team from China, led by Prof. SONG Zhan and Dr. YE Yuping, has developed a large-scale 3D facial dataset and an innovative AI model to improve the realism of humanoid robots. Unlike traditional methods that rely on 2D texture mapping or synthetic 3D faces, their approach processes raw 3D facial scans directly, addressing inaccuracies caused by texture misalignment and digital model discrepancies. The team created a comprehensive database of around 200,000 high-fidelity 3D facial scans, including multi-expression and dynamic 4D datasets, making it one of the largest structured collections of real 3D human facial data to date. This dataset was recognized by Fujian Province’s 2025 High-Quality AI Dataset Program. The researchers introduced a curvature-fused graph attention network (CF-GAT) that analyzes unordered point clouds representing facial geometry without surface textures. By incorporating a geometry-driven sampling strategy and encoding curvature as a geometric prior within the model’s attention