人形机器人理想丰满，现实骨感？行业预测明年卖10万台，机构只看到1.2万台

Core Viewpoint - Morgan Stanley warns that the market's short-term expectations for humanoid robots are overly optimistic, facing challenges in application scenarios, hardware costs, and software intelligence [1] Group 1: Market Predictions - There is a significant divergence between industry companies and analysts regarding future humanoid robot sales, with some predicting demand in China will reach 100,000 units by 2026, while Morgan Stanley forecasts only 12,000 units [1] - By 2030, Morgan Stanley estimates the demand will only grow to 114,000 units, highlighting concerns over limited current capabilities, unattractive ROI, and various product development challenges [1] Group 2: Application Scenarios - Short-term applications will focus on specific, structured B2B verticals, with early implementations likely in commercial services and certain industrial tasks rather than general-purpose robots [2] - The efficiency of humanoid robots in industrial applications is currently only 20-30% of human efficiency, raising doubts about ROI in this sector [2] - Commercial service scenarios, which require lower efficiency and precision, are seen as more realistic early breakthroughs for humanoid robots [2] Group 3: Cost and Competition - The key driver for commercialization is cost, with customer expectations for complete robot prices ranging from 100,000 to 200,000 RMB [2] - Overemphasis on price and premature engagement in price wars could compromise product reliability and performance stability, ultimately harming the industry's healthy development [3] Group 4: Hardware Challenges - Humanoid robots face significant hardware challenges, including structural disconnection between system integrators and component suppliers, leading to unclear performance specifications and high customization costs [4] - Quality consistency and yield rates of core components remain major obstacles, with many parts still relying on manual assembly, resulting in performance variability [5] Group 5: Software Challenges - The current biggest bottleneck lies in software, where achieving general capabilities similar to GPT requires vast amounts of multimodal data, which is currently scarce [6] - The effectiveness of "Sim2Real" remains debated, and computational power is a structural bottleneck, with future models potentially needing thousands of TOPS, exceeding current edge device limits [6]