a16z 对谈 Tutor：16 美元一小时的机器人，卖给九成从没买过机器人的工厂

信源说明：本文基于 a16z Podcast 一期节目整理（视频 ID lXkjXSul6yg，全长约 45 分钟，英文转写 23 段、0s–2697s 已完整读取，文中数字与引语均依英文转写）。嘉宾为 Tutor Intelligence 联合创始人兼 CEO Josh Gruenstein，身份经核实（公司官网、其个人主页及 a16z 投资公告；转写自述五年半前创业，与公开资料 2021 年 2 月注册成立吻合）；主持人身份从转写判断为 a16z 投资人（自述投资前访谈过其客户、a16z 发布过供应链与机器人报告），姓名未见于转写。利益相关：Tutor Intelligence 为 a16z 投资组合公司——a16z 于 2026 年 11 月 2 日（本文成文前一日）宣布领投其 7500 万美元 B 轮，本场访谈即配合该轮发布，读者请按投资人对谈自家被投创始人的口径取信。文中全部数字与案例均为节目口述，以「节目称／Gruenstein 认为」分层标注，除注明「经核实」者外本文未做独立核实；ASR 疑似错词以「转写疑作」标注，文末附校正表与完整转写原文及全量中文翻译。同频道强相关旧文：a16z AD50 再工业化机制拆解——彼时已展开的国防首购、制造中层 co-engineering、资本「提前于需求」等背景本文不再重复，本期补的是其中缺的一块：自动化需求侧（机器人公司）自己怎么长出来。

先说结论

这期访谈表面上聊机器人技术，真正的主轴是一套「绿色地带」（greenfield）生意逻辑。Gruenstein 的起点数字：美国 2025 年只部署了约 3.8 万个工业机器人工作单元（节目称，对比全球约一百万，且不含存量旧机），十分之九的美国工厂和仓库没有任何机器人——所以几乎全部潜在客户都是第一次买机器人的人，而他们默认「机器人不适合我」。Tutor 的解法不是先讲模型，而是先卖价格：每小时计费、零资本开支的机器人劳动力（节目称 16 美元/小时，官网标「14 美元/小时起」经核实），首通电话到签约平均 40–60 天，机器人到场当天投产。数据策略同样反直觉：不追「从零数据造智能」的算法，而是让付费部署本身「排放」出训练数据——远程人类「导师」（tutors）教机器人干活，数据的经济价值由客户正在付钱这一事实背书。由此推出全场最重的三个判断：机器人必须是模型、硬件、供应链、客户、产品合一的全栈平台，丢掉任何一层就消灭九成取舍空间；美国作为机器人绿地市场反而免去「拆掉重建」的包袱，中国押注的上一代专用自动化在他看来事后或成「巨大错误」；以及机器人公司的跨代资产不是任何一次模型架构，而是「飞轮与循环本身」。

一、九成工厂没有机器人：先卖「16 美元一小时」，再谈 AI

Gruenstein 给出的市场底数（均为节目称）：美国一年部署约 3.8 万个工业机器人工作单元；若十分之九的工厂和仓库没有任何机器人，「那么几乎整个市场都是绿色地带」。这些潜在客户了解老一代自动化长什么样，并且「假设机器人不适合自己」。Tutor 打破这个假设的方式不是谈 AI、模型或导师，而是冷启动电话：「嘿，你考虑过 16 美元一小时的机器人吗？」——对方从没考虑过，因为这对其是全新命题，于是会议约到了。他强调，真正让铺开速度「比传统工业自动化行业快多个数量级」的是商业模式带来的激励对齐：通用机器人不定制、从工厂直发客户现场，客户不承担任何经济风险。节奏是：第一通电话到签约平均 40–60 天，签约到机器人进场干活约 40 天，到场数小时到一天内投产。

需求侧的底色是劳动力而非技术。Gruenstein 描述的搬运岗位（节目称）：三个月的流失率，干这活的平均年龄五十五岁、即将退休；「这些地方根本招不到人」。主持人由此补了一句本场最生动的成本对标：机器人真正的对手是「一种你喂它米饭就能干活、效率高得惊人的东西——人体」。所以客户来电不是黑色星期五式的救急，而是战略性的劳动力规划：企业知道自己员工的年龄结构、知道需求在长期趋势上行、知道履约成本在涨，而它们不是有高毛利缓冲可压缩的软件公司——「生意要么成立，要么不成立」。客户要找的是长期负责「机器人这件事」的伙伴：做除臭剂的就专注做除臭剂、做蜡笔的就专注做蜡笔，机器人交给专家。主持人也交代了一个尽调发现：与多数机器人公司扎堆追财富五十强不同，Tutor 相当多数的客户是本地中小企业——Gruenstein 的解释很平：中小制造与物流企业就是这个市场的大多数，服务市场就得这么卖。

落地之后，人的态度怎么变？Gruenstein 说运营群体最直接：「它能干活。我知道所有需要这个活的地方——全都装上。」于是从单站点单台机器人扩到多站点十几台。一个组织细节：机器人行业里买家不是用户——买的是 COO 或总经理，用的是天天和机器人并肩的产线工人；Tutor 不追求整厂全自动，只自动化部分岗位，所以必须经营工人与机器人之间的关系。他有学术界的现成工具箱（HRI，人机交互研究），最典型的一招是命名：每台机器人都有名字，复购客户获得给新机器人命名的权利——有人用孩子的名字、球星、流行歌星命名机器人（命名体系最初用飓风名单，绕太多圈后改成技师随喜）。他还引用一条推文概括工厂经理层面的心态：「机器人就是机器人，除非它能干活——一旦能干活，它就是机器。」依恋的根源是机器人真的在帮忙：工人少干重体力活、升到更高岗位，经理的日子「好过一千倍」。

二、Tutor 这个名字：付费部署才「排放」出有价值的数据

公司名来自一个思想实验。Gruenstein 的推理链：地球八十亿人都拥有机器人该有的那种基础物理智能；假想一个无限强大的软件能把任何一个人连接到任何一台机器人（他称之为机械降神式的「无限机器人智能」）——即便把这个智能白送给你，问题也没解决：机器人制造还是家庭作坊式产业，软件不存在，客户也不存在（机器人干不了客户要的活）。反过来，LLM 的成功靠的是「人类智能的互联网」以自然排放的方式造出了数据集——人们在社交媒体发帖、写邮件，副产品就是训练数据；而物理劳动不联网：「我端起咖啡杯，没人记录我指尖的全部压力和位姿」。他的结论：与其找一种从零数据凭空造出智能的算法（「那是一场对我们 rigged 的比赛，计算机视觉不是这么赢的，NLP 也不是」），不如先建一张真实世界的机器人网络，用人类遥操作者——「导师」——教机器人新任务；部署本身在赚钱的同时，像互联网一样「排放」出机器人训练数据。这被他称为一种独特的正和游戏：不是巨型资本项目，边建网络边把所有人的商品做便宜。

这套数据观里最值得移植的是价值判定标准。主持人的质疑代表了行业默认路线：生成式 AI 的奇迹是「分布进、分布出」——用图像训练、人们就要图像，用文本训练、人们就要文本；机器人所需的能力可能刚好在训练分布之外。Gruenstein 的回应把问题换成「数据从哪来」：如果你给所有人头上绑摄像头、派去印尼各地采集数据，没有任何「看不见的手」保证那些数据值钱；而机器人在真实客户现场干着客户付费的活，其数据「必然有经济价值」——赚钱的生意多半也在产出数据，经济激励天然对齐。主持人再追一层（反方）：经济上值钱的活可能维度极高，解掉九成后，最后一成也许要人顶很久。Gruenstein 承认「那在经济上可能是完全可以接受的」，并补一句「我们不是纯粹主义者」：模拟、互联网级数据都在工具箱里，用来拓宽分布。

