在具身智能这场高热度的创业潮中，多数玩家正涌向本体制造与大模型架构的“军备竞赛”。然而，曾任北美顶尖机器人实验室研究员的00后学者杨世琦却持相反判断：行业真正的命门不在于硬件，也不在于一味追逐模型创新，而在于如何实现训练数据规模化地供给与复用。
杨世琦在今年6月创立行元智能后，并未急于发布机器人整机或宣讲模型的宏大叙事，直到10月27日，行元智能将推出跨本体数据引擎“Chorus众行”。
Chorus试图将智能眼镜、胸戴相机、UMI设备等多源人类操作视频进行清洗与三维重建，从而能够将单份人类示范跨本体适配至多种机器人，尝试打通从人类行为到机器人动作的底层基础设施。
在Chorus发布前一周，他不仅剖析了具身智能行业中的“Demo迷思”，还指出当前行业仍处于“非标自动化”阶段。透过行元智能坚持“不做硬件本体”的策略，杨世琦讲述了团队以 System Scaling（系统规模化）为技术路径，贯通数据、模型与真实部署，探索机器人基础模型的规模化发展的底层逻辑。
创业前先诊断具身行业
2025年7月回国后，杨世琦花了数月时间深入调研国内具身智能的真实进展，发现了科研与产业上的部分差异。
在北美顶尖实验室时，杨世琦是全球最早将人类数据用于机器人训练的研究者之一。在他的认知中，正确的研究路径，应是发布第一个模型或产品时，已经能推算出下一代数据如何采集、何时面市以及如何实现规模化扩展；相比之下，他看到了行业中的部分团队更热衷于追逐大模型架构的热点，频繁发布视觉-语言-动作模型（VLA）或世界模型（World Model），但系统如何持续迭代、能力如何规模化，往往缺乏明确的路径。
“具身行业普遍偏重以Demo或Paper的形式来反映机器人学习能力，甚至部分演示背后的真实模型能力也需要打个问号。”杨世琦表示。
基于对“Demo繁荣”的警惕，行元智能选择跳出整机制造的思维定式。
正如大语言模型（LLM）的基础是“词元”（Token），而机器人基础模型要解决的是“行动”（Action）维度的智能。行元的名字也是由此而来，杨世琦希望，未来能进一步探索和定义机器人的“Action Token”。
“大语言模型已经在理解、记忆和规划上取得突破，但机器人最终必须把理解转化为动作。”杨世琦团队要做的，是从人类数据出发，建立模型与数据体系，而不是再造一台机器人。
把人类数据“扩容40倍”
机器人训练数据短缺是行业老生常谈的痛点。早期依赖遥操作采集数据，成本高昂且场景受限；后来的UMI等方案虽脱离了机器人本体限制，但仍未能彻底盘活现实世界中海量、持续产生的人类行为数据。
“人类本身就是一种具身形态，拥有丰富的运动经验和灵巧操作能力。”杨世琦认为，直接利用人类视频是破解数据瓶颈的必然方向。但人类会做，不代表机器人能直接照搬。人类视频存在手部遮挡、视角变化以及轨迹不完整等问题，加之在身体结构、运动方式等存在差异，若未经处理，绝大多数据会在质量筛选中被直接剔除。
行元智能推出的Chorus引擎，核心是要解决机器人学习中的数据规模化问题：如何让有限的人类经验，转化为大规模复用的机器人训练数据。
为此，Chorus通过原子动作识别、语义标注、三维运动重建、轨迹优化及跨本体映射，将不同来源的人类操作数据纳入统一处理管线，使同一份人类经验能够适配不同机器人本体，从而扩大有效训练数据的供给规模。
在当前测试配置下，Chorus实现了约85%的有效数据转化率。通过适配16种机器人本体，1小时经过预筛选的人类视频可生成约13.6小时的跨本体动作轨迹；进一步结合3倍指令增广，最终实现约40倍的有效训练数据扩容。
不过，杨世琦并不认为数据规模化就是简单地增加训练数据，数据与模型的Scaling需要协同设计，只有让数据的规模与质量匹配模型的训练需求，才能进一步探索机器人学习中的Scaling Law。
当然，模型能力是否随着数据规模的扩大而持续提升，最终还是要在真实场景中验证。这也是行元智能强调System Scaling的原因，需要从数据规模化出发，让模型训练和真实的场景部署相互反馈，持续推动机器人能力迭代升级。
不做本体，让模型跑得更快
在战略选择上，行元智能划定了明确的边界：坚决不做硬件本体。
“中国已经拥有成熟且庞大的机器人硬件供应链。”杨世琦解释，自建本体不仅需要投入巨额资本，还可能拖慢软件模型的迭代速度。更重要的是，工业客户需要的不是一台外形完整的人形机器人，而是一个能稳定完成特定工序的系统。
这种战略定位决定了行元智能的技术路径：不做本体，也不绑定单一硬件。
行元智能的解决思路是，模型不必等到具备近乎完美的自主能力才进入实际应用。当模型达到一定可用水平后，就可以尝试部署到真实场景中；遇到无法处理的情况时，由人及时介入，完成纠正，再将这些经验回流至训练系统。
他反复强调两件事：数据能复用多少次，模型多久能迭代一轮。
“大量工厂客户根本不关心你用了什么模型、什么算法，他们只关心你能不能把活干了。”杨世琦直言。
在真实工厂环境中，机器人极易遭遇训练数据未覆盖的分布外（OOD）状况，导致任务中断。为此，行元智能设计了一套人机协作机制：当机器人卡住时，不懂技术的普通工人可通过简单设备在2至3秒内快速接管，纠偏后交还系统继续执行，而人工接管形成的反馈会持续反哺模型。
通过这种方式，行元智能试图在任务成功率与人工成本之间取得平衡。在物流等场景中，初期可由一名工人监督一台机器人（“一拖一”），随着模型成功率提升再向“一拖三”演进。
对行元智能而言，这也是System Scaling与单纯扩大模型或者数据规模的区别，机器人能力的提升，不仅取决于训练阶段投入了多少数据，更取决于整个系统能否在真实应用中持续迭代。
未来3年是具身智能的技术窗口期
谈及当前的具身赛道，杨世琦并不避讳产业正处于资本热潮与市场宣发的竞争阶段，但作为创业新兵，他看到的是具身智能还有足够大的技术空间。
他觉得，现阶段行业应当合力把技术落到实处，而非相互争夺名义上的领先位置。
对于现今产业的发展状态，杨世琦给出客观判断：具身智能本质上仍处于“非标自动化”阶段，距离机器人真正广泛承担通用劳动力，可能还需要三到四年的沉淀。
当然，这也是每家具身智能企业实现突破的技术窗口期。
具身行业并不缺乏关于未来的想象，也不缺少惊艳的Demo，但从机器人完成一个任务到持续学习，还有很长的路要走。对杨世琦和行元智能而言，Chorus是这一探索的起点，从人类经验出发，让数据得以规模化复用，让模型在训练与真实部署中持续迭代。他们期待的，不只是机器人能够完成更多任务，而是让机器智能逐渐拥有自主成长的能力。（本文首发于钛媒体APP，文｜钱方）
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。
