文 | Tech商业
近期，字节跳动在其 Seed Model 部门密集更新了公开的招聘信息。数十个专业岗位被同时放出，覆盖面极广：从硬件设计、灵巧操作、减速器控制，到物理模拟和数据算法工程。
这释放了一个极其明确的信号：字节不再只满足于通过财务投资参与具身智能，而是正试图亲自下场，打通一条面向人形机器人的“软硬件+数据”全栈管道，推动产品从“制作演示”真正走向“大规模生产”。
与此同时，宏观背景极其宏大：中国在2025年制造了全球80%以上的人形机器人。摩根士丹利预测，中国2026年、2027年人形机器人出货量分别约为5万台和10万台；工信部口径则显示，2026年上半年产量已超4万台，预计全年突破10万台。在这个关键节点，字节的入局顺理成章，却也充满挑战。当习惯了“敏捷迭代”的互联网巨头，撞上“重资产、零容错”的硬核制造业，一场关于技术、供应链与商业闭环的残酷大考才刚刚开始。
从大脑到身体
字节入局人形机器人，绝非心血来潮。2025年6月，字节系基金锦秋基金参与了国内头部机器人公司宇树科技的C轮融资。但从“财务投资”转向“亲自下场”，意味着字节对AI的终局有了更深的焦虑与渴望。
在大模型时代，字节凭借豆包等产品已经在“大脑”（算法与模型）层面证明了自己。但如果没有一个能在物理世界里移动的实体，AI的边界就永远被锁在屏幕里。大模型拥有丰富的世界知识，却无法直接理解摩擦力、惯性以及真实的因果关系。将AI具象化为可交互的物理实体，是必由之路。
从字节放出的岗位来看，其野心不可谓不大。硬件设计、灵巧操作、减速器控制，说明字节试图吃透机器人的底层物理结构与运动控制；物理模拟、数据算法，则依托于字节极其强大的算力储备和算法基因。
在字节的版图中，具身智能不是孤立的产品，而是其整个AI生态的自然延伸。他们试图打通“大模型推理（Seed）—物理模拟训练—硬件本体执行”的闭环。在这个闭环中，字节最大的优势在于软件和数据。依靠庞大的算力集群，字节可以在虚拟环境中构建“仿真世界”，通过强化学习让机器人在仿真中“试错”，再将能力迁移到现实世界。这恰恰是传统硬件公司不具备的能力。
量产鸿沟
尽管字节的算法能力毋庸置疑，但摩根士丹利等机构近期的报告却泼了一盆冷水。报告指出，从实验室里的“演示视频”到真正进入工厂替代人工，规模部署的时间表正在后移。企业的采用意愿仍在，但从试点走向规模部署需要更长时间。核心阻碍有三：数据短缺、硬件不可靠，以及机器人在预设测试之外频繁发生故障。
这是所有互联网大厂跨界做硬件必然要面对的鸿沟。
互联网的逻辑是“小步快跑，快速迭代，容忍Bug”。但制造业和物理世界的逻辑是“一次做对，安全至上，零容错”。要让机器人在真实的工厂流水线上拧螺丝、搬箱子，你需要喂给它海量的真实物理交互数据。而在现实中，采集高质量数据的成本极其高昂。
在仿真环境里训练出来的机器人，一旦遇到摩擦力变化、光线变化，或者一个不规则物体，往往就会当场“死机”。这就是行业里最著名的“Sim-to-Real”（仿真到现实）鸿沟。
此外，人类的手指拥有极高的自由度和触觉反馈，现有的机器人灵巧手在寿命、力度控制上，依然面临极高的工程挑战。电机、减速器、传感器等核心零部件的可靠性，直接决定了机器人的寿命和成本。
字节需要招募的，不再仅仅是坐在格子间里写代码的极客，而是懂机械工程、材料学、供应链管理的实干家。将最顶尖的AI算法，与最沉重的齿轮和液压装置结合，没有“大力出奇迹”的捷径，只有一步步踏踏实实地做工程试错。
供应链底座
在这场竞赛中，中国制造业展现出了极其恐怖的底座实力。
据相关数据显示，中国在2025年制造了全球80%以上的人形机器人。这意味着，从供应链的完整度、制造成本的压缩空间，到产品迭代的试错场，中国都是全球具身智能最肥沃的土壤。在长三角和珠三角，几乎可以找到制造一台人形机器人所需的所有核心零部件，且响应速度极快。
与此同时，宇树科技（Unitree）等中国本土独角兽也在近期发布了最新一代先进的模型和产品。从四足机器狗到人形机器人，中国企业在运动控制（四肢）上已经走在了世界前列。
字节此时的加入，无疑将加剧这一赛道的竞争。大厂入场，不仅能带来充裕的资金，更能将大模型最前沿的智能注入到行业之中，加速整个产业的成熟。但挑战同样巨大：如何在众多已深耕多年的本土机器人企业中脱颖而出，如何利用自身的软件优势形成差异化，是字节亟待解答的难题。
卖铲人红利
在巨头们轰轰烈烈的造“人”运动中，谁最有可能在短期内实现商业闭环？
摩根士丹利给出了一个极其冷酷的答案：当科技巨头斥资数十亿美元追逐科幻人形机器人时，目前唯一有把握的赢家，是向它们出售齿轮、轴承和液压装置的传统供应商。
这完美复刻了AI大模型时代的“卖铲人”定律。当字节、特斯拉、Figure等公司在软件和本体上打得头破血流时，那些默默造减速器、伺服电机、传感器的核心零部件厂商，正在稳扎稳打地收割着巨头们溢出的红利。因为无论哪家大厂率先突破量产，都必然需要向这些底层供应链企业采购核心零部件。在硬件革命的早期，卖铲子永远是最确定性的生意。
闭环难题
抛开技术层面的挑战，人形机器人面临的最核心问题依然是：商业闭环。
大厂投钱做研发，最终是为了赚钱。人形机器人的成本极高，一台动辄几十万上百万的售价，让绝大多数企业望而却步。在当前技术下，人形机器人能做的事情，往往与它的高昂成本不匹配。
字节的入局，或许会走一条不同的商业化路径。依托其庞大的内容生态和商业场景，字节可能会优先把机器人放进内部物流、仓储等已验证环节，再逐步探索直播、内容制作等方向。目前可证实的是，字节物流机器人已量产超千台，主要用于抖音电商仓的分拣与运输，并有顺丰、比亚迪电子等外部客户；人形机器人进入直播、内容制作仍只是想象空间。先通过内部消化降低成本，再逐步向外部企业输出。
但这依然是一条漫长的路。从实验室的演示视频到量产交付，从工厂的流水线到千家万户，还有无数个关于齿轮、电机、数据和成本的无眠之夜需要熬过。
结语
字节跳动全力招募机器人工程师、构建具身智能硬件团队，是中国科技巨头向硬科技深水区迈进的一个生动缩影。
这条路极其漫长，甚至比大模型还要烧钱。它提醒所有人：在物理世界，互联网式的“魔法”很难立刻奏效，所有的算法最终都要接受重力的检验。
字节的入局，或许只是这场宏大变革的一个开始。在这场竞赛中，真正的赢家，将是那个能在算力与重力之间，率先找到商业平衡点的企业。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Man, Tech Business
Recently, byte beats an intensive update of open recruitment information in its Seed Model department. Dozens of professional posts were released at the same time, covering a wide range of areas, from hardware design, smart operations, speed-reducing controls, to physical simulation and data algorithms.
This sends an extremely clear signal that byte is no longer content to be involved in smartness through financial investment, but is trying to end up in person by opening a “soft and hardware + data” whole-barrel pipe for human robots and pushing the product from “produce” to “large scale production”.
At the same time, the macro-level context is extremely ambitious: China created more than 80 per cent of the human robots in 2025. Morgan Stanley predicts that in 2026 and 2027, respectively, China will have about 50,000 and 100,000 man-shaped robots, while the Ministry of Industry and Communications's caliber indicates that in the first half of 2026 it has produced more than 40,000 and is expected to exceed 100,000 throughout the year. At this critical node, the entry of byte is a logical but challenging one. A brutal test of technology, supply chain and business closure has only just begun when the Internet giants, used to be “agile” and hit “heavy assets, zero tolerance” hard-core manufacturing.
From brain to body.
byte human-shaped robots are not in the mood. In June 2025, byte was a member of the Fund's Kinyu Fund, which participated in the C-wheel financing of the leading robotics company Oki Tree Technology in the country. But the shift from “financial investment” to “personal exit” meant that byte had a deeper anxiety and desire for the final outcome of AI.
In the era of LLM, byte has proven itself at the “brain” (calculations and models) level with products such as Doubao. But without an entity that can move in the physical world, the boundaries of AI are locked in the screen forever. LLM is rich in world knowledge, but does not understand friction, inertia and true causality. The transformation of AI into an interactive physical entity is an imperative.
In terms of the posts offered by byte, its ambition cannot be insignificant. Hardware design, smart operation, speed-reducing control, which describes byte ' s attempt to penetrate robotic bottom physical structure and motion control, and physical simulation, data algorithms, which depend on byte ' s extremely powerful algorithmic reserves and algorithmic genes.
In byte, intelligence is not an isolated product, but a natural extension of its entire AI ecology. They tried to get through the close of the “LLM reasoning” - Physical simulation training - implementation of the hardware matrix”. The greatest advantage of byte in this ring lies in software and data. Based on a large array of algorithms, byte can build a “simulate world” in a virtual environment, where robots can “try wrong” in a simulation through enhanced learning and then migrate their capabilities to the real world. This is precisely what traditional hardware companies do not have.
