Commentary: China’s Humanoid Hype Faces a Reality Check

The World Artificial Intelligence Conference concluded in Shanghai after a three-day event, marking its eighth year and demonstrating remarkable progress in robotics, with an unprecedented display of moving humanoid robots throughout the exhibition hall[para. 1]. Compared to the previous year, when most robots, such as Tesla’s Optimus, remained stationary, this year’s conference featured dynamic demonstrations. Visitors watched Unitree Robotics’ humanoid robots box each other, AgiBot’s robots play folk music, Galbot’s robot deliver soda, and PsiBot’s robot compete in mahjong—all illustrative of significant advancements among Chinese robotics firms, except for Tesla’s still immobile Optimus model[para. 2][para. 3].

Over 150 humanoid robots were exhibited, marking the largest showcase in China’s history, though an executive noted their operational capabilities varied significantly. Many robots were either controlled remotely by humans or pre-programmed to perform specific actions, while only a minority could make autonomous decisions independently[para. 4]. According to Yao Maoqing, president of AgiBot’s embodied intelligence division, 2024 will be a pivotal year, as technological demonstrations must now prove their worth in real-world scenarios to generate meaningful business results amidst increasing competition in China’s humanoid robotics industry[para. 5].

The recent surge in enthusiasm for humanoid robotics was driven by Tesla CEO Elon Musk, whose unveiling of the Optimus prototype in 2022 set off industry-wide momentum. Tesla has continually advanced its humanoid robot line, projecting to produce 5,000 Optimus units in 2025 and increase annual output to 50,000 in 2026. These projections, however, may face delays following the June departure of project leader Milan Kovac[para. 6].

In China, scaling up mass production remains a formidable challenge. AgiBot’s Yao explained that shipping even a few thousand units is challenging due to the immature state of upstream component suppliers, many of whom are small and midsize firms. However, established electromechanical companies are now entering the market, which could relieve some production bottlenecks over time[para. 7].

Beyond hardware, software limitations—particularly the lack of large datasets—present substantial obstacles. Unlike large language models, the humaniod-robotics field lacks vast freely-available data, leading to what Yao calls a “data desert”[para. 8]. AgiBot has responded by setting up a data-collection factory in Shanghai, where nearly 100 robots gather between 30,000 and 50,000 data entries each day to address this gap[para. 9].

Galbot, on the other hand, utilizes synthetic data to train its robots. Its founder Wang He stated that the embodied intelligence field is still in its infancy, and predicted that overcoming data scarcity may remain a challenge for the next 10-20 years. Most leading companies are unlikely to ship more than a few thousand robots annually in the near future, as trillions of data points might be required to achieve general intelligence in robots[para. 10].

This data shortage directly impedes the “intelligence” of models and limits the real-world utility of humanoid robots. As a result, experienced software companies are increasingly collaborating with robotics hardware firms. On Sunday, SenseTime introduced its Wuneng embodied intelligence platform for robotics capabilities, and Tencent launched its Tairos platform to support planning, perception, and interaction for humanoid robots, partnering with several prominent robotics companies[para. 11].

Despite notable progress, industry experts like Tencent’s chief scientist Zhang Zhengyou note that deploying humanoid robots is a highly complex, multidisciplinary effort. Most humanoid robots still serve primarily in research, data collection, or as guides, and Zhang likened the industry’s maturity to the era before mobile “brick phones,” suggesting widespread household use remains distant[para. 12].