← all writing

The Data Wall: Distill Humans, Then Close the Loop

A robot is a few dozen numbers changing over time. Every arm, hand, wheel and leg boils down to two kinds of joint, one that rotates and one that slides, chained together by rigid links. The stream of numbers those joints produce is 0.004% of what the robot’s cameras take in.

So here’s my bet. Robotics isn’t a separate science waiting for its own Einstein. It’s a sub-problem of general intelligence, and the companies building general intelligence just walked through the door. I got some sharp pushback on the first version of this post, and the field moved fast in the meantime, so the bet is a bit more careful now. Five claims, each with numbers behind it:

  1. Action is just another modality, though not an easy one. General models bring the world knowledge. Turning it into reliable contact is still its own problem.
  2. Without enough data, exploring algorithms mostly burns time. In 2019, 13,000 years of simulated experience bought one Rubik’s cube at 20% success. In 2016, 3,000 robot-hours bought one skill: grasping. Past a certain data scale, algorithm differences shrink into the noise. The teams with a real shot are the patient ones building AI-native, end-to-end hardware.
  3. Human distillation is the on-ramp. The device and pipeline that turn human skill into training data. You can measure how efficient it is, and the gaps between today’s interfaces run from 3x to over 100x.
  4. The field is splitting in two. One camp scales capture toward 10 million hours on giant checks. The other hunts for the loop: one complete, paid deployment, even a small and slow one.
  5. The long-term moat is the loop. A system that keeps turning real-world experience into capability you can deliver. Capture is the entry, the model is the converter, and finished tasks are the only test.

Start with the object itself. Here’s a whole robot arm. Six numbers. Hit play.

一台机器人,就是几十个随时间变化的数字。 不管是机械臂、手、轮子还是腿,拆到底都只有两种关节:一种转,一种滑,中间用刚性连杆串起来。这些关节产生的数字流,只占机器人摄像头输入的 0.004%。

所以我的判断是:机器人不是一门等着自己的爱因斯坦出现的独立学科。它是通用智能的一个子问题,而做通用智能的那几家公司,刚刚推门进来了。这篇的第一版发出去后挨了几条很到位的批评,行业这几个月也跑得飞快,所以现在的判断比原来谨慎了一点。五个论点,每个后面都有数字撑着:

  1. 动作只是又一种模态,但不是一种简单的模态。 通用模型带来对世界的知识,可要把它变成可靠的接触控制,仍然是一个独立的问题。
  2. 数据不够时,探索算法大多是在空烧。 2019 年,一万三千年的仿真经验换来一个魔方,成功率 20%。2016 年,三千个机器人小时换来一项技能:抓取。数据过了某个量级,算法之间的差异只剩统计方差。真正有机会的,是沉下心打磨 AI 原生、端到端硬件的团队。
  3. 人类蒸馏是入口。 也就是把人的手艺变成训练数据的那套设备加流程。它的效率可以量化,现有各种接口之间差距从 3 倍到 100 倍以上。
  4. 这个领域正在分成两条路。 一边拿巨额资金把采集规模往一千万小时推;另一边在找闭环:一个完整的、有人付钱的部署,哪怕规模小、速度慢。
  5. 长期的护城河是闭环。 一套能持续把真实世界的经验转化成可交付能力的系统。采集是入口,模型是转化机制,真正完成的任务才是唯一的检验标准。

先从机器人本身说起。下面是一整条机械臂,一共六个数字。点一下播放。

A note on that arm, because it’s an example in its own right. Claude Opus 5.5 wrote the whole thing, zero-shot in the sense that matters. It got a one-paragraph description. No demonstrations, no motion capture, no robot data. The kinematics, the reach, the pour, the wipe: it worked all of them out on its own, and checked its own screenshots along the way. Now imagine it running 10x faster, and in 3D instead of 2D. I think models like this are fully capable of beating every embodied model that exists today.

关于上面这条机械臂,多说一句,因为它本身就是一个例子。它是 Claude Opus 5.5 写出来的,而且在最关键的意义上是零样本。 只给了它一段文字描述,没有示范,没有动作捕捉,也没有任何机器人数据。运动学、伸手抓取、倒水、擦桌子,全是它自己想出来的,过程中还会看自己的截图来检查。现在想象一下,它快 10 倍,而且从 2D 换成 3D。我觉得,这类模型完全有能力超过现存的任何一个具身模型。

Strip away the marketing and every robot is the same object: rigid links, connected by joints that either rotate or slide. A 7-DOF arm’s entire configuration is 7 numbers. A dexterous humanoid is about 50. The pose of the hand is just those joint values pushed through one rigid transform $A_i \in SE(3)$ per joint:

\[T_{\text{hand}}(q) = A_1(q_1)\,A_2(q_2)\cdots A_n(q_n)\]

Everything a robot does is a trajectory $q(t)$ through that space. Wiping, folding, pouring, plugging in a cable. We can code any trajectory we want. The hard part was never the joints. It’s choosing the trajectory. And that needs knowing what a cup is, how a towel drapes, where you want things, and what happens if you push too hard. That’s world understanding, and world understanding is general intelligence.

The bandwidth makes the point brutal. Fifty joints at 50 Hz is 2,500 numbers a second. Three cameras at 512 × 512 RGB and 30 fps is 70.8 million numbers a second. The action is 0.004% of the stream.

It’s also a smooth, low-dimensional waveform, which is exactly why Physical Intelligence’s FAST tokenizer squeezes it with a frequency transform (the DCT), the same family of trick JPEG and MP3 use. Play with it below. Six coefficients out of fifty and you can barely tell the difference. To a frontier lab that already models speech and video, the action head is a rounding error, at least in bandwidth.

1. 机器人就是一串关节加连杆

把营销话术剥掉,所有机器人都是同一种东西:刚性连杆,用要么转、要么滑的关节连起来。 一条 7 自由度机械臂的全部状态就是 7 个数。一个灵巧人形机器人大约 50 个。手的位姿,不过是把这些关节值,每个关节乘一个刚体变换 $A_i \in SE(3)$:

\[T_{\text{hand}}(q) = A_1(q_1)\,A_2(q_2)\cdots A_n(q_n)\]

机器人做的每一件事,都是这个空间里的一条轨迹 $q(t)$。擦桌子、叠衣服、倒水、插线,都是。任何轨迹我们都写得出来。难的从来不是关节,而是选哪条轨迹。 这需要知道杯子是什么、毛巾怎么垂下来、你想把东西放哪、用力过猛会怎样。这叫对世界的理解,而对世界的理解就是通用智能。

带宽一对比,这一点就很残酷。50 个关节、50 Hz,是每秒 2,500 个数。三个摄像头、512 × 512 RGB、30 帧,是每秒 7,080 万个数。动作只占整个数据流的 0.004%。

而且它是一条平滑的低维波形,这正是 Physical Intelligence 的 FAST 分词器能用频域变换(DCT)把它压扁的原因,跟 JPEG、MP3 是同一家族的把戏。下面自己拖着玩:五十个系数里留六个,几乎看不出区别。对一个已经在建模语音和视频的前沿实验室来说,动作头只是个零头,至少在带宽上是。

Now the caveat I owe you, because it was the first thing people pushed back on, and they’re right: small output doesn’t mean easy to learn. A chess move is one square, and chess is still hard. Folding a shirt may come down to a few numbers per control step, but to pick them you have to read occlusion, friction, tension and contact state off pixels, and guess what the cloth does after you move. The 0.004% doesn’t make the action model a trivial decoder.

So the honest version is this. General models bring strong priors for knowledge and perception. Whether those priors become reliable physical control is still its own technical problem. The action is cheap in bandwidth, not in difficulty, and that’s exactly why the data matters so much.

