新智元报道
启元机器人刚刚下线5000台！
启元机器人具身研究中心承接底层算法研发，对外锚定一个清晰目标：做出第一个卖给个人的物理 AI机器人。
在行业普遍把目光投向远期人形机器人的同时，启元团队判断，Physical AI 是大语言模型走向真实物理世界的关键中间环节，2027 年或将迎来这一赛道的关键突破窗口。
已经量产上市的 Q1、T1 是启元面向 C 端交出的第一代硬件产品，但启元并不把这两台机器视作家务作业机器人的终局。
Q1 定位小尺寸全身力控人形机器人，T1 主打轮足‑四足可变形形态，二者更多解决硬件稳定性、安全入户、陪伴交互的问题；真正面向家务作业能力的轮式机器人 H 系列仍在迭代，与之配套的 「启元大脑」，才是整套体系的核心。
对比海外路线，π 路线依靠真机反复试错，任务成功率高，但很难规模化、泛化有限；UMI 类手持采集容易放大样本规模，泛化佳，但任务成功率很难自然拉满。
启元试图把两条路线融合，保障真机成功率的同时，通过便携式采集解决规模化问题。
启元的技术框架建立在快慢脑分层之上，这套思路借鉴人类双系统认知逻辑，把机器人的思考与执行拆开处理。
GPT‑6 Astra 以及其他大语言模型被定位为 S2 慢脑，承担高层任务理解、逻辑推理、长任务规划、记忆调度，相当于机器人的思考中枢。
在视频号演示案例中，机器人既要理解大富翁游戏复杂规则，又能中途切换叠衣服任务，完成操作后回到游戏继续推进，这类跨任务长周期的认知能力，正是慢脑的价值所在。
但团队明确，通用大模型不善于直接接管底层动作执行。如果完全依靠 GPT 类模型做抓取、整理桌面这类精细操作，每一步细微位移都需要反复思考推理，单次抓取耗时可达半小时，同时带来极高 Token 成本。
而且其他更复杂的操作难以实现，思维大脑不具备「作业手感」。高层规划交给慢脑，真实世界的动作生成，要交给自研 S1 快脑 —— 视觉‑语言‑动作模型 VLA。
VLA 负责叠衣服、整理桌面、柔性物体操作，接收慢脑下发的任务指令，快速输出高频连续动作。类似人类大脑分区，二者形成协作：慢脑定目标，VLA 负责落地执行。
GPT 更像场外通用教练，真正上场完成实操的运动员，是 VLA、机器人本体，以及围绕本体沉淀下来的遥操作、UMI示范数据。通用大模型迭代可以放大现有模型能力，但无法替代这套和硬件深度绑定的具身能力。
团队同时提醒，短期很难出现一套可以通吃全部硬件的 VLA 模型。类比自动驾驶，即便场景复杂度低于机械臂物理交互，行业至今也没有真正跨车型通用的自动驾驶模型。
机器人要面对摩擦、形变、柔性物体、密集接触，VLA 必须和特定本体做耦合训练，模型、数据、硬件三者必须绑定迭代。未来不管基座模型来自内部自研还是外部合作，都必须在自有硬件平台完成数据采集、训练与验证。
当前启元对 GPT 仅做 API 调用，长期来看，如果通用大模型能够读懂视频流、空间关系与动作序列，就有机会充当强化学习里的Reward模块，充当通用监督信号，帮助加速家务任务成功率的提升。
具身机器人的一大现实瓶颈来自数据采集。如果只是单纯堆示范数据不等于补齐模型短板，难点在于精准定位模型不会什么，针对性补充样本。
传统 HG‑DAgger 模式及其衍生方法，需要机器人先执行任务，出错之后再由人类接管纠正，纠正数据回流模型训练以不断修正模型效果，让成功率逼近100%。
然而这套流程高度绑定真机，迭代速度被硬件执行速度束缚，采集效率很难拉高，难以将百台、千台乃至万台机器部署到不同的地方采集。
启元最近推出Hil-UMI范式，试图解开真机枷锁。UMI本身是一套无本体便携式夹爪，不需要完整机器人本体到场，人类操作者手持便携夹爪直接完成任务示范。这个方式适合规模化数据采集，但过去并不能指导人类针对模型薄弱环节采集数据。
Hil-UMI进一步增加薄弱环节识别机制：人类数采员一边采集数据的过程中、模型分析哪些时候哪些动作需要进一步加强、指挥人类数采员补录数据。
在整理桌面的对照实验中，Hil-UMI的数据采集效率达到传统真机 HG‑DAgger 的5.63 倍，任务最终完成度也更高。
左图为四项任务平均分随累计采集时间的变化；右图为整理桌面任务中各方法随训练阶段的表现
这套机制不再被动等待机器人犯错，采集阶段就瞄准模型盲区。新采集的数据先更新模型，新旧样本混合参与训练，形成 「识别短板‑定向采集‑模型更新‑再识别短板」 的闭环。
正如之前所说，UMI夹爪本身具备一定本体解耦能力，更换法兰盘即可适配不同机械臂。这为未来众包采集埋下伏笔：不需要大量整机设备运送到各地，只分发便携式采集单元，就可以在成千上万不同家庭、不同环境下收集人类操作轨迹。
启元还反复强调：数据质量优先于数量。劣质样本不止消耗人力，更会大量浪费 GPU 算力，造成训练 Loss 异常。因此算法团队和数据采集团队从源头协同，把模型需要的先验知识前置到采集环节。
模仿学习存在能力天花板。团队拿 AlphaGo、大模型 RLHF 做类比：只学示范数据很难突破人类水平，机器人同样需要后训练、强化学习。
HIL-UMI范式的核心并不在于使用哪个底座模型，而是证明，在不过度依赖真机的前提下，依靠UMI采集链路，也可以开展后训练，提升任务成功率。
按照现在的爬坡节奏，启元希望 2027 年自身数据规模冲击百万小时级别。但团队清醒，百万小时距离真正通用具身智能仍有距离。想要接近 GPT‑6/GPT‑7 级别的物理世界泛化，需要千万甚至上亿小时量级数据。
以叠衣服为例，想要做到家庭场景稳定泛化，至少要覆盖一万个家庭、几十万件不同衣物。数据能力提升也不是简单线性增长，当单一场景数据饱和之后，继续增加小时数，边际收益快速衰减，关键要看任务、物体、环境的多样性。
算力层面，启元现在处于数千卡集群规模，与大模型公司相比并不多，但具身智能训练和大语言模型存在本质差异：训练主体是高冗余视频流，帧与帧之间大量信息重复，有效信息密度低。在 5000 卡规模、40% MFU 利用率的条件下，传统不做特殊优化的训练，跑完一个 Epoch 就要 3‑6 个月。