In this high-hot wave of smart entrepreneurship, most players are rushing to create an “arms race” with LLM structures. However, Yang Shi Qi, the post-00 scholar who was a North American top robotic laboratory researcher, decided the opposite: the real life of the industry is not in hardware, nor in the pursuit of model innovation, but in how to achieve the scaling up and reuse of training data.
When Yang Se Qi created his branch intelligence in June of this year, he did not rush into the grand narrative of the publish robot machine set-up or modeling, until October 27th, when he will launch the cross-site data engine "Chorus March".
Chorus attempted to clean and reconstruct multi-source human-operated videos such as smart glasses, chest-dressing cameras, and UMI equipment, so as to be able to adapt a single human demonstration across the body to a variety of robots and try to access the underlying infrastructure from human behaviour to robotic action.
In the week before Choruspublish, he had not only analyzed the “Demo myths” in smart industries, but also pointed out that the industry was still at a “non-standard automation” stage. In keeping with the strategy of "do not do the hardware ", through line-based intelligence, Yang Shi Qi described the team's technological path, using System Scaling (system scaleing), to connect data, models and real deployments, and to explore the underlying logic of the size development of robotic models.
Diagnosis of specific industries before starting a business
Upon his return in July 2025, Yang Shi Qi spent several months studying the real progress of the country's intellectual capacity and discovering some of the differences between research and industry.
At the top laboratory in North America, Yang Se Qi was one of the world ' s first researchers to use human data for robotic training. In his perception that the right path to research should be the first model or product of publish, it had been possible to figure out how the next generation of data would be collected, when it would be marketed and how it would be scaled up; in contrast, he had seen some teams in the industry more keen to pursue the hot spots of the LLM architecture, frequent publish visual-linguistic-action models (VLA) or world models, but there was often no clear path for how the system would last, how capacity would be scaled.
“The prowess industry generally prefers to reflect robotic learning in the form of Demo or Paper, and even some of the real modelling capabilities behind the demonstration need to be asked.
On the basis of vigilance over the "Demo boom", a swath of intelligence chooses to jump out of the whole mechanism of thinking.
Just as the Big Language (LLM) model is based on the word Token, the robotic base model addresses the intelligence of the action dimension. That is why the name of the linen is called, and Yang Shi Qi hopes that the “Action Token” of robots can be further explored and defined in the future.