这里我欠你一个限定条件。这是第一版最先被人挑出来的地方,而且挑得对:输出维度小,不等于学起来容易。 下棋只输出一个落子位置,难度照样很大。叠一件衣服,每个控制步也许只有几个数,但要选出这几个数,你得从像素里判断遮挡、摩擦、张力、接触状态,还得预判动作之后布料会怎么变。不能因为动作只占视觉带宽的 0.004%,就认定动作模型是个微不足道的解码器。

所以更老实的说法是:通用模型能提供很强的知识和感知先验,但这些先验能不能变成可靠的物理控制,仍然是一个独立的技术问题。动作在带宽上便宜,在难度上不便宜。这也正是数据为什么这么要紧。

2. General intelligence eats robotics

If the action is cheap in bandwidth, then the expensive part of the robot brain is the world model. And the world model is the thing frontier labs already train on tens of thousands of GPUs and the biggest video corpora on earth. NVIDIA’s Cosmos alone was trained on 20 million hours of video (NVIDIA). In 2026 every major lab stopped watching and moved in:

Lab 2026 move Source
OpenAI Robotics division launched May 31, led by Aditya Ramesh; Altman in September: “we will definitely do a humanoid”; hiring for data-acquisition stations and managed operator workforces Forbes, The Rundown
Anthropic Held acquisition talks with Physical Intelligence in spring, per The Information; no deal closed, and PI’s CEO denied the later rumor Yahoo / TechCrunch
Google DeepMind Shipped Gemini Robotics 2 on Apptronik, Boston Dynamics and Franka hardware (July 30) Forbes
DeepSeek Bought ~\$21M of Unitree shares in its Shanghai IPO (August); agreed to lead a seed round valuing robot-AI lab Winverse at ~\$400M (October) Softonic, Dealroom

Read OpenAI’s job posts and the strategy is plain. They’re hiring for collection facilities, operators, rigs, throughput and data quality, not just model researchers. OpenAI wound down its first robotics team in 2021, citing a lack of data. The second attempt starts by building a human data factory. The labs have reached the same conclusion I have: data first.

2. 通用智能吃掉机器人

既然动作在带宽上只是个零头,那机器人大脑里最贵的部分就是世界模型。 而世界模型,正是前沿实验室已经在用几万张 GPU 和地球上最大的视频语料训练的东西。光 NVIDIA 的 Cosmos 就用了 2,000 万小时视频训练(NVIDIA)。2026 年,每一家大实验室都不再观望,直接下场了:

实验室 2026 年的动作 来源
OpenAI 5 月 31 日成立机器人部门,由 Aditya Ramesh 领导;Altman 9 月说「我们肯定会做人形机器人」;在招数据采集站和成建制的操作员团队 Forbes,The Rundown
Anthropic 据 The Information 报道,春天曾与 Physical Intelligence 谈过收购;没有成交,PI 的 CEO 也否认了之后的传闻 Yahoo / TechCrunch
Google DeepMind 7 月 30 日在 Apptronik、Boston Dynamics 和 Franka 的硬件上发布 Gemini Robotics 2 Forbes
DeepSeek 8 月宇树在上海 IPO 时认购了约 2,100 万美元的股份;10 月同意领投机器人 AI 实验室 Winverse 的种子轮,估值约 4 亿美元 Softonic,Dealroom

看看 OpenAI 的招聘启事,战略一目了然。他们招的是采集设施、操作员、采集设备、吞吐量和数据质量,而不只是模型研究员。OpenAI 2021 年解散了第一支机器人团队,给出的理由就是数据不够。第二次尝试,第一步是建一座人类数据工厂。 这些实验室得出的结论跟我一样:数据先行。

3. Without enough data, exploration mostly burns time

Every year some new lab raises money on the same pitch: our new algorithm will let robots learn on their own. Here’s how I think about that. If you’ve trained RL agents, you know what the start of a run looks like: the policy’s entropy is high, the actions are close to random, and almost every rollout is wasted. A company exploring new algorithms before it has enough data is in that high-entropy phase, as a company. Most of the motion goes nowhere. And the field has already watched this play out, at scale, with the best teams in the world:

Year Experiment What it cost What it bought Source
2021 OpenAI closes its robotics team Five years of work A verdict: not enough data Forbes
2019 OpenAI one-handed Rubik’s cube 13,000 years of simulated experience; 64 V100 GPUs plus 920 machines of 32 CPU cores, for months One task, 20% full-solve success on the hardest scrambles VentureBeat
2017 DeepMind locomotion from scratch Agents learned from simple rewards like “move forward without falling” Robust movement that often didn’t look human; follow-up work imitated human motion capture to get natural skills DeepMind
2016 Google “arm farm” 800,000 grasp attempts, about 3,000 robot-hours on 6–14 arms over months One skill: grasping IEEE Spectrum

Put the Rubik’s cube next to today’s best human-distilled model. 13,000 years is 114 million hours of experience, for one task, at 20%. Generalist’s GEN-1 was pretrained on 500,000 hours of human interaction across hundreds of tasks, and it adapts to a new robot and task with about one hour of robot data (Generalist). That’s 228x less experience, for a model that generalizes instead of memorizing one puzzle.

Why exploring from zero loses: the space is too big. Take a 50-joint body, a 10-second task at 50 Hz, and only 10 possible values per joint per step. The number of possible trajectories is

\[10^{\,50 \times 500} = 10^{25{,}000}\]

There are about $10^{80}$ atoms in the observable universe. You could try random trajectories until the universe ends and not make a dent. A human demonstration doesn’t search that space at all. It points at the thin slice that works. Loosely, human distillation is importance sampling, with a billion years of evolution as the proposal distribution. It’s an analogy, not a theorem, but the intuition holds: a demo hands the model a piece of search that a human already did.

Watch it happen in a toy. Red is random exploration trying to get a hand through a slot to the cup. Teal is a policy fit to five human demos. Then crank the sliders on the calculator and watch the exponent explode.

3. 数据不够时的探索,大多是在空烧

每年都有新实验室拿同一套说辞去融资:我们的新算法能让机器人自己学会。 我是这么看的。训过强化学习的人都知道一次训练刚开始是什么样子:策略的熵很高,动作跟随机差不多,绝大部分 rollout 都是白跑。一家在数据还不够的时候就去探索新算法的公司,整家公司都处在这个高熵阶段。 大部分折腾都不会有结果。这件事,整个领域已经用全世界最好的团队、按最大规模演示过一遍了:

年份 实验 花了什么 换来什么 来源
2021 OpenAI 关闭机器人团队 五年的工作 一个结论:数据不够 Forbes
2019 OpenAI 单手魔方 一万三千年的仿真经验;64 张 V100 加 920 台 32 核 CPU 机器,跑了几个月 一个任务,最难的打乱下完整复原成功率 20% VentureBeat
2017 DeepMind 从零学走路 只给「往前走、别摔倒」这种简单奖励 动作很稳,但经常不像人;后续工作改为模仿人类动作捕捉,才学出自然的动作 DeepMind
2016 Google「机械臂农场」 80 万次抓取尝试,6–14 条机械臂跑了几个月,约 3,000 个机器人小时 一项技能:抓取 IEEE Spectrum

把魔方跟今天最好的人类蒸馏模型摆在一起看。一万三千年是 1.14 亿小时的经验,换一个任务、20% 的成功率。Generalist 的 GEN-1 用 50 万小时、横跨几百种任务的人类交互数据做预训练,换到一台新机器人、一个新任务上,只要大约一小时的机器人数据就能适配(Generalist)。经验少了 228 倍,而且得到的是一个会泛化的模型,不是一个只背下了一道题的模型。

