Commentary: China’s Humanoid Hype Faces a Reality Check

The three-day World Artificial Intelligence Conference recently concluded in Shanghai, marking its eighth edition with the most robots ever displayed at the venue. The event highlighted rapid progress in humanoid robotics, which was evident not only in the number and diversity of robots showcased but also in the increasing demands for real-world demonstrations of their capabilities [para. 1]. Compared to last year—when robots were mostly static displays à la the Tesla Optimus prototype—this year's expo featured interactive performances: Unitree's humanoids boxed, AgiBot's robots played folk music, Galbot's robots fetched drinks, and PsiBot's offered to play mahjong. Notably, while Tesla’s Optimus remained inactive for safety or proprietary reasons, most Chinese robotics firms had to prove their progress on-site, meeting a growing expectation for practical, dynamic demonstrations [para. 2][para. 3].

Organizers reported over 150 humanoid robots on display, setting a record for industry debuts in China. However, most of these robots still only partially operate autonomously. According to an executive, there are three categories of show robots: those remotely controlled by humans, models pre-programmed for specific actions, and only a handful with genuine autonomous decision-making. The pressure is mounting as 2024 progresses; companies must now move from showpieces to real-world results, distinguishing themselves by technological superiority and revenue potential [para. 4][para. 5].

The boom in interest around humanoid robotics is partly due to Tesla CEO Elon Musk, whose company introduced its first Optimus prototype in September 2022. While that early prototype couldn’t walk, by late 2023, a second generation had emerged, and a third is expected in 2024. Musk predicts that with advances in AI, learning from human demonstration will accelerate robot capability growth, and he has set ambitious targets: 5,000 Optimus robots in 2025 and up to 50,000 annually by 2026. However, with the recent departure of the project’s leader, Milan Kovac, in June, meeting these milestones may now face delays [para. 6].

Chinese companies confront significant mass production barriers as well. AgiBot's leadership concedes that producing thousands of robots is still daunting, mainly due to supply chain immaturity—many essential parts suppliers are small, scaling up slowly. Yet, as larger electromechanical firms join the sector, these capacity bottlenecks may ease over time [para. 7].

Beyond hardware, the most acute challenge is software, particularly access to sufficient data for model training. Unlike large language models, which benefit from the vast troves of internet data, embodied intelligence struggles with limited, relevant datasets [para. 8]. To help close this gap, AgiBot has opened a Shanghai data-collection facility, running nearly 100 robots and generating up to 50,000 data entries daily [para. 9]. Another company, Galbot, relies heavily on synthetic data, but its founder, Wang He, believes genuine general intelligence in robots may require trillions of data points—a milestone likely decades away. Even leading companies might only produce a few thousand robots this year due to these constraints [para. 10].

The lack of data directly curtails the intelligence and usefulness of humanoid robots. This challenge has encouraged partnerships with software companies with expertise in AI, autonomous driving, and robotics. At the conference, SenseTime launched the Wuneng intelligence platform, while Tencent introduced its Tairos platform, partnering with major robotic manufacturers to provide advanced AI planning and perception capabilities [para. 11].

Despite these advances, Tencent's chief scientist Zhang Zhengyou notes that humanoid robots are still in a research phase, performing limited tasks such as data collection and guiding tourists. He likened the current stage to the pre-brick-phone era of mobile phones, suggesting mass adoption of humanoid robots in daily life remains a distant prospect [para. 12].