导师本身则被当成一个单位经济学问题来管：像算 H200 的服务成本与占空比那样，去算「需要从人类拿到、而模型暂时给不了的那种异构智能画像」的成本；为从模型里抠出最后一成成色要做多少工程，对比直接用导师，是一道算术题。一个有趣的质量信号：投资前访谈中，大多数客户根本分不清什么时候是导师在操作、什么时候是模型在跑——Gruenstein 认为这是好事：「你想要的反馈是我以规模生产的洗发水便宜了五分钱；如果反馈是这机器人挺好看，那大概是台很差的机器人。」至于外部模型浪潮：他说各家实验室因为直接的市场激励、用 CAD 与 3D 建模数据做后训练，补上了 GPT 级模型与「动作扩散 transformer 头」之间空间推理的最后一公里——「我们被向前弹射了很多年」；因为机器人是系统问题，前沿模型的每次进步对他是加速器而非替代威胁。Tutor 2021 年用卷积神经网络起家，历次范式转移都经历过：「永远有价值的是数据流」，架构技巧只要能加速飞轮就值得用，而「有跨时代价值的资产是飞轮与循环本身」。

三、全栈是被迫的：供应链「太小、而且不对」

主持人问了个连续谱问题：从「买台 Unitree 写 prompt」到「自己机加工零件」，你在哪？Gruenstein 的回答暴露了美国机器人供给侧的真实状态（节目称）：2025 年美国部署的 3.8 万个工作单元里，Unitree 这类新形态机器人在生产环境的部署「大概是零」——有就会体现在数字里；而 Fanuc、ABB 造的机器人大多不在美国生产，即便在美国造，也是为系统集成商、为快速重复动作设计的物件，「不是你会拿来与前沿模型协同设计的那组取舍」。结论：现有机器人供应链「既太小、又不对」，Tutor 被迫从零设计并自制整机——相机、执行器等部件仍从 OEM 采购（他们规模更大），但「会跑、带机械臂、能干活」的机器人本体，在马萨诸塞州 Watertown（经核实为公司总部所在地）规模化自产。

两款机器人是「硬件可制造性 × 模型真实能力 × 市场缺口」三方协同设计的样本。Cassie 是「头蛮牛」：单臂加大量吸盘，跑两亿到四亿参数量级的基础模型——这直接继承了 2021 年「基础模型能预测可供性（affordances）与抓取」的世界观；它对准的缺口是，美国实体经济里一切都要装上托盘或装进纸箱运输，于是存在巨量徒手搬运的活。Sonny 则是「工业半人形」：仿佛看着一台 Unitree 说「我要它在工厂里以人类的速率和可靠性跑三到十年」，于是得出另一组设计决策——轮式而非腿式（室内工业地面是抛光的）、载荷更高、移动更快、满足可在人周围作业的工业协作安全标准。节目称波士顿有一百台 Sonny 在「数据工厂」里采集数据、推进研究，真实世界还有一支增长的 Sonny 机队做更灵巧的双手操作；在制造与物流场景内，这两款机器人「今天大概能干 80% 到 90% 的活」。干不了的也讲得清楚：出不了门——不是移动算法问题，是硬件（轮子过不了碎石路），建筑与垃圾填埋场回收这类场景要等腿式形态进入其「具身形态万神殿」；家庭则是「设备太工业」加上经济账对不上（工厂里一台机器人创造的价值，对照家庭付费意愿）。

Gruenstein 顺手拆掉了大众对制造业的想象：《How It’s Made》里那种无人连杆机器是「美国制造物流的前 0.01%——如果存在，多半已经被拍进节目了」；真实画面是百万平方英尺厂房里的人站在传送带和折叠桌旁徒手作业。他还给了一个更反直觉的事实：任何你买过的、装在盒子里的东西，都是人装进去的——包括 iPhone 的包装，在最高规模化程度上也是人。产品方法论上，主持人观察到多数机器人公司还停在 v1，Tutor 已有两代且快速迭代；Gruenstein 说他们刻意把数字产品的方法搬进硬件：自制整机让制造流程能每两周推送变更，机器人代码每日或每两周一个大版本；产品职能与客户成功、实施职能坐成一个团队；研究团队一端嵌进客户问题空间拉取特定数据，一端打「登月」赌注——模型侧一有突破就围绕它协同设计产品。而多数同行的死法被他说得很直白：拿大客户一纸「闪亮的承诺书」，花两年造一套高度定制系统，大概率做不成（做产品不可能一次成型），做成了客户需求也可能没了——「这种事我眼看连续发生了五到十次；要是发生在我们身上就太难堪了」。所以多元客户群是逼出「通用机器人」产品形态的强制函数：两周跟一家中型客户签约、机器人随即干活，且这与其他客户用的是同一套产品。

四、拆与建：中国押注上一代，美国绿地免拆

主持人把话头引向工厂选址与「一厂镇」的未来，Gruenstein 给出本场最重的地缘判断（观点级，非事实）：机器人将是美国与其他国家之间的「伟大的均衡器」——像中国这样投资了上一代工业自动化的国家，某些方面因此非常成功，但事后看可能是巨大错误，因为那一代自动化做的是专用机器；美国正在用通用机器人做工业产能的绿地建设，在中国「他们将不得不拆掉重建」，而美国不需要以同样的方式拆。配套的另一个预测：如果出现「无限灵活的劳动基底」，工厂数量会像云计算加软件工具让企业数量爆炸那样变多——任何人都将能在物理世界开办并规模化一门生意。

他自己对供应链现实的表述则克制得多：2026 年的今天「无论谁怎么说」，全球供应链高度互联互依，中国公司在投美国、美国公司在投中国，所有人都在建设自己的工业产能（主持人提到 a16z 发布过供应链与机器人报告、中国在其中非常核心，得到的正是这段回应）。Tutor 自己的垂直化逻辑是量变引起的：机器人行业是家庭作坊，起步时子组件全从 OEM 采购；量长到占对方产能实质比例后，「他们不再是 OEM，成了光荣的合同制造商」，关系动力学随之改变——收进来还是指挥他们做，答案是「某种组合」。但他把结论拔到了产业层：Tutor 要成功，美国就必须有一条成熟的机器人供应链；而因为自己是这条供应链需求的很大一部分，去建它既是义务也是机会——这就是在 Watertown 造机器人、并继续把更多组件垂直整合到美国的原因。这正接着 AD50 那期拆过的再工业化机制往下走：那边讲的是资本与国防需求如何「提前建造」产能，这期补上了需求侧样本——一家机器人公司如何倒逼供应链本土化。