为什么从零探索会输:空间太大了。 取一个 50 关节的身体,一个 10 秒、50 Hz 的任务,每个关节每一步只允许 10 种取值。可能的轨迹数是

\[10^{\,50 \times 500} = 10^{25{,}000}\]

可观测宇宙里大约有 $10^{80}$ 个原子。在这么大的空间里随机试,试到宇宙终结也试不出个所以然。人类示范根本不去搜,它直接指给你能成功的那一小片。打个比方,人类蒸馏就像重要性采样,提议分布是十亿年的进化。 这只是类比,不是严格的数学结论,但直觉没错:一条示范,等于把人已经替你搜过的那部分问题直接交给了模型。

下面用一个玩具看看。红色是随机探索,想让手穿过一条缝够到杯子;青绿色是拟合了五条人类示范的策略。然后去拖下面计算器的滑块,看指数怎么爆炸。

The AlphaGo Zero exception proves the rule. Also in 2017, DeepMind showed that pure self-play beat human data at Go. It worked because Go meets three conditions: the simulator is the real game, experience is free and infinitely parallel, and the reward (win or lose) is exact. The physical world fails all three. Simulators drift from reality, real experience costs money and breaks things, and “tidy the kitchen” has no clean reward.

Now the part that comes from my own work training on very large datasets. Once data passes a certain scale, the differences between algorithms mostly shrink into the statistical variance. Swap the architecture, change the loss, add a clever trick: the curves end up inside each other’s error bars. Below that scale, every trick looks decisive. Above it, the data decides. Most companies betting on an algorithm will run out of runway before their data ever gets there.

That doesn’t mean nobody else has a chance. A robot isn’t just a model. It’s hardware and software together: engineering, productization, and a long grind of polishing in the real world. If anything, I think the teams that win from here are the patient ones. They admit they won’t disrupt the industry overnight, and they spend the time building AI-native hardware. By that I mean fully end-to-end hardware: sensors that hand the model raw data, actuators that take the model’s output directly, and no hand-written layer in between.

And here’s my hot take. The only mistake we’ll make is not trusting end-to-end enough. Give the model the raw data. Hand it the actions completely. Drop the human-designed algorithms in the middle. It sounds riskier, but it actually makes the problem simpler: there’s one thing to improve instead of a stack of hand-tuned modules fighting each other. Our only job is to give the AI a feedback loop. That’s also where RL finally earns its keep: not exploring from zero, but learning from real outcomes on top of a strong prior. Hold on to that loop. It comes back in section 6.

AlphaGo Zero 这个例外,恰好证明了规律。 同样是 2017 年,DeepMind 证明了在围棋上纯自我对弈能打败人类数据。它之所以行,是因为围棋满足三个条件:模拟器就是真实的比赛;经验免费而且可以无限并行;奖励(赢或输)是精确的。物理世界三条全不满足。模拟器跟现实有偏差,真实经验要花钱、会弄坏东西,而「把厨房收拾好」没有一个干净的奖励。

下面这段来自我自己在海量数据上训练的经验。数据量一旦过了某个量级,不同算法之间的差异,基本只会体现在统计方差里。 换个架构、改个损失函数、加个巧妙的技巧,最后的曲线都落在彼此的误差范围之内。数据没到那个量级的时候,每个技巧看起来都像是决定性的;过了那个量级,说了算的是数据。大多数押注算法的公司,会在自己的数据达到那个量级之前就倒下。

但这不等于别人就没有机会了。机器人不只是一个模型,它是软硬件结合的工程化、产品化,加上在真实世界里一点点打磨的过程。我反而觉得,接下来能赢的是那些沉得下心的团队:承认自己不可能一天之内颠覆行业,把时间花在打磨 AI 原生的硬件上。我说的是完全端到端的硬件:传感器把原始数据直接交给模型,执行器直接执行模型的输出,中间没有人写的那一层。

我还有一个暴论:我们唯一会犯的错误,就是不够相信端到端。 把原始数据丢给模型,把输出的动作完全交给 AI,放弃中间那些人写的算法。听起来更冒险,其实反而让问题变简单了:要改进的只剩一个东西,而不是一堆互相打架、各自手调的模块。我们唯一需要做的,就是给 AI 提供反馈回路。 这也是强化学习真正派上用场的地方:不是从零开始瞎探索,而是在强先验之上,从真实的结果里学。记住这个回路,第 6 节还会回来。

4. The on-ramp: the human distillation interface

If models commoditize and from-scratch exploration is a furnace, the first variable that decides who gets anywhere is this: how efficiently can you turn human skill into robot-usable data? Call the device and pipeline that does this the distillation interface: a teleop rig, a handheld gripper, a sensor glove, a head-mounted camera. Its efficiency is

\[\eta = \frac{r \cdot f \cdot d}{c}\]

where $r$ is capture speed relative to a bare human hand, $f$ is fidelity (how much of the signal a robot can actually use: hand pose, contact, force), $d$ is diversity (distinct homes, objects and tasks per thousand hours), and $c$ is cost per captured hour.

The speed term has actually been measured. In the UMI paper, people arranging cups produced about 231 demonstrations in 15 minutes with bare hands, 111 with the UMI handheld gripper (about \$371 including its GoPro), and 35 by teleoperation (UMI). On a dynamic tossing task, teleop produced zero successful demos.

Interface Speed vs bare hand Hardware Cost per human-speed hour Hours a \$250M budget buys
Teleoperation 15% ~\$20,000 rig ~\$198 1.3M
UMI handheld gripper 48% \$371 ~\$52 4.8M
Bare hands on camera 100% a camera ~\$25 10M, but no action labels

The cost column assumes a \$25-an-hour operator and hardware amortized over 4,000 hours. Per \$1,000 of hardware, UMI yields about 171x more demonstrations per hour than teleop. Feed the 3.8x data advantage through the scaling law in section 5 (error ∝ D−0.5) and the same check buys about 2x lower error. No algorithm paper delivers 2x with that kind of certainty.

Drag the budget around. The glove’s speed is the one number nobody has published, so it’s a slider. Set it to whatever you believe.

4. 入口:人类蒸馏接口

如果模型会变成大路货、从零探索又是一座熔炉,那决定谁能起步的第一个变量就是:你能多高效地把人的手艺变成机器人能用的数据? 我把干这件事的设备加流程叫作蒸馏接口:遥操作台、手持夹爪、传感手套、头戴摄像头,都算。它的效率是

\[\eta = \frac{r \cdot f \cdot d}{c}\]

其中 $r$ 是相对于人徒手的采集速度,$f$ 是保真度(机器人真正用得上的信号有多少:手的位姿、接触、力),$d$ 是多样性(每一千小时里有多少不同的家庭、物体和任务),$c$ 是每采集一小时的成本。

速度这一项真有人测过。在 UMI 论文里,摆杯子的任务,15 分钟内徒手大约能做 231 条示范,用 UMI 手持夹爪(连 GoPro 在内约 371 美元)能做 111 条,遥操作只有 35 条(UMI)。换成一个动态抛物的任务,遥操作一条成功的示范都没做出来。

接口 相对徒手的速度 硬件 每「人类速度小时」成本 2.5 亿美元能买多少小时
遥操作 15% 约 2 万美元的设备 约 198 美元 130 万
UMI 手持夹爪 48% 371 美元 约 52 美元 480 万
徒手 + 摄像头 100% 一个摄像头 约 25 美元 1,000 万,但没有动作标签

成本这一列假设操作员每小时 25 美元,硬件按 4,000 小时摊销。按每 1,000 美元硬件算,UMI 每小时产出的示范是遥操作的约 171 倍。把 3.8 倍的数据优势代进第 5 节的缩放律(误差 ∝ D−0.5),同一张支票能把误差压低大约一半。 没有哪篇算法论文能这么确定地给你 2 倍。

拖一下预算试试。手套的速度是唯一一个没人公布过的数,所以我把它做成了滑块,你信多少就设多少。

The industry has already voted with its capex:

  • Sunday trains on about 10M episodes from 500+ homes, captured with a \$200 skill-capture glove. Its ACT-1 model uses “zero robot data” (eWeek).
  • Generalist pretrains GEN-1 on 500,000+ hours with no robot data, captured with low-cost wearables. Back at GEN-0 it was already growing by about 10,000 hours a week (Generalist).
  • Tesla moved Optimus training in 2025 from motion-capture suits and teleop to workers wearing five-camera helmet-and-backpack rigs (eWeek).
  • OpenAI is hiring to run data-acquisition stations and operator workforces at multi-site scale.

