In Depth: Humanoid Robots Need a Job

The 2024 World Robot Conference (WRC) in Beijing drew over 200 robotics companies showcasing more than 1,500 products, including advanced humanoid robots that can run, dance, and box. Despite these impressive demonstrations, most robots displayed face significant technical challenges and are limited to simple customer service roles, with industrial applications still in their infancy due to a lack of real-time interaction and complex task abilities. Major technological breakthroughs are needed to expand their utility. Even cost-competitive models, such as Unitree Robotics’ humanoid robots priced as low as 40,000 yuan ($5,566), remain limited in capability, and no companies have achieved profitability at current price points. More sophisticated robots for tasks like folding clothes or making coffee can cost up to 500,000 yuan, making widespread household adoption unlikely unless supply chain costs drop significantly [para. 1][para. 2][para. 3][para. 4][para. 5].

Investor interest, however, remains strong, with high hopes for embodied artificial intelligence (AI)—AI-enabled robots with physical forms—driving rapid financing and numerous firms preparing for initial public offerings. The race is fueled by the belief that embodied AI systems could transform multiple sectors, yet tremendous challenges persist [para. 6][para. 7].

A central debate at the WRC concerned the primary bottleneck in developing embodied AI. While Unitree CEO Wang Xingxing asserts that robot hardware has matured and the real constraint is lagging AI models, others in the industry disagree. Some companies invest heavily in collecting operational and simulated data to train robots, following the vision-language-action (VLA) model approach, which seeks to integrate sensory input, reasoning, and physical response. Yet, developing real-world, high-quality data at scale remains expensive and complex. X Square Robot’s CEO emphasizes both architecture and data, citing decades of trial and error favoring end-to-end, general-purpose models, while Robot Era’s founder stresses model design as the key factor, followed by data quality and hardware [para. 8][para. 9][para. 10][para. 11][para. 12][para. 13][para. 14][para. 15][para. 16].

Despite progress, embodied AI models are not sufficiently advanced to allow out-of-the-box industrial deployment, necessitating manual tuning and reliance on traditional industrial robotics technologies for many tasks. Only a fraction of robot features (e.g., advanced object recognition and dexterity) represent current embodied AI; the rest are improvements on older capabilities [para. 17].

Pricing remains a significant barrier. Even as manufacturers introduce more affordable robot models, most remain high-end products targeted at commercial buyers due to expensive computing systems and advanced components. For example, Digit Huaxia’s humanoid robot starts at 198,000 yuan. Dexterous robot hands, crucial for complex tasks, vary significantly in price and ability, with low-end models (10,000 yuan) insufficient for demanding tasks like coffee-making, necessitating upgrades to models costing 50,000 yuan or more [para. 18][para. 19][para. 20][para. 21][para. 22].

Despite limited commercial applications so far, a funding frenzy is underway, especially after notable public showcases like Unitree’s performance at the Spring Festival Gala and several high-profile funding rounds among Chinese AI developers. Much of the raised capital is directed toward massive data collection efforts, with estimates suggesting over 1 billion yuan may be needed to train a humanoid robot suitable for retail work. Local governments, such as Tianjin’s, are supporting the industry with dedicated AI data facilities [para. 23][para. 24][para. 25][para. 26][para. 27][para. 28].

Nevertheless, concerns of a speculative bubble and industry shake-out loom. Some investors warn that only companies generating substantial revenues (at least 400 million yuan) will survive the next year, with many at risk of failure as the market corrects itself [para. 29][para. 30].