AI News AI资讯 2h ago Updated 1h ago 更新于 1小时前 56

Hardcore Exclusive: Tang Wenbin's 'Origin Intelligence' Merges with Logistics Robot Company and Secures Investment from Zhipu, SenseTime, StepFun, etc. 硬氪独家 | 唐文斌「原力灵机」并购物流机器人公司,并获智谱、商汤、阶跃等投资

When four companies—Zhipu, StepFun, SenseTime, and Alibaba—that almost represent the top-tier strength of domestic large models simultaneously invest in an embodied intelligence startup only a few months old, it is hard to view this merely as a routine financial investment. It seems more like a manifesto: a collective rush to seize the narrative of the next chapter of technology. The protagonist is "ForceMecha," with Megvii co-founder Tang Wenbin standing behind it, along with a group of familia 当智谱、阶跃星辰、商汤科技和阿里巴巴这四家几乎代表了国内大模型顶尖战力的公司,同时将钱投向一家成立仅几个月的具身智能企业时,你很难将其仅仅视为一次普通的财务投资。这更像是一份宣言:关于技术下一章叙事权的集体抢占。主角是「原力灵机」,它的背后站着旷视科技的联合创始人唐文斌,以及一批熟悉的旷视旧部。

85
Hot 热度
80
Quality 质量
75
Impact 影响力

Analysis 深度分析

This storyline is full of the fatalism of technology and the drama of business warfare. SenseTime and Megvii, these two "old rivals" who have battled for years in the computer vision era, unexpectedly "meet" at the entrance of embodied intelligence. SenseTime is betting not only on Tang Wenbin’s personal abilities but also on the extension of Megvii’s technological DNA into the physical world. Meanwhile, Zhipu and StepFun’s involvement reveals deeper anxiety: as the battlefield of large models shifts from cloud-based token processing to the precise execution of "actions" in the real world, focusing solely on the "brain" without engaging the "body" means handing over the initiative in the core battlefield. Model companies have finally realized that artificial general intelligence (AGI) without the ability to interact with the physical world is like walking on one leg—it’s destined to move slowly.

But what truly makes this chess game come alive is that silent acquisition—"ForceMecha" swallowed "Atomix," a logistics robotics company previously incubated by Tang Wenbin. This is far more than a simple business integration; it strikes at the cruelest "Achilles’ heel" of embodied intelligence: data.

The entire industry is talking about the "data flywheel," dreaming of robots becoming smarter with use. The reality, however, is that we are stuck in a logical paradox: training powerful embodied models requires massive, diverse real-world interaction data, especially data on "failures" and "corrections." Yet, if the model isn’t strong enough, robots cannot reliably enter real, unstructured environments to collect this data. It’s a chicken-and-egg problem, a deadlocked loop. "Atomix"’s accumulated real operational records from over 500 projects in warehousing and logistics—serving clients like Uniqlo and Mixue Ice Cream & Tea—and particularly the operational data corresponding to its global second-place sales of pallet four-way shuttles, instantly become the scarce fuel needed to break this "data deadlock." Tang Wenbin’s statement that "picking is the atomic task of embodied intelligence" essentially turns the arduous, grueling work of physical world data collection into a continuously value-generating "data engine."

The embodied large model "DM0" launched by "ForceMecha" proposes a "triple data fusion" technology roadmap (internet semantics, autonomous driving physical rules, and robotic operational data), which aligns precisely with this ambition of integration. It attempts to use a more "engineering-driven" approach to forcefully break down the data barriers between the virtual and physical worlds, between simulation and real operations. Achieving sub-millimeter-level operations and long-horizon tasks with only 2.4B parameters reflects the extension of "chain-of-thought reasoning" into physical space—this may be the essential distinction between embodied models and traditional industrial robot control algorithms: granting machines the ability to "think" and "plan" before execution.

However, we must maintain a sober skepticism. Behind the buzz of financing and mergers lies the collective exploration of the entire industry in the deep waters of technology. Finding the "Scaling Law" for embodied intelligence is far more complex than stacking parameters in language models. It requires high-quality and transferable physical world knowledge that bridges the Sim2Real (simulation-to-reality) gap, addressing fundamental challenges such as contact mechanics, dynamics, and environmental uncertainty. ByteDance’s heavy investment in talent and foreign star company Skild AI’s acquisitions all point in the same direction: integrating data, scenarios, models, and hardware to build an end-to-end closed-loop capability.

"ForceMecha"’s combination strategy represents a pragmatic approach with distinct Chinese characteristics: instead of waiting, it uses existing robotic businesses (Atomix) that generate cash flow to fuel cutting-edge model research, then uses breakthroughs in the model to feed back the intelligent upgrading of hardware. This "fighting while being sustained" model appears particularly practical in an industry that currently lacks mature commercial pathways. However, it also means the company must simultaneously navigate two tough battles: hardware supply chain management and project delivery on one side, and frontier model research on the other—a tremendous challenge.