The critics’ best objection is about the $f$ term. Video alone misses the physical interaction, the push and resistance that teleop captures (eWeek). Fair. That’s exactly why the race is in the top-right corner of that scatter: wearables that record hand pose, contact and eventually force, at human speed, in real homes. Whoever pushes η furthest owns the scarcest input in robotics.

Here’s what my first draft underrated, though. The most valuable part of a human demo usually isn’t the clean, successful trajectory. It’s the judgment around it:

  • Why did they switch grasp points right there?
  • How did they notice the shirt wasn’t fully spread out?
  • Once they saw the mistake, how did they fix it?
  • When did they stop, step back, and look again?

That’s data about judging and recovering, and it’s much closer to deployable skill than a run of perfect motions. Every capture route gets some of it and misses some:

Capture route What it’s good at What it still has to solve
Robot teleoperation Actions map straight onto the robot Speed, cost, scene coverage
UMI-style handheld Scales well, carries action information The tool constrains natural behavior; transfer error
Bare hands and wearables Natural, rich scenes Action labels, contact information, whether a robot can actually execute it

So I still like η for comparing interfaces. It makes clear that a cheap rig, a giant hour count or precise labels don’t prove anything on their own. But it’s not the number I’d optimize. That one is:

For every dollar spent, how much does task success go up in real deployment?

Getting the data and the model learning from it are a long way apart. Ten thousand hours of the same easy fold and ten thousand hours that cover the failure modes that actually bite can be worth completely different amounts.

整个行业已经用资本开支投了票:

  • Sunday 用一只 200 美元的「技能采集手套」,在 500 多个家庭里采了大约 1,000 万条片段来训练。它的 ACT-1 模型用的是「零机器人数据」(eWeek)。
  • Generalist 的 GEN-1 用 50 万小时以上、不含任何机器人数据的人类数据预训练,靠的是低成本的可穿戴设备。早在 GEN-0 的时候,它就以每周大约 1 万小时的速度在增长(Generalist)。
  • Tesla 2025 年把 Optimus 的训练从动捕服和遥操作,换成了让工人戴着五摄像头的头盔加背包(eWeek)。
  • OpenAI 在招人运营多地点的数据采集站和操作员团队。

批评者最有力的反驳落在 $f$ 这一项:光有视频,就漏掉了物理交互,漏掉了遥操作能记下的推力和阻力(eWeek)。说得对。这也正是为什么真正的竞争在那张散点图的右上角:在真实家庭里、以人的速度,记录手的位姿、接触,最终还有力的可穿戴设备。 谁把 η 推得最远,谁就握住了机器人领域最稀缺的那种输入。

不过,有一点是我第一版低估了的。一条人类示范里最值钱的,往往不是那条干净漂亮的成功轨迹,而是它周围的判断:

  • 人为什么在这里换了抓点?
  • 是怎么发现衣服没有完全展开的?
  • 发现错误以后,怎么纠正的?
  • 什么时候停下当前动作,退一步重新观察?

这些数据包含的是「如何判断、如何恢复」,比单纯记录一串完美动作更接近可部署的能力。每一种采集路线都能拿到一部分,也都会漏掉一部分:

采集路线 主要优势 还需要解决的问题
机器人遥操作 动作直接对应机器人 速度、成本、场景覆盖
UMI 类手持设备 易扩展,带动作信息 工具对自然行为的约束、迁移误差
裸手与可穿戴采集 自然、场景丰富 动作标签、接触信息、机器人能不能执行

所以我仍然喜欢用 η 来比较接口:它说明便宜的设备、海量的小时数、高精度的标签,任何一个单独拿出来都证明不了优势。但它不是我真正要优化的那个数。真正要优化的是:

每投入一块钱,真实部署中的任务成功率能提高多少?

「采到了」和「模型学会了」之间,还隔着很长一段路。同样一万小时,重复的简单折叠,和覆盖了真正会出事的失败模式的数据,价值可能完全不同。

5. What the data wall costs

Why does the interface matter this much? Because the bill is enormous. Two numbers frame it: robotics has about 0.5M hours of real manipulation data, and Waymo alone has 11–18M hours of fully driverless driving (271.3M rider-only miles through June 2026, at 15–25 mph; Waymo). Driving outputs two numbers, steering and acceleration. A dexterous body outputs 40–50, and every one of them makes and breaks contact. And driving still isn’t considered solved.

Compute. Training cost is roughly $C \approx 6ND$, where $N$ is parameters and $D$ is hours × tokens per hour. A lean policy model sees about 3.5M tokens per hour of three-camera video. A rich world model at 512 px and 30 fps sees about 83M. On 20,000 B300s, one epoch over 10M hours of rich video takes about a month. At 100M hours it takes about eleven months and \$0.8–1.4B. And if model size grows with data, compute grows with the square of data. Try it:

5. 数据墙要花多少钱

接口为什么这么要紧?因为账单大得吓人。两个数字先定个框:机器人领域大约只有 50 万小时的真实操作数据,而光 Waymo 一家就有 1,100 万到 1,800 万小时的全无人驾驶(截至 2026 年 6 月共 2.713 亿英里纯乘客里程,按 15–25 英里/小时算;Waymo)。开车只输出两个数:方向和油门。一个灵巧的身体要输出 40 到 50 个,每一个都在不停地接触、松开。而开车到今天都还不算解决了。

算力。 训练成本大约是 $C \approx 6ND$,$N$ 是参数量,$D$ 是小时数乘以每小时的 token 数。一个精简的策略模型,每小时三路摄像头视频大约是 350 万个 token。一个 512 像素、30 帧的丰富世界模型大约是 8,300 万。用 2 万张 B300,在 1,000 万小时的丰富视频上跑一个 epoch 大约要一个月。到 1 亿小时,大约要十一个月,花 8 到 14 亿美元。而如果模型大小跟着数据一起涨,算力就按数据的平方涨。自己试试:

Reliability. A task with $n$ steps and per-step success $p$ succeeds with probability $p^n$. Sunday’s table-to-dishwasher demo had 33 distinct interactions (68 in total): 72% end to end at 99% per step, 97% at 99.9%. A full “clean the kitchen” is hundreds of interactions. Here’s what that feels like when you actually run the episodes:

可靠性。 一个有 $n$ 步、每步成功率为 $p$ 的任务,整体成功的概率是 $p^n$。Sunday 那个从餐桌到洗碗机的演示有 33 种不同的交互(总共 68 次):每步 99% 时,端到端成功率 72%;每步 99.9% 时是 97%。完整地「收拾厨房」要几百次交互。真把一局一局跑出来,是这个感觉:

The scaling law turns that into a data bill. Generalist’s GEN-0 post shows loss falling as a power law in pretraining data, and one secondhand report puts the exponent at about 0.5, i.e. $L \propto D^{-0.5}$ (AIbase). I haven’t seen Generalist state that number itself, so treat it as a working guess. Then the big assumption, stated out loud: if per-step failure rate falls the same way, each 10x cut in per-step error costs 100x more data:

Per-step reliability Data needed (from today’s 0.5M hours) Rich-model epoch cost at 10B params
99% ~0.5M hours ~\$4–7M
99.9% ~50M hours ~\$0.4–0.7B
99.99% ~5B hours ~\$38–70B

Treat that table as a scenario, not an invoice. Training loss, per-step failure rate and long-task success are three different quantities, and linking them takes assumptions I can’t prove. Even so, I think the shape survives being off by an order of magnitude: each extra nine costs a lot more experience than the last one, and compute grows faster than data. A pipeline that’s 4x more efficient isn’t a 4x edge. It moves the finish line by years. And the cheapest way to buy the next nine is to aim the data at where the robot actually fails. Which brings us to the split.