这催生了对视频训练基础设施的迫切需求。不能简单复用图片、文本大模型的训练管线，需要做筛选、压缩、表征重构。
未来的目标是把上下文窗口从十几秒拉长到小时、甚至天级别，让模型可以完整消化机器人一整天的任务经历，这才会带来 In‑Context Learning 层面的质变。
消费级家务机器人最大陷阱，是实验室演示与家庭真实体验之间的鸿沟。
启元机器人给出一条硬性判断：面向普通家庭的 C 端功能，任务完成度需要做到95% 以上才具备商业化条件，89%、94% 对用户而言依然是不合格半成品，消费者不会为需要频繁人工介入的产品付费。
因此，公司策略不是一开始追求全任务通用，而是集中资源先打穿一个家庭 Killer Application。以衣物处理举例，完整闭环包含收衣、送洗、晾晒、折叠归位全链路，单点碎片化功能没有产品价值。
Q1、T1 已经量产，但它们并非原生面向家务作业的本体。未来承担家务作业的是新一代轮式 H 系列，拥有更强手臂、负载与自由度。
启元Q1
启元T1
除直接家庭场景，团队也在寻找 「家庭同构环境」：酒店、样板间这类结构化空间，任务复杂度可控，可以作为商业化验证场。
另外一条长线路径是 Robot‑Ready Home。机器人很难一步适应所有存量家庭，反过来，先控制好泛化边界、并让家居环境适配机器人。
启元在 2026 上半年和尚品宅配成立合资子公司，探索和家装、地产结合的方案。先在边界可控样板间打磨能力，再向普通家庭渗透。版本会滚动迭代，计划 2026 年底发布大脑端初步成果，后续持续对外放出 Zero‑Shot 泛化基准测试。
To C 本体还有一层容易被忽略的价值：它本身就是 Physical AI 时代的基础设施，类似大模型时代的 GPU。当终端规模走到万台、十万台级别，在合规、用户明确授权前提下，终端回流的视频、图像数据规模，会远超外包采集或者 B 端项目所能拿到的数据。
当然，Q1/T1 现阶段的数据不能直接拿来训练，需要等待采集设备配套改造，再通过 「机器人养成计划」 引导用户自愿参与数据贡献。
C 端硬件量产爬坡同样是高门槛。从十几台样机到千台、万台交付，良率、供应链、制造和模型迭代协同，每一步都是难关。启元认为自己的优势，是已经率先跑通消费机器人的规模量产链路。
市场普遍存在一种疑问：当 GPT‑7、GPT‑8 能力进一步跃升，甚至内置 VLA 能力，会不会直接吃掉具身机器人创业公司的生存空间。
启元的判断是，GPT 这类大模型属于行业公共资源，所有人都可以调用。
真正拉开差距的不是 「能不能用上强通用模型」，而是手里有没有配套的本体、高质量操作数据集、完整训练闭环。
语言大模型提供高层认知，但缺少 「燃料」—— 也就是和硬件绑定的物理交互数据，再强的规划能力也无法落地。OpenAI 自身也要布局硬件，正是出于同样逻辑。
算力不是当下最主要瓶颈，千卡到万卡集群现阶段已经基本够用。行业真正稀缺的是高质量具身数据、更高效率的数据采集与后训练范式。国内在数据供给上具备地缘优势。
对比海外路线，π 路线依靠真机反复试错，任务成功率高，但很难规模化、泛化有限；UMI 类手持采集容易放大样本规模，泛化佳，但任务成功率很难自然拉满。启元试图把两条路线融合，保障真机成功率的同时，通过便携式采集解决规模化问题。
完整的 Physical AI 体系分为三层：S0 运动控制、S1 执行层 VLA、S2 高层认知规划。
启元机器人团队有大规模大模型训练经验，处理过数十万小时级别数据。未来的理想形态类似苹果模式：硬件终端、数据入口、模型能力三者深度耦合，而不是单纯输出算法模型。
风险同样清晰摆在面前：百万小时数据能否如期爬坡；训练基建的工程难度；C 端硬件大规模量产良率；家务场景能不能跨过 95% 体验门槛；Robot‑Ready Home 和地产家装的商业模式能否跑通。
Physical AI 是一条长链条，任何一个环节滞后，都会拖累整体落地节奏。
对于启元而言，Q1、T1 只是入场券，物理AI的真正的战争发生在硬件、数据、模型互相咬合的闭环之中。
编辑：Aeneas
Read more
不喜欢
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New Ji-won.
The droids just dropped 5,000!
The initial robotic body research centre is responsible for the development of bottom algorithms and has a clear objective: to make the first physical AI robot to sell to individuals.
While the industry is generally looking at long-term human-shaped robots, the Enlightenment team judges that Physical AI is the key middle link between a large-language model and a real physical world, or a critical breakthrough window in 2027.
Q1, T1, which is already on the market in volume, is the first generation of hardware to be handed over to the C end, but Kaimoto does not see these machines as the end of a domestic robot.