The Digital Divide
Despite the unquestionable ability of byte, recent reports by agencies such as Morgan Stanley have spilled a cold water. The report indicates that the time frame for deployment is moving back from “demonstration videos” in the laboratory to real access to the plant to replace labour. The willingness to adopt remains, but it will take longer to move from pilot to scale. There are three core barriers: data shortages, unreliable hardware and frequent failure of robots outside preset tests.
This is the gap that all large Internet factories have to face when they cross-border hardware.
The logic of the Internet is “slow-paced, fast-tempered, tolerant Bug”. But the logic of manufacturing and the physical world is that “one at a time is right, one at a time is safe and one at a time”. You need to feed the robot a big amount of real physical interactive data to screw and move the case on the real plant line. In reality, the cost of collecting quality data is extremely high.
The robots trained in the simulation environment tend to be “dead machines” on the spot once they encounter friction, light or an irregular object. This is the most famous “Sim-to-Real” gap in the industry.
In addition, human fingers have a high degree of freedom and touch feedback, and the available robotic smart hands continue to face extremely high engineering challenges in terms of longevity and strength control. The reliability of core parts such as electrics, speed-retarders and sensors directly determines the life and cost of robots.
byte needs to be recruited, not just as a hacker who writes code in a grid room, but as a layman who knows mechanical engineering, material science, supply chain management. Combining the best AI algorithms with the heaviest gears and hydraulic devices, there is no shortcut for “demonstrating miracles” and only a step-by-step engineering error.
Bottom of the supply chain
In this competition, Chinese manufacturing has shown the most horrific undersea power.
According to relevant data, China manufactured more than 80 per cent of the human robots worldwide in 2025. This means that China is the most intelligent soil in the world, from the integrity of the supply chain, to the condensed space of manufacturing costs, to the faulty field of product over time. In the long and bead triangles, almost all the core parts required for the manufacture of a human robot can be found and the response is very rapid.
At the same time, indigenous Chinese unicorns such as Unitree have recently developed the latest generation of advanced models and products. From a four-leg machine dog to a human robot, Chinese companies are already at the forefront of the world in terms of motor control (tangular limbs).
The accession of byte at this time will undoubtedly intensify competition on this track. The entry of large plants will not only generate sufficient funds, but will also inject the cutting edge intelligence of LLM into the industry and accelerate the maturity of the whole industry. But the challenge is equally enormous: how to emerge from many indigenous robotics that have been working for years, and how to use their own software advantages to make a difference is an urgent challenge for byte.
Selling forklifts.
Who is most likely to achieve commercial closure in the short term in the burgeoning “man” movement of giants?
Morgan Stanley gave an extremely cold answer: when the technology giant spent billions of dollars chasing science fiction robots, the only sure winner is the traditional supplier of gears, bearings, and hydraulic devices.
This is a perfect recapitulation of the law of the AZLLM era of "selling people." When companies such as byte, Tesla and Figure were bleeding on their software and on-the-board, the core parts manufacturers of silent brakes, servers and sensors were steadily reaping the dividends of the giants. Because, regardless of which major plant is the first to break through, it is inevitable that core parts will have to be procured from these bottom supply chain enterprises. In the early days of the hardware revolution, the sale of shovels was always the most certain business.
The closed ring puzzle.
Leaving aside the challenges at the technical level, the most central problem for human robots remains: commercial closure.
Large factories invest in research and development, ultimately to make money. The cost of human-shaped robots is extremely high, with hundreds of thousands of dollars at which the vast majority of businesses are deterred. Under current technology, what human robots can do is often not commensurate with their high cost.
The entry of byte may follow a different commercial path. Based on its vast content ecological and commercial landscape, byte may prioritize the placement of robots into proven links such as in-house logistics, warehousing, etc., and gradually explore the direction of live broadcasting, content production, etc. It is now confirmed that byte logistics robots have produced more than 1,000 units, mainly for the sorting and transport of Douyin power warehouses, and have external customers such as Sunfung, Biadi electronics; human robots enter live and content production remains only imaginary space. Reduced costs through internalization and then progressively exported to outside firms.
But it remains a long road. From laboratory presentations to mass delivery, from water lines in factories to thousands of households, there are countless sleepless nights for gears, electrics, data and costs.
Concluding remarks
byte beats to recruit robotic engineers and build a body-smart hardware team, which is a vivid microcosm of China ' s technological giants moving towards a hard-tech deep-water zone.
The road is extremely long, even burning money more than LLM. It reminds everyone that in the physical world, Internet-style “magics” are difficult to work immediately, and all algorithms are ultimately subject to gravity tests.
The entry of byte is perhaps just the beginning of this grand change. In this competition, the real winner will be the one who can lead the way in finding a business balance between calculus and gravity.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.