Articles

The article details the Morpheus lifelike robot face system released by Zhao Hao's team at Tsinghua University Institute for AI Industry Research, in collaboration with several institutions. The system aims to overcome the limitations of traditional mechanical faces in emotional expression. It integrates 'hybrid drive structure,' 'self-learning inverse modeling,' and 'speech-driven emotional animation' to generate 25 delicate expressions in real-time. The article explains how these technologies enable robots to understand and learn emotions, and provides experimental data to verify its excellent performance in emotion recognition and lip-sync. Morpheus's software and hardware will be fully open-source, providing a technological foundation for the next generation of emotional robots, with significant industrial feasibility and reusability.

The article provides a detailed review of the development history of High-Performance Computing (HPC) in China, particularly the key contributions of the Institute of High-Performance Computing (Institute of HPC) at Tsinghua University. The article emphasizes that HPC is a critical enabler for the advancement of large language models, encompassing machine setup, system software enhancement, large-scale parameter storage, and distributed computing. Tsinghua's Institute of HPC pioneered the shift to purchasing workstations to build cluster systems in the 1990s, and later focused on software optimization and applications for large-scale clusters, accumulating rich experience in computation, storage, and communication. The article showcases the breakthroughs and innovations of the Institute of HPC in supercomputing, graph computing, storage, and current LLM training and inference optimization (such as Bagua (a training system), Chitu (an inference engine), Mooncake, KTransformers, SuperFS, MadFS, and other systems) through the contributions of core figures and their teams, including Zheng Weimin, Chen Wenguang, Zhai Jidong, Wu Yongwei, Zhang Mingxing, and Lu Youyou. The article emphasizes that 'Hardware-Software Co-design' is the core research ideology of the Institute of HPC, and points out that it plays a decisive role in reducing the training and inference costs of large language models, leading to high demand for graduates from Tsinghua's Institute of HPC among top domestic technology firms.

The article discusses in detail the severe challenges faced by the dexterous hand industry in the field of Embodied Intelligence, pointing out that the technology is far from mature, but market price wars have escalated prematurely. Dexterous hand startups face pressure from both upstream component manufacturers and downstream complete machine manufacturers, limiting their survival space. The article deeply analyzes the core pain points such as 'High Biofidelity' (disputes over degrees of freedom and transmission schemes), 'Advanced AI Integration' (insufficient intelligence, lack of data), and 'Earliest Mass Production' (scattered commercial customers, unfocused scenarios). Ultimately, all controversies point to the only consensus of 'cost reduction', and the industry is caught in a vicious cycle of 'wanting to reduce costs but lacking scale'. The article believes that although domestic substitution and design optimization have promoted cost reduction, technological upgrades lag behind price wars, leading to low-priced, impractical products. At the same time, the article also mentions attempts to seek breakthroughs through the deployment of arm-hand systems on production lines and data open source.