Q1 The small-sized body-force-controlled human-shaped robot, T1, is a four-footer deformation, and the two more address hardware stability, safe entry, and companion interaction; the series of wheel-based robots H, which is really oriented towards domestic work, is still iterative, and its accompanying "start brain" is at the heart of the entire system.
In contrast to overseas routes, the Zen route relies on real machine error and the task success rate is high, but it is difficult to scale up and extend; the UMI-type hand-held collections can magnify the size of sample and expand well, but success rate is difficult to fill naturally.
It seeks to integrate the two routes and to ensure that the real machine success rate is accompanied by portable collection to solve the problem of scale.
The technological framework of the Enlightenment is based on the fast and slow brain stratification, an approach that draws on the dual-system cognitive logic of human beings and removes robotic thinking and execution.
GPT – 6 Astra and other large language models have been positioned as S2 slow brain, with high-level tasks of understanding, logic, long-term mission planning, and memory scheduling, which are equivalent to robotic thinkers.
In the video number demonstration case, robots understand both the complex rules of the Monopoly game and the ability to change clothes midway and move back to the game when they are operational, and the cognitive ability to cross the long mission cycle is the value of slow brain.
However, it was clear to the team that General LLM was not good at taking over the ground operations directly. If you rely exclusively on GPT-type models for fine operations such as capturing, sorting desktops, you need to think repeatedly about reasoning for every step of micro-division, which can take up to half an hour for a single grab, with very high Token costs.
And other, more sophisticated operations are difficult to achieve, and the mind doesn't have a sense of "operational hand." High-level planning is given to slow brain, real-world action generation, self-study S1 fast brain -- visual-language action model VLA.
VLA is responsible for folding clothes, collating desktops, handling soft objects, receiving slow-drive mission instructions and rapidly exporting high-frequency continuous operations. Similar to the human brain division, the two form a collaborative effort: slow brain targeting, VLA is responsible for landing.