五、把机器人当劳动力做生意：单位经济学与研究组织

Gruenstein 的价值公式只有一行：「机器人的价值 = 它能做的工作量 ÷ 它的成本，其他都是评论。」据此，只要能在美国实体经济高度流动的劳动力市场里以更低价格提供同等能力，市场就是全部体力劳动。成本结构正在搬家：传统机器人是一笔大 capex 加十年回本的 ROI 题，而 Tutor「现阶段持续成本占比更大，这让很多人意外」——钱越来越多花在持续的推理与遥操作上。获客也产品化了：官网定价计算器让客户输入运营数据，估算「初始部署两台机器人一年省十万、一百万、一亿美元」，且不是五年定制工程路线图，而是「两个月后开始省、不压任何资本开支」。产业侧的不连续点他举了一个精确案例（节目称）：一家威斯康星住房工厂造房目前比传统现场施工贵约 5%，但只要比同质量现场建造便宜哪怕 1%，开发商的全部需求会立刻切换到工厂造房；垃圾填埋场的物料回收同理。他的收束愿景是把 Tutor 做成「机器人基底」——其他企业借以改造物理经济的劳动力平台。

研究侧的组织同样围绕「真实世界优先」。他按 LLM 的演化类比：数据来源高度异构，但「数据金字塔的顶端是『做你想做之事』的真实世界强化学习轨迹」；为此在波士顿建了一百台机器人的 Data Factory One（节目称是全美最大的半人形数据采集场），专门研究大规模真实世界强化学习——因为强化学习「在小批量下基本不奏效」，只有大机队才玩得起学术界与工业界都还没玩过的那些方法。组织哲学收在最后一段：「这件事上没有任何地方可以推卸责任——不管你愿不愿意，你都要与每一层打交道，问题只在于你是否理解它。」团队遍布全美（尤其 go-to-market 与技师），而这份工作的独特之处是「早上在白板前聊模型架构，下午在产线看机器人干得怎么样」。

机制 → 二阶影响 → 读者启发

1. 绿色地带市场：分发与定价胜过产品参数。信号：九成工厂无机器人、买家全是首次购买者（节目称 3.8 万工作单元/年）→ 直接原因：存量自动化成本结构（大 capex、系统集成商、定制工程）把中小企业排除在外 → 得失：按时计费让中小厂首次买得起「劳动力」，传统集成商与 capex 式自动化厂商的渠道被绕开 → 二阶影响：机器人采购从设备预算转入劳动力预算，需求随人口结构（55 岁、三个月流失率）变成战略性、周期性 → 观察指标：首购客户占比、签约周期（Tutor 口径 40–60 天）、复购扩张速度 → 启发：面对从未用过品类的市场，卖「熟悉的东西」（每小时劳动力）而不是「新的东西」（AI 机器人）。
2. 数据飞轮：让数据成为收入的排放物。信号：物理劳动数据不联网（咖啡杯没人记录）→ 直接原因：遥操作导师让「干活」与「采数据」合一，客户付费保证数据经济价值（对照：印尼绑摄像头无「看不见的手」）→ 得失：拥有付费部署的公司同时拥有数据护城湖，纯采集/纯仿真路线贬值 → 二阶影响：前沿实验室的 3D 空间后训练外溢，加速而非替代部署型公司；tutor 比例成为成本曲线的关键旋钮（H200 占空比式算法）→ 观察指标：导师工时占比/模型自主率、单位任务数据成本 → 启发：评估任何 AI 数据战略时先问一个问题——这份数据是不是某笔真实收入的副产品。
3. 全栈平台：不成熟市场的强制形态。信号：供应链「太小而且不对」（Fanuc 为集成商设计、Unitree 生产部署≈0）→ 直接原因：跨层取舍空间决定前进速度，丢弃一层等于消灭九成取舍 → 得失：自研整机者（Tutor 在 Watertown）掌控协同设计；OEM 在量变后沦为「光荣的合同制造商」→ 二阶影响：水平→垂直是市场不成熟的表现（类比早期计算机），成熟后再回到平台化，让非机器人企业能在其上生长 → 观察指标：制造变更周期（两周）、机型覆盖场景比例（80–90% 口径）、美国本土机器人供应链成型度 → 启发：垂直整合是阶段性武器，为的是将来「越建越少取舍」，而不是终身形态。
4. 地缘窗口：绿地免拆 vs 存量重置。信号：中国深度投资上一代专用自动化（嘉宾观点）→ 直接原因：通用机器人与专用机器是两代资产，旧产能对新任务是被锁定而非可复用 → 得失：美国绿地建设「不必拆」，存量自动化强国面临重置成本（「他们将不得不拆掉」）→ 二阶影响：机器人劳动基底若成立，工厂数量像软件企业一样膨胀，选址与「一厂镇」逻辑重构 → 观察指标：中美机器人在对方市场的投资流向、美本土机器人产能（如 Watertown 类工厂）扩张、每小时机器人价格对当地工资的剪刀差 → 启发：技术代际切换时，「没有历史资产」可能从劣势变成免拆红利——但这同时是嘉宾为自家市场路线背书的利益相关判断。

转写校正表

账本高优先级 43 条均在正文有落点，未采用 0 条。

转写原文（修正版）

[0s - 120s] No matter what anybody says, today in 2026, we have a highly interconnected, interdependent global supply chain. Both U.S. companies and Chinese companies are thinking very hard about what does the long-term strategy of this look like, and ultimately, everybody trying to build their industrial capacity. We see a lot of robot pitches. Very few people thinking about what the robot is supposed to do at the end of all of this. Almost everybody we’re selling a robot to has never purchased a robot before. Almost all of the market is greenfield. The reality is, you need to build a full-stack platform. Tutor Robots have helped manufacture and ship over a quarter of a billion products to American consumers. We’re training frontier robot AI models, manufacturing robots at scale, and directly deploying into the largest businesses in the supply chain. How does product innovation work in robotics? Most companies are just on v1, whereas you’ve got two versions of the robot and you continue to iterate really fast. So, Josh, thanks for for joining us. So, we’re with Josh Gruenstein, who is the CEO and founder co-founder co-founder of Tutor Intelligence, and we’re going to be talking about robots today. And so, maybe before we jump in, just give a quick background, maybe of yourself and then the company. Yeah, so I’m Josh. I started Tutor about five and a half years ago with the core thesis that if you could just start deploying robots out into the world and to all the places in the physical economy where robots are needed the most, you could start to build the datasets, the models, and the supply chain that’s needed to proliferate robots throughout the physical economy. So we’re about 160 people, based primarily in Boston, and we’re training frontier robot AI models, manufacturing robots at scale, and directly deploying into the largest businesses in the supply chain. And maybe talk a little bit about like you know we see a lot of pitches. There’s a lot of companies that talk about doing robotics but don’t actually deploy robots or have not yet. Yeah, and you’re one of the few companies that’s actually deployed robotics at scale. So maybe just talk about like the state of the business. Yeah, we have robots deployed across the United States.

[120s - 240s] Everywhere from small mom and pop businesses to Fortune fifty, Fortune one hundred businesses. Claude told me last night that it estimates Tutor robots have helped manufacture and ship over a quarter of a billion products to American consumers. So, like presumably, most American consumers have are a Tutor customer a couple layers down. I think this comes from kind of the origin story of the business, which is us looking at the state of robot learning. It was obvious to us that robot intelligence and robot software was this bottleneck, right? Robots for sixty years has been this custom engineered machines that can’t scale. Yeah, so we we started the company with this thesis that if we could just start deploying robots, we could build the labeled data sets that help us train these models and deliver this value to the American consumer. We see a lot of robot pitches. Ninety nine percent of them start with some insights about the model and how you’re going to build the model, or some insight about the data shape of the data that is. And very few people thinking about what the robot is supposed to do at the end of all of this. You started from the reverse, and you had the insight that you could actually tutor the robot, so to speak, the name of your company. Where did that insight come from? It came from kind of this thought experiment, where yeah, we had this insight that hey, there are eight billion people on Earth who all have this baseline physical intelligence that we want our robots to have. And so imagine you had infinitely powerful software that could connect any human being on Earth to any robot and kind of embody that robot with intelligence. That’s kind of like this Deus Ex Machina of like suddenly you have unlimited, infinite robot intelligence. Let’s say you take that for granted. Now what? Right. You know the problem isn’t actually solved. Right. You still need to actually manufacture robots at scale. Right. Robots are this cottage industry today in the United States.

[240s - 360s] That’s a rounding error, even in the most automated industries like manufacturing logistics. So that doesn’t exist. Certainly, none of the software exists, and none of the customers exist because, you know, again, the robots can’t do any of the jobs that customers want robots to do. So robotics is not, I think, any one layer of the stack. It’s not just about intelligence. It’s not just about hardware. It’s not just about the customer. It’s about kind of figuring out how to put all these three pieces together, and it’s the combination of all three that actually lets you build, in my view, a successful platform at any one of those layers. Yeah, internally we talk about robotics being the ultimate systems company, yeah, yeah, as opposed to being one, one or the other. You know, it’s funny about this though, which is, um, and it’s a question I wanted to ask you is, I mean, you’ve got a pretty deep technical background, right? You know, and like, uh, I mean, MIT robotics, yeah, is that fair? Um, but whenever I talk to you, I feel like I’m talking to an operations guy that talks about like, yeah, they’re talking about unit economics and operations and supply chain. So I mean, and even a go to market guy, go to market guy. So it’s kind of you’ve done this, like the transformation is complete. And so, like, maybe this is kind of an open ended question, but like I’ve always wondered, like, is the harder problem here actually that stuff? No, but but I think you have to do all of it is kind of the thing. And like I, when I was just a roboticist, right, and only working on the technology side of this problem, I think my strength as a roboticist was that it was very full stack, and I was willing to think about the manufacturing, the hardware. I was willing to think about FPGAs and ASICs, and like if you view robotics in the same way that I do, that it’s this like ultimate systems problem. The only way you can make forward progress is by figuring out all of the clever tradeoffs across as many layers of the stack as you can. And if you throw out one of the layers of the stack, and you’re like, well, we’re not going to touch that, you’ve eliminated ninety percent of your tradeoff space to make progress. Interesting. And then starting a business, it became really clear to me that like, oh, okay, robotics is not actually full stack like model, hardware, software.

[360s - 480s] It’s full stack, like model, hardware, software, supply chain, customers, product, right? It’s like exactly. And similarly, you ask the question, like, why hasn’t robotics worked? Right? There’s been a lot of robot startups. I would argue no robot startup has been like really massively economically successful. Why is that? And I, I think it’s because one, you don’t have this core primitive, which is intelligence, and that lets you kind of bring, the the, unit economics of the system to a certain level. But two, it’s that it’s the super sort of disaggregated market, and you have infinite players, and they’re all trying to work with each other. And the reality of what we need in robotics is to be able to make tradeoffs across the full stack. You know, it’s interesting. So this is an observation that we have had, and it’s it’s interesting to articulate it this way, which is most robot companies start horizontal, and they all go vertical. Every one of them, like they literally start like we have a model or whatever, and then you know they start like buying 3D printers. And do you think that that is kind of because the market is so immature, like you had to do in the early days of cars, or like you think it’s actually inherent in this domain? Like, like even in 20 years, you will have to verticalize just because the economy of movement or something requires that. No, I think it’s immaturity in the market, and you can see similar things in like the early computing market. And so, to ask the question then becomes, okay, how do you build kind of the general platforms that let people build businesses that aren’t robotics businesses, right? We have all these incredible software businesses today, and the people who run them don’t have to think about chip architectures, right? How do we make that possible in the physical world with robotics, and how you go about building that platform? And I agree with you. People have tried to solve that problem by thinking about very narrow layers of the stack, and in our view, the reality is you need to build a full stack platform. And the only way to know that that full stack platform really works is to actually have successful business units operating on top of it.

[480s - 600s] I mean, I mean, are you open to like it being the case that actually the physical world is different than the digital world, and the reason we horizontalized compute is because you have clean abstractions that are generally applied, and that’s just not the case for like navigating, you know, physical space. Or I think the beauty of something like a humanoid is it’s the same; it’s that clean abstraction, right? You think about like what is a CPU, and it’s this general computing machine, and it’s not perfectly optimized for every use cases, right? But it’s general enough that you can run most computing work. Oh, but that’s so that’s your horizontal component with the human. But for example, would you use the same, you know, brain or you know, like subcomponents to make a humanoid as you would to like make a drone or to make some other form factor? No, but you can make that same analogy in compute, right? We have CPUs, we have GPUs, we have ASICs, right? So I think you track the maturity of the market, and you start in this world of specialization, custom machines, because that’s the only thing the market can support. You move towards general purpose, and in robotics, this is going to look like humanoids and similar form factors, and then eventually you kind of break past to, you know, that respecialization as the market is big enough. Why is this thing called Tutor? Yeah, intelligence. So Tutor was founded kind of with this thesis that there is a whole internet of human intelligence that is networked, right? That is posting on social media, that’s writing emails, and that network of human intelligence for LLMs and other kind of digital world models, by natural emission, has created this beautiful dataset that we can use to train general purpose models. And the issue is our work in the physical world is not similarly networked, right? Like I pick up my coffee cup, and nobody knows, right? Like nobody recorded all of the pressures of of my fingertips and where they were and the poses, and so we don’t have this asset. Um, and the core idea behind Tutor is: what if instead of trying to find kind of this algorithm where from no data we could.

[600s - 720s] Builder of an intelligence of us, we were like, yeah, that that seems like a problem that’s rigged against us, right? That’s not how this worked in computer vision. It’s not how this worked in NLP. What if we could build this network of robots out in the real world? We could connect them with human teleoperators, which we call tutors, that teach the robots how to do new tasks. And in the same way, the internet kind of by emission created this data set to train these frontier models. We can do the same with robots, and we can build this internet of robot training data. And there are all these beautiful kind of emergent properties. We’re like, okay, suddenly it’s not this massive capital project; it it is to go build this network. But also, you’re creating incredible economic value as you’re doing it, right? You’re making everybody’s goods cheaper as you deploy these robots and collect this data. And so, it’s very positive sum, I think, in this very unique way. And so, when we thought about, okay, what are we going to name this business? You know, it’s not strictly artificial intelligence. There’s some artificial intelligence, but it’s ultimately like this tutor intelligence that we’re building that was connecting and controlling these robots. Yeah, that’s an amazing insight. And the funny thing is, when we talked to your customers before we invested, most of them didn’t seem to know or didn’t even seem to care that there was a tutor, or when the tutor was operating and when the your learning model was operating, and that credit to you guys that you’ve made those very very seamless too. Yeah, this is to me a signal that you know the robots are good, right? You want the feedback to be like my shampoo is that I’m manufacturing at volume. It’s now five cents cheaper than it was. If the feedback is like the robot is pretty, it’s like ah, probably really bad robot. Is the goal to not have tutors in the future, or does it matter? We think of it like at a very pragmatic level of like, what is the operational complexity of tutors? What is the economic impact of tutors? Right? Like, think about what is the cost to serve models with an H200, right? And you can math that out. And what’s the duty cycle and utilization of that H200? And you can think very similarly about

[720s - 840s] Human intelligence, right? What is sort of the heterogeneous profile of intelligence that you really want from humans that you can’t get yet from your models? And then it becomes a unit economics question, right? Like, how much engineering do you have to do to get that last nine out of your models versus using tutors? And really, you know, going back to what we’re talking about, you want to be able to make as many trade-offs here as you possibly can to make a business like this work. I mean, you know, it’s tempting to say that like the economic marvel that happened with generative AI can also happen here. And the kind of the the marvel that happened with generative AI is like you you train these you know these models with you know some distribution of data, and it turns out basically the same distribution is economically viable. Yeah. It’s like you train it on images, and people want images. You train it on text, and people want you know it’s like kind of distribution in distribution out. It’s not clear that this is actually the case with with these robots, right? You could train it on like you know a distribution of data, but then like actually what you need is just slightly out of that distribution. So like how I don’t know, you probably have the best visibility into how this scales and what the convergence is and whether you can actually remove tutors. Like, what is your intuition on on that? I think this is one of the big questions, and I think really it ultimately comes down to like what is the data and where you’re getting it from. Yeah. And the magic of the internet, right, is like if you are a company and you provide the application layer tools for software engineers to accelerate their software development, you are naturally in flow of all of this really, really, really economically valuable work, right? If you can make money as a business, then you’re probably also making data as a business, right? Those like economic incentives are aligned. And similarly for robots, if you’re like, okay, we’re just going to strap cameras on everybody’s heads and send them throughout Indonesia to collect data doing X, Y, and Z, you you have no guarantee, right? You you can guess, but there’s no you know invisible hand of the market that’s guiding you to ensure that that data is actually economically valuable. But if it’s actually this like.

[840s - 960s] Like kind of emission byproduct. If you have robots in real customer sites that customers are paying for for those robots to do that task, you know for sure that that data is economically valuable. Doesn’t it kind of depend on also the dimensionality of the solution set, though? Like it could be the case that like the economically valuable stuff is so high dimensional that you could like you could solve ninety percent of it, but the last ten percent, you know, you’ll you’ll need a human being to do for a while, just because like getting that data set is just going to require too many tutors over too period of time. Yeah, and that might be fine, right? You might be willing to accept that on the economics. That also might be a place where other forms of data matter. I think we’re not purists. Again, we’re willing to take any tradeoff, right? We love using simulation to kind of widen these data distributions. We love using like internet scale data to widen these distributions. So we have a lot of tools in our toolbox other than that real world robot data. And talk a little bit more about the simulations and these other tools and the impact models like Astra are having on that. Yeah, that seems to be another cottage industry of excitement for the robotics. Yeah, world, right? How are you using that? It’s it’s a sea change, right? I think we lived in this world for a while where everybody knew that all of this general knowledge about the world was baked into these models, but they couldn’t really express it, right? They didn’t have that last mile of spatial reasoning to connect the dots from the stuff that like you know a GPT model knows to something that an action diffusion transformer head can know, and that’s changed, right? The labs are now post-training their models on three D spatial data like CAD and like three modeling, which by the way they’re doing because there’s this direct economic incentive because there’s a real market for it, so they can immediately monetize it, and it turns out that’s the exact spatial reasoning that our models need. So I think we’ve been catapulted many years ahead in the future because we’re starting to be able to use.

[960s - 1080s] All of that general intelligence that lives in these big frontier models for robots, and because it’s a systems problem, the net effect for you is acceleratory. Yeah, as opposed to slowing you down or saying somebody’s going to come and eat your lunch. Yeah, no, we uh, and we’ve seen a couple of these happen, right? We started this business in 2021. We were doing this initially with convolutional neural networks, right? So like we’ve we’ve gone through every sea change, and and the thing that is always valuable is the flow of data, right? And the flow of data is is always the thing that uniquely you can do, and it’s like yeah, those like the million architectural tricks, like do them if it helps you accelerate the flywheel, right? If it helps you deploy more robots, collect more data, manufacture at a bigger scale, do more tasks, like they’re worth doing. Um, but but ultimately, we think the thing that has kind of the long-standing generational value in a business is the flywheel and loops themselves. Can Can you talk a little bit about what you mean when you say you build robots? Yeah, I think on one end of the continuum, it’s like yes, you buy a Unitree, yeah, you know, and you like prompt it, and the other end, you know, like whatever you’re like machining parts, like where are you in this? Yeah, so um, for context, for everybody, in 2025, the United States deployed uh 38,000 industrial robot workcells, which is tiny compared to roughly a million. You’re not counting the Fanuc’s and the old ones, the Fanucs, right? But but that’s it, right? Nobody nobody deployed any Unitrees in the United States into any. They would have shown up in those numbers, but there were probably zero. Um, and so the supply chain that exists for robots is both too small and, as you called out, wrong, right? Like the robots that Fanuc and ABB are building, they’re mostly not building the United States, but but even when they are building them in the United States, they’re building these objects that are designed for systems integrators that are designed for faster repetitive motions. They are not cut from the set of tradeoffs that you would want to co-design with frontier models, and so we’re kind of starting from scratch.

[1080s - 1200s] On the hardware side, so Tutor, in order to deploy robots at scale, has actually been forced to design and manufacture our own hardware. For us today, like we’re not building actuators, right? So we’re not building cameras. There’s a whole layer of stuff that we’re sourcing from OEMs, where they’re manufacturing at a much larger scale. But the actual robot itself, of you know, the thing that drives around and has the arm on it and can do all the tasks, we actually are building at scale in Watertown, Massachusetts. Can you talk about the morphology of the robot? Is that what is that called? Morphology, anatomy? You you can use whatever. What do they look like? What do they look like? We make two. So one robot we make. Look at this website, Martin. Yeah, go ahead. Yes. So Cassie, Cassie is kind of a brute, right? Cassie is designed to do heavy duty work. It has a single arm and a lot of suction cups, and it’s powered by foundation models in like the 200 to 400 million parameter range, and kind of came out of this world that Tutor was started out of in 2021, of like what can foundation models do for robots? They can predict kind of affordances, grasps that you can go then execute. So it’s this co-design of the hardware that you can manufacture at scale, what your models can actually do, and actually a big market need where in the United States anything in the physical economy has to be transported on a pallet or in a cardboard box. So there ends up being an enormous amount of physical human handwork of transporting pallets, cases by hand. And so Cassie is kind of the result of that co-design, and we manufacture it at very large scale, and we deploy it to a big diversity of businesses across manufacturing logistics. Our second robot is called Sonny, and Sonny is kind of an industrial semi-humanoid, right? It’s as if you said you were looking at a Unitree, and you were like, I need this robot to run for three to ten years in a factory at human rates and reliability, and so you end up with a slightly different set of design decisions.

[1200s - 1320s] Sonny is wheeled, not legged, because most indoor industrial settings have polished ground floors. It has a much higher payload capacity and can move much faster, and has industrial collaborative safety standards that enable it to operate around humans. And so, we have a hundred Sunnies operating in Boston in our data factory, collecting data for internal use and advancing our research agenda. And then we also now have a growing fleet of Sunnies out in the real world, doing more dexterous bi-manual manipulation. And do you see how broad do you see this product line of different robot form factors? Is it can you capture yeah a big chunk of the useful market with three of them or do you need thirty of them? Depends how you define the useful market. Like where do you want to go? If you want to just live inside of manufacturing logistics, Sonny and Cassie today can probably do eighty percent to ninety percent of the work. Right? These are designed for those cuts. What maybe just describe what that work is? Yes, because yeah, I’m not sure it’s probably understood. Yeah, so most people have this mental picture of what sort of the industrial base or manufacturing logistics look in the United States, which comes from the TV show How It’s Made. I know if either of you two have seen this, of course, of course, yeah. And you see these like Rube Goldbergy and machines of levers, and it’s like, yeah, you don’t see any humans in those pictures, right? It’s somebody built a machine that is the factory, and that does exist. And if it exists, they probably filmed it and put it on How It’s Made, right? Because that’s the top 0.01 percent of American manufacturing logistics. And so, what’s below that? And you walk into these million square foot facilities, and you see a bunch of people standing around on conveyor belts or folding tables doing things by hand, and it’s kind of shocking for anybody who doesn’t come from this industry and community. And so, you look at a robot like Cassie, and it’s like you know a huge amount of work, which is backbreaking work that’s really hard to hire for. There’s a three month turnover where your average

[1320s - 1440s] Age of the worker doing that work is fifty-five years old and about to retire, and it’s like, okay, we we could use a Cassie to do that backbreaking work. And similarly, Sonny is like basically anything you ever bought that came in a box, a human had to put that in a box, which is a crazy revelation for most people. Anything that requires any degree of assembly, right? Like even you know iPhones, you know, you would assume it’s a robot, you know, putting that packaging together. No, it’s a person, even at the highest level of scale. Then to his question, which is so like you have like the single arm with the sucker, you’ve got the two armed humanoid. Yeah. What can’t they do? Yeah. What can’t they do? Where do you go from there? Yeah. So they can’t go outside, and outside is really really interesting. Was that just like a locomotion issue? It’s not locomotion. It’s the hardware, right? These these robots have wheels, and they have like a little bit of traction. You know, they have like these articulating hind legs, but they can’t drive over gravel. Right? There’s huge opportunities of robots in construction, robots reclaiming materials from landfills. We you know we need robots with legs for that, so I think that’ll eventually be part of our pantheon of embodiments. And then similarly, robots in the home—you probably don’t want a Cassie or Sonny in your home. Just it’s a little bit too industrial of a device. Yeah, they look scary. Yeah. Let’s talk about the customer adoption side. How how has that been? Obviously, not strictly from a I don’t have people, so I’ll take whatever I can get. Yeah. To once they have it in their assembly line or whatever it is, how does adoption go in terms of their mindset? Yeah, it’s an interesting challenge because almost everybody we’re selling a robot to has never purchased a robot before, and that kind of comes out of these stats, right? If the United States is only deploying thirty-eight thousand robots a year, if nine out of ten U.S. factories and warehouses don’t have any robots, then almost all of the market is greenfield. And these people, you know, understand what this old guard of automation looks like, and they assume a robot doesn’t work for them. And so the way we break through that actually.

[1440s - 1560s] Isn’t by talking about AI or models or tutors or anything like that. We call them up and we’re saying, “Hey, have you ever considered a robot for sixteen bucks an hour?” And they haven’t, right? That that’s a new proposition for them, right? So you can book a meeting and then you can kind of explain to them this thing that we’re doing and get people interested and involved and kind of by creating that business model where there is that true incentive alignment between us as the robot vendor and the customer, we’re able to run distribution of robots at velocities that you know are many orders of magnitude faster than traditional industrial automation industry. We have an average time from that first call to a robot signing contract of forty to sixty days. We have an average time from that signed contract to robots actually in the customer site of about forty days doing work. We stand up robots in production within a day, often within the first couple hours that a robot is on site. So we’re able to move very, very, very quickly because not only is the robot this generally intelligent, generally capable machine that we’re not customizing, we’re just you know shipping directly from the factory to the customer, but we’re not putting any sort of economic risk on the customer, so we’re able to get them on board. And once they deployed, how do their attitudes about labor versus machines change? The operations community is really, really fantastic because they see a thing work and it’s like, “I, it works.” Right, like that’s it, right? It’s, it’s like, yeah, I, I can see obviously it does the job. I know all the places where this job needs to be done. Let’s go put it in all those places. So we’ve been able to scale up very, very quickly, both with small businesses and to some of the largest companies in the supply chain, from you know one robot in one site to you know a dozen plus robots across many sites. It is interesting though, because like like a lot of gen AI was like, there’s a social element where like people get very attached to their models and like.

[1560s - 1680s] It’s hard to swap models, and so, and listen. You probably have more experience deploying robots than most people alive at this point. Like, are you saying the same thing? Like, is it are they viewed like equipment, or like they starting to be viewed like beings and coworkers? Yeah. So I saw a very funny tweet, which was, “A robot is a robot unless it works, in which case it’s a machine.” I got to figure out the attribution. But is it real? I guess it. It’s it’s pretty real. Is that what it is? Like they just viewed as like a piece of equipment once it works. Certainly, at like the factory manager level, yes, I think there’s still the layer of like, interestingly, in robotics, your buyer is not your user, right? Your buyer is like the COO or the general manager, and the user is like a factory worker, right? Who’s working with the robot all day long, and there you actually do need to cultivate the relationship between those workers and the robots because, you know, we don’t fully automate 100% a factory. We we automate some jobs in a factory or warehouse. And do you have a philosophy in like how you even approach this? Like, do you like try to make them look fun? Do like make them look more like coworkers or less like coworkers? You make them talk like like how do you kind of view this philosophically? Yeah. So very very luckily there is a whole school of academic research called HRI human robot interaction. No idea. There you go. Yeah, a lot of people have spent a lot of time thinking about this, including a bunch of people from from my academic background, and there are a whole bunch of tricks. So one trick is every robot has a name, and so that lets us actually build relationship with the robots. It also mostly our manufacturing technicians. Okay, so so you name it, not the. It’s not like a dog. Is when you get a new dog, you can rename it. So when a customer expands with Tutor, when they when they do a repeat order, then they’re allowed to name the robots. Oh, you’re not supposed like rename a boat because that’s bad luck. You got to keep the boat name, but you can rename a dog. So like. Yeah, so the new robots get to be get to be named. That solves it, and we get customers. They get a higher sense of ownership now. They’ve named it exactly right, and we get like really strong signals. Like we had people name the robots after their kids.

[1680s - 1800s] Right, people name the robots after their favorite football player, their favorite pop star. People get very attached because the robot ultimately helps, right? It makes their jobs easier. If you’re a line worker in the facility, you’re not doing backbreaking work as much. You’re able to be elevated into a higher position, and none of these places can hire. And then, if you’re a factory manager, your job is just a thousand times easier. What naming scheme do you use? Yeah, from when you we started off doing hurricane naming, and then we kind of that’s not yeah. You didn’t read the HRI book, literally. Well, it’s I don’t know. It’s the customer doesn’t know that Brittany is you know NMA twenty six or whatever equals two. It’s um, it’s uh, but we actually stopped because we kind of went through the hurricane. We went through the alphabet too many times to be able to keep track of it. Okay, so what are you using now? Uh, I think the technicians just just pick whatever brings them joy. We we have a really wonderful and privileged to employ a big staff of people in the United States who are helping to build these robots, and they are also naming them. Awesome. The other thing that surprised me when we first did the work before investing most of the time when we talk to robotics companies, they’re going after these humongous companies, right? Yeah, the big Fortune fifty type companies that seem to have the budget for lots of experiments, et cetera. Yeah. Whereas I think a significant majority of your customers are all these local businesses, almost. Yeah. And what is that? I mean, is that a message in there, or was it a conscious strategy, or? Yeah, it’s a lot of things. One is that’s most of the market, right? The vast majority of U.S. businesses in manufacturing logistics are small businesses or mid-market businesses. So if you want to serve the market, that’s what it takes to serve them. I think we also started this company at a time we’re seeing kind of all of these very academic, supply chain oriented robotic startups kind of get started in.

