3D视觉派：斯坦福具身智能人物关系图

This article traces the intellectual lineage of embodied AI research through key Stanford figures and their networks. Lu Cewu (卢策吾) pursued similar questions from different angles. After studying vision at Hong Kong Chinese University under Jia Jiaya, he questioned how to integrate AI's scattered subdisciplines into unified embodied agents — essentially, robots. Working under both Fei-Fei Li and Leonidas Guibas between 2015-2016, Lu conducted crucial early experiments. He and Zhu Yuke (朱玉可) explored object relationships and visual understanding, laying groundwork for embodied AI research. Lu later founded Feixi Robotics in 2016 with Wang Shiquan (王世全). The article documents a striking asymmetry: while 2018-2021 saw explosive embodied AI research in North America, Chinese academia remained largely unaware. Lu presented embodied AI at VALSE 2020 with audiences dwindling to single digits. The ecosystem now includes Su Hao's SAPIEN simulator (2020) and HillBot robotics company (2024); Lu Cewu's Qionche Intelligence (2023), splitting from Feixi to focus on embodied AI "brains"; Wang He's Yinhe Universal Robots (2023), claiming 95% success rates using pure simulation data. The article emphasizes that truly comprehensive embodied AI requires four technical stacks: vision, graphics, learning, and control — a rare combination held by this Stanford-connected cohort. Su Hao developed ShapeNet, a large-scale 3D dataset containing over 3 million models across 3,135 categories. By 2017, Su Hao and Qi Ruizhen developed PointNet using ShapeNet, establishing deep learning's viability for 3D point clouds. The work attracted autonomous driving applications and generated over 10,000 citations.<br/>---END---