GPT is more like an off-site general coach, a real player who comes to practice, a VLA, a robotic body, and a remote operation, UMI demonstration data, surrounding the sedimentation of the body. The generic LLM anecdotal is able to amplify existing model capabilities, but is not a substitute for the physical capability of this set and hardware depth binding.
At the same time, the team cautioned that it would be very difficult in the short term to have a VLA model that could eat all the hardware. Comparable automatic driving, even if the scene is less complex than the physical interaction of the mechanical arm, does not yet have a truly universal cross-carriage automatic driving model.
In order for robots to face friction, shape change, soft objects, intensive contact, VLA has to be trained in combination with a specific body, and models, data and hardware have to be tied to an iterative pattern. In the future, data collection, training and validation will have to be completed on the own hardware platform, regardless of whether the base model comes from in-house self-study or external cooperation.
The current start-up is an API call only, and in the long term, if General LLM can read the video stream, spatial relationships and action sequences, it will have the opportunity to serve as a Reward module in intensive learning, serving as a universal monitoring signal to help accelerate the advancement of success rate for household tasks.
A major practical bottleneck for robotics comes from data collection. If a mere stack of demonstration data is not equal to the completion of the model panel, the difficulty is that precision positioning models do not do anything and are targeted to complement sample.
The traditional HG-DAgger model and its derived methods require robotics to perform their tasks before humans take over to correct them, correct data flow model training to constantly modify model effects and bring success rate closer to 100 per cent.
However, the process is highly bound to the real machine, the iterative speed is constrained by the speed of hardware implementation, and collection efficiency is difficult to increase and it is difficult to deploy hundreds, thousands or thousands of machines to different locations for collection.
Kaimoto recently launched the Hil-UMI paradigm in an attempt to untie the real chain. The UMI itself is a portable set of hand-held claws, which do not require the presence of a complete robotic body, and human operators carry hand-held claws directly to the mission. This approach is appropriate for scale data collection, but has not guided human collection of data on model weaknesses in the past.
Hil-UMI further increases vulnerability identification mechanisms: the process of collecting data by humans, modelling when actions need to be further strengthened and directing humans by humans.
In a comparison experiment to organize desktops, Hil-UMI data collection was 5.63 times more efficient than the traditional HG-DAgger, and the final completion of the task was higher.
Left figure averages the four tasks with the cumulative time of collection; right chart collates the performance of the desktop tasks with the training phase
The mechanism is no longer passive waiting for a robot to make a mistake, and the collection phase is aimed at the model blind. The newly collected data first update the model, the old and new sample mix of training, and form the closed circle of the "Identification Short Panels - Directional Collection Model Update - Identification Shortboard."
As stated earlier, the UMI claws themselves have a certain level of introspection capability, and the replacement of the French plate would fit different mechanical arms. This provides an ambush for future crowd packages: human operational tracks can be collected in thousands of different homes and environments without the need to transport large quantities of machine equipment to local locations and with portable collection units only.
It was also repeatedly emphasized that data quality takes precedence over quantity. The poor quality of sample is not only a labor drain, but also a huge waste of GPU calculus, causing training for Los. The algorithm team and the data acquisition team therefore work together from the source to pre-empt the a priori knowledge required for the model to be collected.
imitation of the existence of a competence ceiling. The team uses AlphaGo, LLM RLHF as an analogy: modeling data alone is very difficult to reach, and robots need post-training and intensive learning.
The core of the HIL-UMI paradigm is not the bottom model, but rather the proof that, without over-reliance on real machines, the UMI pick-up chain or post-training can be conducted to upgrade task success rate.
At the current rate of climb, it is hoped that in 2027, the scale of its own data will hit a million hours. But the team is awake, and a million hours is still a distance from real universal intelligence. It takes millions, if not hundreds of millions of hours, to get close to GPT-6/GPT-7 levels of physics.
In order to stabilize the family landscape, for example, folding clothes must cover at least 10,000 families and hundreds of thousands of different clothing items. Nor is there simple linear growth in data capacity, where the number of hours continues to increase after a single scene is saturated and marginal gains decline rapidly, depending on the diversity of tasks, objects and environments.
At the arithmetical level, the initials are now on the scale of thousands of calories, not much compared to LLM, but are fundamentally different in terms of smart training and large language models: The main subject of the training was a high redundancy video stream, with a large amount of information duplicated between frames and frames, and a low level of effective information density. Under 5,000 card sizes and 40% MFU utilization, traditional training without special optimization takes 3 to 6 months to run an Epoch.
This has created an urgent need for video training infrastructure. The training lines of LLM, which cannot be simply duplicated, need to be screened, compressed and reformed.
The goal of the future is to stretch context window from a few seconds to an hour, or even a sky level, so that the model can fully absorb the robot's entire day of mission experience, which will lead to a qualitative change at the In-ContextLearning level.
The biggest trap for consumer-grade domestic robots is the gap between laboratory demonstrations and the real experience of families.
The droids give a hard judgement: C-end functions for ordinary families, which require more than 95 per cent completion of the mission to be commercialized, 89 per cent, 94 per cent of which remain unqualified semi-finished products for users, and consumers do not want paid for products that require frequent manual intervention.
The corporate strategy, therefore, is not to start by pursuing mission-wide universality, but to pool resources to penetrate a family, Killer Application. In the case of clothing treatment, for example, the complete closed ring contains a full chain of clothing collection, washing, drying and folding, and single-point fragmentation has no product value.
Q1, T1 have been produced in volume, but they are not home-oriented. In the future, it is the new generation of rotational H series with stronger arms, loads and freedom.
KaimotoQ1
Initial T1
In addition to the direct home scene, the team is also looking for "family co-constructive environments": structured spaces like hotels, templates, tasks that are manageable and can serve as commercial validation sites.
The other long line is Robot-Ready Home. It is difficult for robots to adapt to all households in stock one step at a time and, in turn, to control borders and adapt robots to home environments.
In the first half of the year 2026, the initial capital was set up as a joint-venture subsidiary to explore a combination of home and property. The ability to grind between controllable templates at the border first penetrates ordinary families. The version will be rolling in an iterative fashion, and it is planned that the initial results of the publish brain end of the year 2026 will be followed by the continuous external release of the Zero-Shot generic benchmark test.
ToC has another easy value: it is the infrastructure of the Physical AI era itself, similar to the GPU of the LLM era. When the terminal size reaches 10,000 or 100,000 levels, the size of the video and image data flow of the terminal, subject to compliance and clear user authorization, will go well beyond what is available from outsourced collection or B-end projects.
Of course, Q1/T1 data at this stage cannot be used directly for training, and needs to wait for the retrofitting of the collection equipment to guide users to volunteer data contributions through the robotization program.
C The terminal hardware is also a high threshold. From a dozen prototypes to thousands and thousands of deliveries, each step of good rates, supply chains, manufacturing and modelling is difficult to achieve. The initial advantage is that it has taken the lead in running through the mass chain of consumer robots.
There is a general doubt in the market that when GPT-7, GPT-8 capabilities are further elevated, or even built-in VLA capabilities, they will simply eat the living space of robotic start-ups.
The initial judgement is that GPT, such as LLM, belongs to the public sector and is accessible to all.
What really opens the gap is not "can we use a powerful generic model" but do we have a physical set, a high-quality operational data set, a complete training closed loop.
Language LLM provides high-level awareness, but lacks "fuel" -- that is, physically interactive data tied to hardware -- and no strong planning capacity. Open AI itself has the same logic of layouting hardware.
The calculus is not the main bottleneck of the moment, and the kilocalories to the Vanka clusters are almost adequate at this stage. What is really scarce in the industry is a high-quality, high-efficiency data collection and post-training paradigm. Domestically, there is geographical advantage in the availability of data.
In contrast to overseas routes, the Zen route relies on real machine error and the task success rate is high, but it is difficult to scale up and extend; the UMI-type hand-held collections can magnify the size of sample and expand well, but success rate is difficult to fill naturally. It seeks to integrate the two routes and to ensure that the real machine success rate is accompanied by portable collection to solve the problem of scale.
The complete Physical AI system is divided into three tiers: S0 Movement Control, S1 Executive Level VLA, S2 High Level Awareness Planning.
The team has extensive LLM training experience, processing hundreds of thousands of hours of grade data. The ideal shape of the future is similar to the apple model: hardware terminals, data portals, modelling capabilities, and not simply output algorithm models.
The risks are equally clear: whether millions of hours of data are able to climb on schedule; the difficulty of training infrastructure; the probability of large-scale production of C-end hardware; whether household scenes can cross 95 per cent of the experience threshold; and whether the business model of Robot-Ready Home and real estate home can run.
Physical AI is a long chain, and any lag in any chain can drag down the whole rhythm.
For starters, Q1, T1, is just entry tickets, and the real war of physics AI takes place in a closed circle of hardware, data, models that bite each other.
Editor: Aeneas
Read more
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.