[1800s - 1920s] And then not really succeed, and we we had the opportunity of being like, okay, what went wrong here? And one of the big things I think that robotics companies get wrong is they’re not actually really product companies. They’re like, okay, some really big company has a very specific problem. They’re willing to write us a shiny piece of paper that says they’re going to expand with it. We’re going to spend two years building a very custom system to solve their problem. We’ll deploy it. Chances are it won’t work because, like anybody who’s built products knows that you can’t just one and done a product. Right, doing product requires iteration. And then even if it does work, who’s to say that customer still has that problem? And so, boom, the company’s done. And I feel like I saw this happen like five to ten times in a row to the point of like, oh my god, if this happened to us, like that would just be embarrassing. And so, I think working with a diverse range of customers is this beautiful forcing function for building the right shape of product, which to us is a general robot, right? We don’t want to build a Rube Goldberg machine that works really well for big industrial customer A or B. We want to build a platform that is generally capable robot bodies running generally capable robot models that can immediately get to work. And if you can, you know, sign a contract in two weeks with a mid-market customer and have a robot running immediately after that, and that’s a totally different type of product and business than your other customers, you know, you’re being successful. How do you think, in the limit, what the industry looks like, assuming you continue to be successful? Is it just that you know this work gets cheaper faster, or does like somehow quality go up, or can they do more things, or like kind of like what is like the economic conversion to the productive conversion? Yeah, I think everything gets better. So there’s a whole set of goods that if they are cheaper, we will buy more of them, and if you believe those goods are valuable, that’s good for everybody. There are also a bunch of discontinuities in the world. Like I was talking with somebody the other day who worked on the line in a housing factory.

[1920s - 2040s] in I think Wisconsin, and they were describing how the houses that they’re building are about five percent more expensive to build them in a factory than to do traditional construction on site. But if these houses were even just one percent cheaper than kind of the equivalent quality of house manufactured or constructed on site in the traditional way, all of the demand from developers would immediately go to that other shape of building these homes in a factory. And the same thing can be said about like reclaiming materials and landfills and a whole bunch of other markets. So I think our hope is like if Tutor is really successful, we can provide this robot substrate, this this labor force that other businesses can use to transform the physical economy and make the world ultimately a cheaper and better place to live in. Will you ever go in the home? Yeah, we’re we’re not opinionated. We’ll go anywhere. But I would have betted that would be kind of a much tougher technical problem, just from like I don’t know complexity of task and safety and and the economics of it, right? Like how many? What is the economic value Tutor robot delivers in a factory, and how many consumers are willing to pay for that in their home? And it doesn’t quite work out. But again, that that comes back to our our overall philosophy is take trade offs across as many layers of the stack as possible to get a foothold. Start building the platform, and as you build the platform, you have to take fewer and fewer trade offs over time. How does product innovation work in robotics? I mean, most companies are just on V one, whereas you’ve got two versions of the robot and you continue to iterate really fast. Well, how what is it different? How is it different than like digital product innovation or other categories? We’re trying to make it as similar as possible. So we’re manufacturing our own robots. This gives us the ability to own that manufacturing process and figure out how do we push changes to the manufacturing process every two weeks. We are shipping code to robots either every day or.

[2040s - 2160s] Major releases every two weeks, and it built a huge amount of infrastructure to support that level of iteration. We’ve built a think a really unique shape of product function where our product function in our organization actually sits right next to our customer success and implementation functions, and operate actually as one big team. So they’re as close to the customer as humanly possible, which I think has been really successful for us. And then we also have a research team which is able to be both, on the one hand, deeply embedded in kind of the customer problem space, and we’re able to pull all sorts of levers to get specific types of data from customers, but also able to take these moonshots. And you know, if we have a breakthrough on our modeling side that lets robots do a whole new shape of thing, we can kind of co-design a product around that in a way that I think most companies couldn’t and get that out into the field much quicker. If I if I’m listening to this and I have a warehouse, yes, and I want to get a robot, do I just go to the website and like put in my email? That’s how does this work? You can go to our website. You can put an email. We actually have cool pricing calculators where you can actually put in some operational data about how your warehouse or factory runs today, and you can get an estimate of hey, with an initial deployment of two to robots, we could be saving a hundred thousand dollars a year, a million dollars a year, a hundred million dollars a year, and that’s not like okay, this is a roadmap over five years of custom engineering. It’s like no, you could save that a million dollars a year two months from today without putting capex down. And have you found that there is like demand for? I mean, it’s just such a sci-fi thing. Is there demand out there? There is so much demand, and it’s not even demand for robots because again, like people don’t know that the shape of robot exists or that it works or that it can do this job. So it’s it’s ultimately like you know we we have this framework of the value of a robot is you take how much work a robot can do and you divide it by its cost, and that’s it. Right, everything else is is commentary, and so.

[2160s - 2280s] So if we’re able to compete in these highly liquid labor markets that exist across American physical economy, and we can offer something that delivers greater equivalent capability at lower cost, the market for that is the whole market, right? It’s all physical work done. And what? Just one quick, wonky question on this, which is: How much when you say cost do you mean amortizing the upfront cost versus like the ongoing cost to operate it? It’s both, which is very interesting and different, right? Traditionally, we’ve thought about robotics as ROI, right? Big capex expense, and then ten years to make it back, and now it’s actually like no, more and more of the cost is going into ongoing inference teleoperation. But does that also include like maintenance and break-fix and all of that stuff? And then what? And like, what is the dominant cost, or are they roughly equal? We drive down all of them over time. No, no, I’m just saying, like from from your standpoint, when you like pencil out your Cogs, how much of is it the upfront cost versus the ongoing? Yeah, it is more ongoing at this point, and that’s surprising to people. Interesting. Oh, why? Yeah, not surprising to me. I mean, you’re really competing with like something that you feed rice to, and it’s very, very efficient, right? Like the the human body. Yeah. So, the when you get called the customers, is it because Black Friday or something’s coming up and they need a bunch of units done in a really short time, or is this actually now part of their actual labor planning cycle? No, it’s strategic, right? It’s these businesses I think are faced with a really hard choice, which is they know the age of their workforce, they know that demand is ramping up due to all these secular trends, and the cost basis of fulfilling that work is going up, and they’re like, how does our business pencil out? Because you know these are not software businesses that have these you know huge gross margins that you can you can play with and can compress a little bit, and it’s like it’s fine. It’s like no, the business either works or it doesn’t, and so I think.