缩放律把这件事变成了一张数据账单。 Generalist 在 GEN-0 的文章里给出的是损失随预训练数据量按幂律下降,而一篇二手报道说这个指数大约是 0.5,也就是 $L \propto D^{-0.5}$(AIbase)。我没看到 Generalist 自己公布这个数,所以只当它是个工作假设。然后还有一个大假设,我明说出来:如果每一步的失败率也按同样的规律下降,那每步错误率每降 10 倍,就要多 100 倍的数据:

每步可靠性 需要的数据(从今天的 50 万小时算起) 100 亿参数丰富模型跑一个 epoch 的成本
99% 约 50 万小时 约 400–700 万美元
99.9% 约 5,000 万小时 约 4–7 亿美元
99.99% 约 50 亿小时 约 380–700 亿美元

这张表请当成假设情景看,不是确定的账单。训练损失、单步失败率、长任务成功率是三个不同的量,把它们连起来需要我证明不了的假设。但我觉得,就算差一个数量级,形状也不会变:每多一个 9,要的经验都比上一个多得多,而算力涨得比数据还快。一个效率高 4 倍的管线,不是 4 倍的优势,它能把终点线往前挪好几年。 而买下一个 9 最便宜的办法,是把数据对准机器人真正失败的地方。这就引出了行业的分岔。

6. The field splits: scale versus the loop

The last few months moved fast, and the progress is solid.

  • Generalist shipped GEN-1.5 in August, still on 500,000+ hours of human interaction. Its trick is “physical prompting”: show it one 3–12 second demo and it picks up a new task with no retraining, at 59% average success across ten tasks, or 83% after ten gradient steps on five minutes of data (Generalist).
  • Dyna pretrained DYNA-2 on over a million hours of egocentric human video, no robot actions, and rebuilt its data pipeline from 14,000 to 440,000 episode-hours a week to digest it (Dyna). Then Dyna-2.1 put that policy inside an agent on a new wheeled robot and ran an uncut hour of commercial laundry (wash, dry, fold, shelve) with no human stepping in (Dyna).
  • PI opened its API in October around a loop: fine-tune on your data, test on the robot, let the failures decide what to collect next (PI). Its research line, π*0.6 and the online RL work after it, learns from deployment rollouts and human corrections.

Put next to each other, these sort the field into two groups.

Path 1: scale. Generalist, plus OpenAI, Anthropic and the other LLM giants. They have, or can raise, the money to push human capture toward 10 million hours (Generalist alone raised \$400M in June and was reportedly valued at \$3B by August) and bet that a big enough prior turns into reliability on its own. That’s the Waymo-scale playbook applied to hands, and it’s a legitimate bet.

Path 2: the loop. PI (which could switch to path 1 any time it wants), Dyna, Sunday. They’re hunting for a loop that works in deployment and pays for itself: one complete application, even if it’s small-scale, in limited scenes, slow, and the success rate isn’t great yet. Dyna’s robots already fold napkins and towels for restaurants and hotels. They’re all charging toward performance-as-a-service: the customer pays for finished work, not for a robot. Nobody publishes per-task pricing yet; Dyna reportedly charges a monthly fee per robot. But that’s the direction. (The lines blur, too. Dyna also talks about a path to 10 million hours. The difference is what each company treats as the finish line.)

I think path 2 is reading the customer right. Customers don’t need more dancing robots, or agents that sing and paint. They already have PineDrama for entertainment. What they need is a robot that runs on its own, robustly, and finishes every task it’s assigned. No human stepping in. Nobody feeding it supplies. Nobody cleaning up its mess. It gets the job done, then it’s ready for the next one, and the one after that, so smoothly you barely notice it’s there. That’s what I mean by ultimate autonomy. It’s the same thing we want from the personal agents that already run errands for us in the digital world (I’m a big fan of Instinct), except with hands.

The loop is how you get there:

deploy → log where it fails → capture exactly the missing experience, recovery included → train → verify the fix → back into the field.

Say you’re folding laundry in 10,000 homes, 100,000 garments. The homes give you a chance to cover the distribution. They don’t give you a moat on their own. The moat is the routing: which garments in which states it fails on → targeted capture → train → verify → back to the homes. That also changes how you capture human data in the first place. Keep the task context, the state before and after, and every failure and recovery, so the human data lines up with what the robots report back.

Here’s a toy version. Both strategies get the same capture budget every week. Blue spends it the usual way, on whatever is cheap to collect. Teal spends it on last week’s failures. Watch how many whole loads finish with nobody touching them.

6. 行业分岔:拼规模,还是拼闭环

这几个月行业跑得很快,而且进展都很扎实。

  • Generalist 8 月发布了 GEN-1.5,底子仍然是 50 万小时以上的人类交互数据。它的绝活叫「物理提示」:给它看一段 3 到 12 秒的示范,不用重新训练就能学会一个新任务,十个任务平均成功率 59%;在五分钟的数据上走十步梯度,能升到 83%(Generalist)。
  • Dyna 用一百多万小时第一视角的人类视频预训练了 DYNA-2,不含任何机器人动作;为了消化这些数据,它把数据管线的处理能力从每周 1.4 万个片段小时提到了 44 万(Dyna)。接着 Dyna-2.1 把这个策略装进一个智能体,放到一台新的轮式机器人上,在商用洗衣房里一镜到底跑了一个小时(洗、烘、叠、上架),全程没人插手(Dyna)。
  • PI 10 月开放的 API 就是围着闭环设计的:用你的数据微调,上机器人测试,让失败来决定下一步采什么(PI)。它的研究线 π*0.6 以及之后的在线强化学习工作,都是从部署时的执行记录和人的纠正里学习。

把这些放在一起看,行业分成了两拨。

第一条路:拼规模。 Generalist,加上 OpenAI、Anthropic 和其他大模型巨头。它们有钱(或者融得到钱),能把人类采集往一千万小时推(光 Generalist 6 月就融了 4 亿美元,据报道 8 月估值到了 30 亿美元),赌的是先验足够大,可靠性自己就会长出来。这是把 Waymo 级别的打法搬到了手上,是一个正当的赌注。

第二条路:拼闭环。 PI(它想走第一条路随时可以走)、Dyna、Sunday。它们在找一个能在部署中跑通、能自己养活自己的闭环:一个完整的应用,哪怕规模还小、场景有限、速度慢、成功率也不高。Dyna 的机器人已经在给餐厅和酒店叠餐巾、叠毛巾。它们都在往按效果付费冲:客户买的是干完的活,不是一台机器人。现在还没人公布按任务计价的价格,据报道 Dyna 是按台按月收费。但方向就是这个。(两条路之间的线也不是那么清楚,Dyna 自己也提了通往一千万小时的路径。区别在于,各家把什么当成终点线。)

