从社交媒体和资本宣传看，人形机器人大热于AI炒作新阶段，Unitree 与 Boston Dynamics 的跳跃、杂技演示以及 Tesla 与 Optimus 的叙事不断被放大，但“会跳舞”与“能替代大量人力”之间仍有巨大鸿沟。根据 Pitchbook，去年相关公司融资超过60亿美元（US$6 billion），是2024年的300倍以上；Morgan Stanley预测到2050年市场规模将达5万亿美元，并出现约10亿台设备。尽管行业愿景常被描绘为对劳动力市场的全面替代，现实仍受制于工程成本、可靠性和实际生产价值验证。
文章的核心经济动机是用通用人形形态替代高复杂度人类劳动。Persona AI 首席执行官 Nicolaus Radford 及 ALM Ventures 合伙人 Modar Alaoui都指出，按理论可让约100台机器人承担约1000名工人的任务，并以24/7运行配合 RAAS（机器人即服务）租赁，减少工会、税负和诉讼压力。中国制造业就业规模仍在数千万人，老龄化推动企业寻求以机器延长工厂寿命。Tony Seba 则把人形机器人称作“人工劳动”，估算相关市场潜力为80万亿美元，并预测5到7年后会开始替代人力。
真正难题在于可量化的工程约束：平衡、续航与控制并非可分割问题；两足站立本身就是持续动态平衡，能耗与动作修正都很高。炫目的舞蹈、跳箱、走位并不等于环境适应能力。Moravec悖论强调，AI在下围棋、数学推理上的优势并不代表其在抓取、折叠、清洁这类看似简单却依赖触觉反馈的人类长期进化技能上更强。文章举例说，像“捡起黄油刀”“折叠衣物”“顺滑擦拭”乃至协助用药、帮人起身仍是关键瓶颈。VLAM与遥操作可借海量重复训练提升技能，但巨额研发成本最终会并入最终产品价格，所谓替代低技能脏活的承诺仍不稳固。
From social media and investor hype, humanoid robots have entered a new phase of the AI hype cycle, with Unitree and Boston Dynamics demos and Tesla’s Optimus narrative repeatedly amplified, yet the gap between “can jump” and “can replace large amounts of labor” remains substantial. Pitchbook reports that last year they raised over US$6 billion, more than 300 times the 2024 level, and Morgan Stanley projects a US$5 trillion market by 2050 with roughly 1 billion units in circulation. Even though the industry narrative is often framed as replacing labor at scale, progress is still constrained by engineering cost, reliability, and proof of real operational value.
The economic rationale is to replace human labor with general-purpose humanoid forms. Persona AI CEO Nicolaus Radford and ALM Ventures partner Modar Alaoui argue that in theory about 100 robots could perform work equivalent to 1,000 workers, operating 24/7 via RAAS (robots as a service) leasing to avoid union, tax, and litigation burdens. In China, where manufacturing employment is in the tens of millions, firms are trying to extend factory life amid demographic aging. Tony Seba calls this “artificial labor,” estimates the potential market at US$80 trillion, and expects substitution to begin within five to seven years.
The core engineering challenge is in quantifiable constraints: balance, battery, and brains are tightly coupled. Even bipedal standing is continuous dynamic control, with high energy use and constant micro-adjustment from sensor feedback. Viral acts—dancing, jumping, acrobatics—do not prove adaptation to changing environments. Moravec’s paradox holds: AI is strong at abstract cognition but still weaker at tasks humans perform with deep sensorimotor learning. Folding laundry, picking up a butter knife, smooth wiping, medication handoff, and helping someone stand are still bottlenecks. VLAMs and teleoperation may improve dexterity through massive repetition, but development costs are likely to be passed into final robot prices, so the promise of replacing low-status manual work remains uncertain.