“Big language models have made breakthroughs in understanding, memory and planning, but robots must eventually turn understanding into action.” The Yang Shi Qi team is going to build models and data systems from human data, not another robot.
Human data "40 times"
The shortage of data on robotics training is a constant pain in the industry. The early reliance on remote operations for data collection is costly and the scene is limited; subsequent UMI programmes, while out of robotic physical constraints, have not been able to fully integrate the large and persistent human behaviour data from the real world.
“Human beings are themselves a physical form, with a wealth of motor experience and smart operational skills.” Yang Se Qi believes that the direct use of human video is a necessary direction for breaking data bottlenecks. But humans can do it. It doesn't mean robots can take it directly. Human videos have hand masks, changes in perspectives and incomplete trajectories, as well as differences in body structure, mode of movement, etc., and, if not processed, the vast majority of them are directly excluded from quality screening.
The Chorus engine, launched by the swath intelligence, is at its core to address the problem of data scale in robotic learning: How to translate limited human experience into large-scale re-use robotic training data.
To this end, Chorus has increased the availability of effective training data by integrating human operating data from different sources into the integrated treatment tube through atomic action identification, semantic description, 3D re-engineering, trajectorization and cross-physical mapping, so that the same human experience can fit different robotics.
Under the current test configuration, Chorus achieved an effective data conversion rate of about 85 per cent. By adapting to 16 robotics, a 1-hour pre-screened human video can generate about 13.6 hours of cross-inverted action trajectory; a further three-fold extension of the instruction will result in an approximately 40-fold expansion of effective training data.
However, Yang Shi Qi does not think that the scaling up of data simply increases training data, that Scaling of data and models needs to be designed in tandem, and that Scaling Law in robotic learning can be further explored only if the size of the data match the training needs of the quality model.
Of course, whether model capacity continues to increase as the data scale expands, and ultimately in real scenes. This is also the reason for the emphasis placed on System Scaling by the Quantum Intelligence, which needs to be based on data scale-up, feedback on model training and real scene deployment, and a continuous drive to upgrade robotic capabilities.
If you don't do it, make the model run faster.
In strategic choice, the linen intelligence defines a clear boundary: it is determined not to be hard-line.
“China already has a mature and large supply chain of robotic hardware.” Yang Shi Qi explained that self-building requires not only a huge investment of capital, but may also slow down the iterative speed of software models. More importantly, it is not a human robot in its outer form that industrial customers need, but a system that can stabilize a particular process.
This strategic positioning determines the technical path of the line-intellectual intelligence: it does not make its own body and does not bind a single hardware.
The solution to this is that models do not have to wait until they have near-perfect autonomy to be applied in practice. When the model reaches a certain level of availability, it is possible to attempt to be deployed to the real scene; in cases where it is unmanageable, people intervene in a timely manner to complete the correction and then retransmit the experience back to the training system.
He repeatedly emphasized two things: how many times the data can be reused and how long the model can be repeated.
"A large number of factory clients don't care what models and algorithms you're using. They only care if you can do your job." Yang, to be honest.
In a real-work plant environment, robots are highly vulnerable to distributional (OOD) conditions that are not covered by training data, leading to disruption of missions. To this end, a human-engineer collaboration mechanism has been designed in linen intelligence: When robots are stuck, unskilled common workers can quickly take over in two to three seconds with simple equipment, and the back-to-back system continues, while feedback from manual takeovers continues to feed the model.
In this way, a linen intelligence attempts to balance task success rate with labour costs. In scenarios such as logistics, a robot (“one by one”) could initially be supervised by a worker and evolve “one by three” as the model success rate upgrades.
This is also the difference between System Scaling and simply expanding LLM or the size of the data in terms of linen intelligence, where robotic capability is enhanced not only by the amount of data invested in the training phase, but also by whether the system as a whole can continue to overlap in real applications.
The next three years are smart technology windows. Period
Turning to the current track, Yang Shi Qi did not shy away from the fact that the industry was at a stage of capital booming competition with the market, but, as a new entrepreneurship recruit, he saw that there was enough technological space for intelligence.
He felt that, at this stage, the industry should work together to put technology into practice, rather than compete for a nominal lead.
With regard to the current state of development of the industry, Yang Shi Qi has made an objective judgment that intelligence is still essentially at a “non-standard automation” stage, and that it may take three to four years for robots to truly take on the general workforce.
This is, of course, the technological window period for every furniture-smart enterprise to break through.
There is no lack of imagination about the future or of amazing Demo, but there is a long way to go from robotics to continuous learning. Chorus is the starting point for this exploration for Yang Se Qi and the Banket Intelligence, allowing data to be scaled up and reused from human experience, and for models to continue to overlap in training and real deployment. What they expect is not only that robots will be able to do more, but rather that they will be able to grow autonomously. (This article was first sent to the titanium media app, in cash)
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.