我觉得第二条路看客户看得更准。客户不需要更多会跳舞的机器人,也不需要会唱歌画画的智能体,娱乐他们已经有 PineDrama 了。他们要的是一台能自己跑、足够鲁棒、交给它的每个任务都能干完的机器人。不用人插手,不用人给它续料,不用人给它收拾烂摊子。活干完,马上准备好下一个,再下一个,顺滑到你几乎察觉不到它的存在。这就是我说的终极自主。它跟我们对数字世界里那些已经在替我们跑腿的个人智能体的期待是同一回事(我是 Instinct 的忠实粉丝),只不过多了一双手。

闭环就是通往那里的路:

部署 → 记下它在哪里失败 → 精确地采集缺失的经验,包括怎么恢复 → 训练 → 验证修好了没有 → 回到现场。

假设你在一万户家庭里叠十万件衣服。家庭数量给你的是覆盖分布的机会,不会自动变成护城河。真正的护城河是这条回路:机器人在哪些衣物、哪些状态上失败 → 定向采集缺失的经验 → 训练 → 验证改进 → 回到家庭继续用。这也会反过来改变你一开始怎么采人类数据:尽量保留任务上下文、操作前后的状态、失败和恢复的过程,这样人类数据才能和机器人回传的部署反馈对上号。

下面是一个玩具版本。两种策略每周拿到的采集预算一样。蓝色按老办法花,什么便宜采什么;青绿色专门花在上周的失败上。看看有多少整筐衣服是从头到尾没人碰就叠完的。

So where does that leave everyone else?

Dead on arrival: labs whose only asset is an algorithm, with not enough data behind it. Ideas leak in months, and past the data threshold from section 3 the edge shrinks into the noise anyway. Their best exit is an acqui-hire.

Squeezed: model-only “robot brain” companies. They sell what the giants will give away. And a general robot API fails the revenue math on the robots that actually exist today:

Step Value
Humanoids shipped, first half of 2026 (Forbes) ~19,100
Annualized ~40,000 a year
Revenue if every new humanoid paid \$2,000 a year to one API ~\$80M a year
Revenue needed for an \$11B valuation at 10x \$1.1B a year
Paying robots needed 550,000, about 14 years of today’s shipments

PI clearly sees this. The Pi API, in beta since October 13, is a fine-tuning service, not a chat-style endpoint: upload your robot data, fine-tune on PI’s compute, run hosted inference. Smart, because it puts PI inside other people’s loops and pulls their deployment data back toward PI. It’s also services-shaped, and PI has reportedly acknowledged network-connectivity problems in some deployments, which is exactly the kind of thing ultimate autonomy can’t tolerate.

Path 2 in the home: Sunday. A \$200 glove instead of a ~\$20,000 teleop rig, worn in 500+ real homes, producing about 10M episodes. Sunday raised \$165M at a \$1.15B valuation and plans to put 50 Memo beta units into homes by Thanksgiving (SiliconANGLE). Its product sense about which chores people will pay to escape is the one thing a lab won’t build fast.

Path 2 at work: Dyna. Paying commercial customers, laundry and napkins at restaurants, hotels and laundromats, plus a million hours of human video underneath. It’s the clearest example of a company treating finished work as the product.

Path 1 at industrial scale: Generalist. From 270,000 hours (November 2025) to 500,000+ (April 2026): about 11,000 hours a week of real interaction, valued at a reported \$3B by August. A lab can copy the GEN architecture in a quarter. It can’t copy years of operational know-how about running human capture at that rate.

Path 1 makes the input the giants can’t print. Path 2 makes the thing nobody can fake: robots that finish the job. The giants will court both.

那其他人怎么办?

出生即死亡:唯一资产是一个算法、背后又没有足够数据的实验室。 想法几个月就会泄露,而且按第 3 节说的,数据一过那个量级,算法的优势本来也会缩进统计噪声里。最好的结局是被「收购式招聘」。

被挤压:只做模型的「机器人大脑」公司。 它们卖的东西,巨头会免费送。而且一个通用机器人 API,在今天真实存在的机器人数量上,收入账算不过来:

步骤 数值
2026 年上半年人形机器人出货量(Forbes) 约 19,100 台
折成全年 每年约 4 万台
假如每一台新人形机器人每年都给同一个 API 付 2,000 美元 每年约 8,000 万美元
按 10 倍收入撑起 110 亿美元估值,需要的收入 每年 11 亿美元
需要的付费机器人数 55 万台,约等于按今天的速度出货 14 年

PI 显然看明白了这一点。10 月 13 日开始测试的 Pi API 是一个微调服务,而不是聊天式的接口:上传你的机器人数据,用 PI 的算力微调,再用它托管的推理。这一步很聪明,因为它让 PI 钻进了别人的闭环,把别人的部署数据往 PI 这边拉。但它本质上是服务生意,据报道 PI 也承认有些部署出现了网络连接问题,而这恰恰是终极自主最容不下的东西。

第二条路,在家里:Sunday。 用 200 美元的手套代替约 2 万美元的遥操作设备,戴进 500 多个真实家庭,产出约 1,000 万条片段。Sunday 以 11.5 亿美元估值融了 1.65 亿美元,计划在今年感恩节前把 50 台 Memo 测试机送进家庭(SiliconANGLE)。它对「人们愿意花钱摆脱哪些家务」的产品直觉,是大实验室短时间内做不出来的东西。

第二条路,在工作场所:Dyna。 有付费的商业客户,在餐厅、酒店和自助洗衣店里叠毛巾、叠餐巾,底下还垫着一百万小时的人类视频。它是「把干完的活当产品」最清楚的例子。

第一条路,工业规模:Generalist。 从 27 万小时(2025 年 11 月)涨到 50 万小时以上(2026 年 4 月):每周约 1.1 万小时的真实交互,据报道到 8 月估值 30 亿美元。一个大实验室一个季度就能抄走 GEN 的架构,却抄不走这些年以这种速度跑人类采集积累下的运营经验。

第一条路生产的,是巨头印不出来的那种输入。第二条路生产的,是谁也装不出来的东西:真把活干完的机器人。巨头会两边都追。

7. Distill humanity first, then augment it

Phase 1: human distillation. Robots learn by compressing human physical behavior, and the interface sets the speed. The data that matters most is deep. Not factory loops of the same pick, but wild, messy, in-home trial and error: the over-grip, the dropped sock, the wrong drawer, and the recovery. Exploration still happens. It’s just that humans do the exploring, at human speed, with human judgment about what’s worth trying. That’s the cheapest exploration on earth, and it’s the opposite of burning GPUs on random actions.

The gate: data becomes a byproduct of revenue. A hand-built Memo costs about \$20,000 today (Sunday). Spread over three years at eight working hours a day, that’s about \$2.30 per robot-hour of capital, against \$25 for paid human capture. Once a robot’s useful work covers its own running cost, revenue starts paying for the data. Not all of it: deployment data still costs takeovers, upkeep, cleaning and training. But the customer is now carrying a big share of the bill. Waymo hit that moment with paid rides.

Phase 2: human augmentation. Past the gate, robots work alongside people, and their own deployments become the main data source. This is where RL finally earns its keep: on top of a distilled prior, on real outcomes, with human corrections. Humans move from demonstrators to supervisors, and the goal is to give the supervisors less and less to do.

Below, the top half is one robot’s hourly ledger, and the bottom half is a deliberately toy timeline. Push the success rate up until the gate opens, then move the gate year and see when fleet hours overtake human hours.