The landscape of China’s embodied intelligence sector is transitioning from the early stage of "diversified demos" to a critical phase of "integrating resources to build closed loops." This move by ForceMecha perhaps signals that the core competitiveness of future players will no longer be merely outstanding algorithms or hardware, but rather the ability to build data pipelines connecting virtual and physical worlds, the engineering capability for rapid iteration, and the commercialization ability to inject model capabilities into real business scenarios. The flames of large models have irrevocably spread into the physical world. This time, leading in parameters alone is not enough—you must first learn to walk amidst the chaos of reality.

当智谱、阶跃星辰、商汤科技和阿里巴巴这四家几乎代表了国内大模型顶尖战力的公司,同时将钱投向一家成立仅几个月的具身智能企业时,你很难将其仅仅视为一次普通的财务投资。这更像是一份宣言:关于技术下一章叙事权的集体抢占。主角是「原力灵机」,它的背后站着旷视科技的联合创始人唐文斌,以及一批熟悉的旷视旧部。

这剧情充满了技术的宿命感与商战的戏剧性。商汤与旷视,这两家在计算机视觉时代缠斗多年的“老冤家”,竟然在具身智能的入口处罕见“会师”。商汤押注的不仅是唐文斌的个人能力,更是看中了旷视技术基因在物理世界的延续。而智谱与阶跃星辰的入局,则透露出更深层的焦虑:当大模型的战场从云端的Token处理,转移到现实世界中对“Action”(动作)的精准执行,只做“大脑”而不触及“身体”,意味着将核心战场的主动权拱手让人。模型公司们终于意识到,没有物理世界交互能力的通用人工智能(AGI),是一条腿走路,注定跑不快。

但真正让这盘棋活起来的,是那场悄无声息的并购——“原力灵机”吞下了由唐文斌早期孵化的物流机器人公司“Atomix”。这绝非简单的业务整合,而是直指具身智能最残酷的“死穴”:数据。

整个行业都在谈论“数据飞轮”,梦想着机器人越用越聪明。现实却是,我们深陷于一个逻辑悖论中:训练强大的具身模型需要海量、多样的真实世界交互数据,尤其是“失败”和“纠错”的数据;然而,如果模型不够强大,机器人就无法可靠地进入真实、非结构化的场景去采集这些数据。鸡生蛋,蛋生鸡,循环卡死。“Atomix”在仓储物流场景积累的超过500个项目、服务了优衣库、蜜雪冰城等客户的真实操作记录,尤其是其托盘四向车全球第二的销量所对应的实操数据,瞬间成了打破这个“数据死结”的稀缺燃料。唐文斌所说的“Picking是具身智能的原子任务”,本质上是将最艰苦、最脏累的物理世界数据收集工作,变成了一个持续产生价值的“数据发动机”。

“原力灵机”推出的具身大模型“DM0”,其宣称的“三类数据大融合”(互联网语义、智能驾驶物理规则、机器人实操数据)技术路线,也恰恰呼应了这种整合的野心。它试图用一种更“工程化”的方式,强行打通虚拟与物理、模拟与实操的数据壁壘。而仅用2.4B参数就实现亚毫米级操作和长程任务,其背后是“思维链推理”向物理空间的延伸——这或许才是具身模型与传统工业机器人控制算法的本质分野:赋予机器在执行前进行“思考”和“规划”的能力。

然而,我们必须保持一种清醒的怀疑。融资与合并的热闹背后,是整个行业在技术深水区的集体摸索。找到具身智能的“Scaling Law”(规模定律),远比在语言模型上堆参数复杂。它需要的是跨越Sim2Real(模拟到现实)鸿沟的、高质量且可迁移的物理世界知识,这需要解决接触力学、动力学、环境不确定性等一系列根本性难题。字节跳动重金挖人、海外明星公司Skild AI并购业务,巨头们的动作如出一辙,都指向同一个方向:整合数据、场景、模型与硬件,构建端到端的闭环能力。

“原力灵机”的组合拳,代表了一种颇具中国特色的务实路径:不等不靠,先用已有的、能产生现金流的机器人业务(Atomix)喂养最前沿的模型研发,再用模型的突破反哺硬件的智能化升级。这种“以战养战”的模式,在行业普遍缺乏成熟商业落地路径的当下,显得尤为实际。但它也意味着,公司必须同时驾驭硬件供应链管理、项目交付和最前沿的模型研究两场硬仗,挑战巨大。

中国具身智能赛道的格局,正从“百花齐放做demo”的草创期,进入“整合资源打闭环”的攻坚期。原力灵机的这步棋,或许预示着未来玩家的核心竞争力,将不再仅是某一项突出的算法或硬件,而是连接虚拟与物理世界的数据管道能力、快速迭代的工程化能力,以及将模型能力注入真实商业场景的商业化能力。大模型的战火,已经不可逆转地烧向了物理世界。这一次,光在参数上领先是不够的,你得先学会在现实的混沌中行走。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

机器人 机器人 大模型 大模型 融资 融资
Share: 分享到: