In Depth: Humanoid Robots Need a Job

The 2024 World Robot Conference (WRC) in Beijing recently concluded, featuring participation from over 200 robotics companies and showcasing more than 1,500 products, including humanoid robots capable of running, dancing, and boxing. Despite these impressive demonstrations, significant technological limitations remain. These humanoid robots can mimic human actions but still struggle to interact dynamically with their environments or perform complex tasks, restricting their practical use mainly to simpler roles like customer service, while broader industrial adoption is just beginning. The advancement of more useful robots will require major breakthroughs in technology [para. 1][para. 2].

According to industry insiders, commercial viability remains an issue, with profitability for even the supposedly market-ready wheeled humanoid robots being distant. Some attempts to make robots more affordable, such as Unitree Robotics pricing models as low as 40,000 yuan ($5,566), mean compromises in capability. At WRC, robot prices ranged from tens to hundreds of thousands of yuan, with more advanced robots (capable of folding clothes or making coffee) costing up to 500,000 yuan. For household ubiquity, costs across the entire supply chain must drop considerably [para. 3][para. 4][para. 5].

Investor enthusiasm, however, persists. Kang Yu from Shoucheng Holdings noted the speed with which embodied AI financing is proceeding, with several firms eyeing IPOs. Embodied AI, which gives intelligent machines the ability to interact physically with the real world, is a key focus [para. 6][para. 7].

A significant debate exists over the main bottleneck for embodied AI: model architecture versus the quality and quantity of training data. Unitree Robotics CEO Wang Xingxing argues that, while hardware is sufficiently advanced, AI models lag behind large language models and data accumulation strategies that worked in language AI are less effective here. Others, like X Square Robot and Robot Era, believe VLA (vision-language-action) models or a mixture of models and high-quality real-world data are critical to progress, with each emphasizing different aspects as the core challenge [para. 8][para. 9][para. 10][para. 11][para. 12][para. 13][para. 14].

Further, production and hardware costs present formidable obstacles. High-end consumer pricing persists due to the need for powerful computing systems and sophisticated components like dexterous hands. For example, Digit Huaxia’s humanoid robot Xing Xing Xia starts at 198,000 yuan, justified by the need for complex interaction. More advanced dexterous hands, which can handle intricate tasks, can cost from 10,000 yuan for basic models to 50,000 yuan or more for models suitable for commercial applications like coffee-making. As tasks grow in complexity, hardware costs climb proportionately [para. 17][para. 18][para. 19][para. 20][para. 21][para. 22].

Despite scalability challenges, funding for embodied AI is surging. Investor interest was boosted by a high-profile dancing robot performance during the Chinese Spring Festival Gala and substantial fundraising rounds among key Chinese robotics startups. Most investment is directed toward data collection—costs for comprehensive data gathering centers can easily reach into hundreds of millions of yuan. Local governments, such as Tianjin, are supporting the industry, as seen with new AI data facilities [para. 24][para. 25][para. 26][para. 27][para. 28][para. 29].

Some voices caution that a financial bubble is forming, suggesting that only firms generating substantial annual revenue (at least 400 million yuan) will survive the coming consolidation. This signals an impending shakeout where only companies with sustainable long-term potential may persist [para. 30][para. 31].