第一阶段:蒸馏人类。 机器人靠压缩人类的身体行为来学习,接口决定速度。最要紧的数据是有深度的:不是工厂里同一个抓取动作的循环,而是在真实家庭里那种野生的、乱糟糟的试错:抓得太用力、袜子掉了、拉错了抽屉,以及之后怎么补救。探索照样在发生,只不过是人在探索,以人的速度,带着人对「什么值得试」的判断。 这是地球上最便宜的探索,跟拿 GPU 去烧随机动作正好相反。

闸门:数据变成收入的副产品。 一台手工组装的 Memo 今天大约 2 万美元(Sunday)。摊到三年、每天工作八小时,资本成本大约是每机器人小时 2.3 美元,而花钱请人采集是每小时 25 美元。一旦机器人干的有用的活能覆盖它自己的运行成本,收入就开始替数据买单。不是全部:部署数据照样有人工接管、维护、清洗和训练的成本。但这笔账的一大块,现在是客户在付。Waymo 靠收费载客跨过了这一刻。

第二阶段:增强人类。 过了闸门,机器人跟人并肩干活,它们自己的部署成为主要的数据来源。强化学习终于在这里派上用场:建立在蒸馏出来的先验之上,用真实的结果,加上人的纠正。人从示范者变成了监督者,而目标是让监督者越来越闲。

下面上半部分是一台机器人每小时的账本,下半部分是一个故意做得很简化的时间线。把成功率往上推,直到闸门打开;然后挪一挪闸门年份,看机队的小时数什么时候超过人采集的小时数。

8. Endgame: open weights, closed humans

To feed their core models, the giants will push robot models toward open release. Any company that tries to hold a closed model moat gets replicated, often with better results, and the replica gets released. You can already see the pattern:

  • NVIDIA released Cosmos under a license that permits commercial use, and published an open reference humanoid on a Unitree chassis (The Decoder, Forbes).
  • Physical Intelligence ships openpi with released weights.
  • DeepSeek showed in early 2025 that a frontier reasoning model can be replicated and released openly within months, and it now has a stake in Unitree.
  • The economics favor it. For a lab selling general intelligence, the robot model is a complement, and commoditizing your complement grows demand for your core product.

Now look at what nobody opens. Cosmos released weights, not its 20M hours of video. No lab has opened a large real-world manipulation corpus. OpenAI is building closed collection facilities, and Waymo never open-sourced its driver.

So the endgame is pretty precise: weights converge to free, while human experience and deployment feedback stay private. A closed robot model is the most exposed asset in the stack. A closed loop, with the network of homes, hands and paying customers behind it, is the least exposed. Same line that separates the dead from the survivors in section 6.

9. What would prove me wrong

A thesis worth publishing should name its own failure conditions. Any one of these breaks part of the argument:

  1. A simulator closes the contact gap. If physics simulation becomes faithful for deformable, contact-rich manipulation, the AlphaGo Zero conditions come back and self-play wins again. Breaks claim 2.
  2. An algorithm wins on a data diet. Some lab hits 99.9% per-step reliability on long household tasks with under 10% of the leaders’ human data. Breaks claims 2 and 3.
  3. The exponent is much steeper than 0.5. If error falls close to linearly with data, the data wall shrinks by orders of magnitude and the interface edge matters less. (Slide α to 1 in the reliability demo and watch 5B hours become 50M.)
  4. A giant opens a large real-world manipulation dataset. That breaks “closed humans” and commoditizes the capture moat.
  5. Scale alone buys robustness. If a path-1 model trained on ~10M hours walks into new homes and is robust out of the box, the loop matters much less and the giants win outright. Breaks claim 5.
  6. End-to-end stalls. If the robots that actually ship reliably keep leaning on hand-written planners and rule layers, my hot take in section 3 is wrong.

What I’m watching over the next 12 months:

  • Interventions per 100 tasks, or hours between interventions, on any paid deployment. That’s the number that actually measures autonomy, and almost nobody publishes it yet
  • Sunday’s 50 beta homes: reliability on long-horizon chores, not single demos
  • Generalist past GEN-1.5: hours per week, and a scaling exponent stated by Generalist itself
  • PI’s deployment fine-tuning: how fast success climbs per week in the field, plus API pricing and paying customers
  • Capture-speed numbers for gloves and wearables, the missing x-values in the interface demo
  • The first force or tactile corpus measured in hundreds of thousands of hours
  • The first fleet whose own working hours outnumber human-captured hours: the phase-two gate

8. 终局:权重开放,人类封闭

为了喂养自己的核心模型,巨头会把机器人模型往开源推。任何想守住闭源模型护城河的公司,都会被复现,复现出来的往往效果还更好,然后被开源。这个模式已经看得见了:

  • NVIDIA 用允许商用的许可证发布了 Cosmos,还在宇树的底盘上发布了一个开放的参考人形机器人(The Decoder,Forbes)。
  • Physical Intelligence 发布了带权重的 openpi。
  • DeepSeek 在 2025 年初证明了,一个前沿推理模型几个月内就能被复现并开源,而它现在手里还有宇树的股份。
  • 经济账也支持这么做。对一个卖通用智能的实验室来说,机器人模型是互补品,把互补品变成大路货,就能拉动核心产品的需求。

再看看没人开放的东西。Cosmos 开放了权重,没开放那 2,000 万小时视频。没有哪个实验室开放过大规模的真实操作数据集。OpenAI 在建封闭的采集设施,Waymo 也从没开源过它的驾驶系统。

所以终局其实挺清楚:权重会趋向免费,人类经验和部署反馈会一直私有。 闭源的机器人模型是整个技术栈里最暴露的资产;一个闭源的闭环,连同它背后那张由家庭、双手和付费客户织成的网,是最不暴露的。这跟第 6 节里区分死者和幸存者的,是同一条线。

9. 什么情况能证明我错了

一个值得发表的论点,应该说清楚自己在什么情况下会失效。下面任何一条成立,论证都会断掉一部分:

  1. 模拟器补上了接触这道坎。 如果物理仿真对可变形、富接触的操作变得足够逼真,AlphaGo Zero 的那几个条件就回来了,自我对弈探索会再赢一次。论点 2 失效。
  2. 某个算法靠少量数据赢了。 某个实验室用不到领先者 10% 的人类数据,在长程家务上做到每步 99.9% 的可靠性。论点 2 和 3 失效。
  3. 指数远比 0.5 陡。 如果误差随数据几乎线性下降,数据墙就会矮好几个数量级,接口的优势也就没那么要紧了。(去可靠性那个演示里把 α 拉到 1,看 50 亿小时怎么变成 5,000 万。)
  4. 某个巨头开放了大规模真实操作数据集。 那「人类封闭」就不成立了,采集这道护城河也会被抹平。
  5. 光靠规模就买到了鲁棒性。 如果一个用约一千万小时训练出来的第一条路线的模型,走进陌生家庭就开箱即稳,那闭环就没那么要紧了,巨头直接通吃。论点 5 失效。
  6. 端到端卡住了。 如果真正可靠出货的机器人,一直离不开人写的规划器和规则层,那我在第 3 节的暴论就是错的。

接下来 12 个月我会盯着的:

  • 任何付费部署里,每 100 个任务要人工介入几次,或者两次介入之间隔多少小时。 这才是真正衡量自主性的数字,而现在几乎没人公布
  • Sunday 那 50 个测试家庭:看长程家务的可靠性,而不是单条演示
  • Generalist 在 GEN-1.5 之后:每周新增多少小时,以及 Generalist 自己公布的缩放指数
  • PI 的部署端微调:在现场每周能把成功率提高多少,以及 API 的定价和付费客户数
  • 手套和可穿戴设备的采集速度,也就是接口演示里缺的那几个横坐标
  • 第一个以几十万小时计的力觉或触觉数据集
  • 第一支自己干活的小时数超过人类采集小时数的机队:第二阶段的闸门

10. Further out: closed hive versus open hive

Now the speculative part. To be clear, it doesn’t follow from the data economics above. It’s a separate bet about society, so take it as a hunch, not a forecast.

Section 8 ended on a moat: weights go free, human experience stays private. Push the logic one step further and that moat dissolves too. A distillation interface is a machine for making tacit labor legible. Every glove, wearable and camera rig turns know-how that was never written down into data. Point the same machinery at materials, processes and supply chains. Once AI can reverse-engineer matter and not just code, secrecy stops being a moat and becomes a liability. The rational response isn’t higher walls. It’s radical openness.

From the outside, that will look a lot like an open-source religion whose god isn’t a person but a protocol. It’s less of a metaphor than it sounds. AI is already a kind of swarm mind: one main brain emerging out of distributed models, data, sensors and agents. Once physical production becomes legible, the only thing that can stand up to a proprietary hive mind is an open hive of equal or greater scale.

Humans don’t become obsolete in that world. They become the hive’s hands: improvising, negotiating, repairing, handling the environments too messy for machines alone. Phase 2 in section 7 is the first draft of this, with humans as supervisors and correctors in the loop. Scale it up and flexible human execution becomes the frontier’s immune system, the part that handles the long tail no dataset has covered yet.

So the real fight isn’t AI versus humanity. It’s closed hive versus open hive: corporate feudalism against a distributed, post-secret civilization.

The open side wins only if it pulls off something very hard: turning humans into agents of a collective mind without extinguishing the human capacity to refuse. If it can, it’s the seed of something much bigger than robotics. If it can’t, it’s digital feudalism with better liturgy.

That’s the stake behind every glove, every dataset and every open-weights release of the next five years. A robot is joints and links, and its priors come from general intelligence. What stays scarce is the skill in human hands, and the loop that keeps turning real-world experience into capability you can deliver. Capture is the entry, the model is the converter, and finished tasks are the only test. Whoever runs that loop fastest survives. What they feed it into decides which hive we end up living in.

10. 再往远看:封闭蜂巢与开放蜂巢

下面是推测的部分。先说清楚:它并不能从前面的数据经济学里推出来,是一个独立的、关于社会的判断。当成直觉看就好,不是预测。

第 8 节停在一道护城河上:权重免费,人类经验私有。顺着这个逻辑再往前推一步,这道护城河也会消失。蒸馏接口,本质上是一台把隐性劳动变得可读的机器。 每一只手套、每一台可穿戴设备、每一套摄像头,都在把从没被写下来的手艺变成数据。把同一套机器对准材料、工艺和供应链。一旦 AI 能逆向工程物质,而不只是代码,保密就不再是护城河,而是负债。 理性的回应不是把墙砌得更高,而是彻底开放。

从外面看,这很像一种开源宗教,它的神不是某个人,而是一套协议。这没有听起来那么像比喻。AI 本来就已经是一种群体心智:一个主脑,从分布式的模型、数据、传感器和智能体里涌现出来。一旦物理生产变得可读,唯一能跟私有的蜂巢心智抗衡的,是一个同等或更大规模的开放蜂巢。

在那样的世界里,人并不会过时。人会成为蜂巢的双手:即兴应变、谈判、修理,处理那些对机器来说太乱的环境。第 7 节的第二阶段就是它的初稿,人作为监督者和纠错者留在回路里。把它放大,人的灵活执行力就成了前沿的免疫系统,专门处理任何数据集都还没覆盖到的长尾。

所以真正的对抗不是 AI 对人类,而是封闭蜂巢对开放蜂巢:企业封建制,对一个分布式的、不再有秘密的文明。

开放的一方要赢,得做成一件非常难的事:让人成为集体心智的执行者,却不磨灭人说「不」的能力。 做成了,它就是比机器人大得多的东西的种子;做不成,它就只是礼拜仪式更漂亮一点的数字封建制。

这就是未来五年里,每一只手套、每一个数据集、每一次开源权重背后押着的东西。机器人就是关节加连杆,它的先验来自通用智能。真正稀缺的,是人手里的那份手艺,以及一个能持续把真实世界经验转化成可交付能力的闭环。采集是入口,模型是转化机制,真正完成的任务才是唯一的检验标准。谁把这个闭环转得最快,谁就活下来。它最终喂给什么样的容器,决定了我们最后住在哪一个蜂巢里。

Appendix: assumptions behind the math

Assumption Value used Basis
Action bandwidth 50 joints at 50 Hz Dexterous humanoid, typical control rate
Vision bandwidth 3 cameras, 512 × 512 RGB, 30 fps Typical multi-camera manipulation rig
Trajectory count 10 values per joint, 50 joints, 500 steps Deliberately coarse; the real space is continuous
Training FLOPs 6 × params × tokens Standard dense-transformer estimate
Tokens per hour, lean 3.5M: 3 cams, 256 px, 10 fps, 16×16 patches, 8x temporal compression My assumption
Tokens per hour, rich 83M: 3 cams, 512 px, 30 fps, 16×16 patches, 4x temporal compression My assumption
B300 throughput 2.25 PFLOPS dense BF16, 40% utilization Flopper
Cloud price \$5.00–9.16 per GPU-hour Reserved estimate; on-demand
Capture labor \$25 per hour, fully loaded Assumption
Hardware life 4,000 hours Two years at 2,000 hours
Capture speeds 231 / 111 / 35 demos per 15 min UMI paper, cup arrangement, read from its figure
Data scaling exponent L ∝ D−0.5 Secondhand report for GEN-0; task-dependent; slide it in the reliability demo
Driving hours 271.3M rider-only miles (through June 2026) at 15–25 mph Waymo

Every demo on this page computes from the rows above, except the loop demo, which is a labelled toy model. Nothing is a product photo; the arm, the workspace and the timeline are drawings.

附录:背后的假设

假设 取值 依据
动作带宽 50 个关节,50 Hz 灵巧人形机器人,常见控制频率
视觉带宽 3 个摄像头,512 × 512 RGB,30 帧 常见的多摄像头操作平台
轨迹数 每关节 10 种取值,50 个关节,500 步 故意取得很粗;真实空间是连续的
训练 FLOPs 6 × 参数量 × token 数 稠密 Transformer 的标准估算
每小时 token 数(精简) 350 万:3 路、256 像素、10 帧、16×16 patch、时间压缩 8 倍 我的假设
每小时 token 数(丰富) 8,300 万:3 路、512 像素、30 帧、16×16 patch、时间压缩 4 倍 我的假设
B300 吞吐 稠密 BF16 2.25 PFLOPS,利用率 40% Flopper
云价格 每 GPU 小时 5.00–9.16 美元 包年估计;按需价
采集人工 每小时 25 美元(含全部成本) 假设
硬件寿命 4,000 小时 两年,每年 2,000 小时
采集速度 每 15 分钟 231 / 111 / 35 条示范 UMI 论文,摆杯子任务,从图中读出
数据缩放指数 L ∝ D−0.5 GEN-0 的二手报道值;随任务而变;可在可靠性演示里拖动
驾驶小时数 2.713 亿英里纯乘客里程(截至 2026 年 6 月),15–25 英里/小时 Waymo

这一页上的每个演示都是从上面这些行算出来的,闭环那个除外,它是标明了的玩具模型。没有任何产品照片;机械臂、工作空间和时间线都是画出来的。

Sources

来源