[2280s - 2400s] Our customers are ultimately looking for a long-term partner that can do the robots, right? If one of our customers is manufacturing deodorant, they’re doing the deodorant, right? And they’re going to be operational experts in deodorant or operational experts in crayons or whatever the thing they manufacture is, and they need a partner who’s the expert on robots. I had a question on the nature of factories, and I mean, you can ship a robot to anybody anytime, not today because of your shortages, but at some point in the future, right? And if your robots become more and more general purpose over time, what does it mean for where factories are located and how people organize around these factories and and you know these one town factories and what one factory town, whatever the expression is, yeah, all of that fundamentally changes now, right? Yeah, I think a lot changes, right? There’s a lot of geopolitics in this. I think one big thing is this is going to be a great equalizer between the United States and other company or excuse me, other countries like China invested in the previous regime of industrial automation, and in some ways that made them very successful, and in other ways, I think in hindsight that might end up being an enormous mistake because in the United States we just discussed we’re doing this greenfield builds of automating our industrial capacity using generally capable robots. In China, they’re going to have to tear down, right? We’re not going to have to do that in the United States in the same way. I think another thing that’s going to shift, which is going to be really interesting, is we’re just going to have way more factories in the same way that we have way more businesses because kind of the cloud and all of the software tooling let people build new businesses. Anybody will be able to start a business in the physical world. Anybody will be able to scale a business in the physical world if we kind of have this infinite flexible labor substrate. You you brought up China, our you know, you know we actually we actually published a report on the supply chain and robotics, and it seemed to be that that China was so central to it, yeah, so.

[2400s - 2520s] You talk a little bit about your dependency on it, and then how that evolves over time, and how much of there is there risk in China to the supply chain. Yeah, I think ultimately, no matter what anybody says, today in 2026, we have a highly interconnected, interdependent global supply chain, and this is one of many examples of this. And I think both U.S. companies and Chinese companies are thinking very hard about kind of what does the long-term strategy of this look like. And it’s Chinese companies investing in the United States. It’s United States companies investing China, and ultimately, everybody trying to build their industrial capacity. Yeah, I’ve I’ve noticed that you know you’ve gotten as far as I can tell, you’ve verticalized more. Is that right? Like you’re taking more in-house, and do you continue? Do you plan to continue to do so? Yeah, well, there’s this tough thing in supply chain where, and again, because robotics is such a cottage industry, is you might start in a world where subcomponents are coming from OEMs, and then you grow your volume to the point of they’re not OEMs anymore. They’re they’re glorified contract manufacturers, right? Ones are doing a material percentage of their volume, and at that point, just the relationship and the dynamics change, right? It’s like is that the right command for that thing? Do I want to be using a command, or do I want to be manufacturing it in-house? Yeah, I mean, Tutor must be one of the largest employers of robots in the United States. You know, as a result, you must be like one of the largest customers for some of these OEMs. You know, as a result, I would imagine over time, like you just continue to like drive, you know, these components yourself. I think that’s does that mean taking it in-house or just telling them what to do? I mean, is it just like are they still suppliers? Some combination, I guess. Some combination. I think ultimately, for Tutor to be successful, we will need a mature robotic supply chain in the United States, and we’re in this position where because we are so much of the demand for that supply chain, I think it’s both our obligation and an opportunity to invest very, very, very heavily to building this capacity in the United States, and that’s why we build robots in Watertown, Massachusetts. It’s why we expect to continue to verticalize and build.

[2520s - 2640s] More and more of our robot componentry in the United States. Again, going back to when we first started talking, I remember it was probably the second meeting when we actually started about the research. Right, most of the meeting because we had such a great business, but that’s a big strength of yours, right? And talk about how you approach it, and there’s all sorts of theories about which robotic learning models will win and so on, so forth. How How are you approaching the robot learning bed? Yeah, in research, I think we’re thinking about how this played out in LLMs and what’s similar and what’s different. And I think one thing that clearly has been very, very important in LLMs is there are lots of heterogeneous sources of data, but ultimately there’s a top of the data pyramid, which is real-world reinforcement learning traces doing the thing you want to be able to do. And so we’ve built our research program in such a way where we are thinking about how do we get the most out of robot experience out in the real world. And so we built this 100 robot fleet called Data Factory One in Boston. This is, to our knowledge, the largest sort of semi-humanoid data collection farm in the United States. And we’re able to use this as kind of this research test bed of how do we do real-world reinforcement learning at scale with robots. You know, reinforcement learning is a technique family where it doesn’t really work at small batch sizes. And so if you have a very large fleet out in the real world, you can start to play with these methods, which I think people haven’t been able to play with either in academia or industry. So I presume, like building a team for this is quite different, right? Than like a you know, Labs. Maybe talk about kind of like the construction of the team and maybe if there’s anybody you’re looking to to hire. We’re looking to hire everybody. Everybody. So we’re looking to hire researchers and engineers, and that’s across every single layer of the stack. Wow. We are also looking to hire technicians to build robots. By the way, every layer of the stack. You mean like software, hardware, materials, electronics, like.

[2640s - 2697s] There, yeah, there’s nowhere that you can abdicate responsibility, right? You’re gonna engage with it whether you want to or not. The question is whether you understand it. Um, so so we’re we’re building that team, and I think that’s a really unique shape of team, not only because it’s so interdisciplinary and collaborative across all those layers of the technical stack, but because I think it’s one of the only places in the universe where you can be on a whiteboard talking about model architecture in the morning, in the afternoon, like you know, looking at a factory line and seeing how a robot is doing. And is this primarily in Massachusetts, or so we have a team across the United States. Um, we have especially folks on the go to market and technicians who are kind of engaging, you know, our customers are everywhere across the United States. All right, awesome. Thank you.

中文翻译（全量）

[0s - 120s] 无论谁怎么说，2026 年的今天，我们都拥有一个高度互联、相互依存的全球供应链。美国公司和中国公司都在认真思考这一切的长期战略会是什么样，而归根结底，所有人都在努力建设自己的工业产能。我们看过大量机器人融资路演。很少有人思考这一切的尽头，机器人到底应该干什么。我们卖机器人的对象里，几乎每个人都从没买过机器人。几乎整个市场都是绿色地带。现实是，你需要打造一个全栈平台。Tutor 的机器人已经帮助制造并向美国消费者发货超过两亿五千万件产品。我们正在训练前沿机器人 AI 模型、规模化制造机器人，并直接部署进供应链里最大的那些企业。机器人行业的产品创新是怎么运作的？大多数公司还停在 v1，而你有两个版本的机器人还在持续快速迭代。那么 Josh，感谢你来参加节目。今天和我们在一起的是 Josh Gruenstein，Tutor Intelligence 的 CEO、联合创始人，今天我们要聊的是机器人。那么在正式开始之前，先简单介绍一下你自己，然后聊聊公司。好，我是 Josh。大约五年半前我创立了 Tutor，核心论点是：只要你能开始把机器人部署到世界上、部署到实体经济中最需要机器人的那些地方，你就能开始积累让机器人在整个实体经济中铺开所必需的数据集、模型和供应链。我们现在大约 160 人，主要在波士顿，我们在训练前沿机器人 AI 模型、规模化制造机器人，并直接部署到供应链中最大的企业。也许可以再谈谈：我们看过很多路演，有很多公司说自己做机器人，但实际上并不部署机器人，或者还没部署过。你是少数真正规模化部署了机器人的公司之一。那就讲讲业务现状吧。好，我们的机器人部署在全美各地。

[120s - 240s] 从小型夫妻店到财富五十强、财富一百强企业都有。昨晚 Claude 告诉我，它估算 Tutor 的机器人已经帮助制造并向美国消费者发货超过两亿五千万件产品。所以大概可以推想，大多数美国消费者往下数两层就是 Tutor 的客户。我认为这源于公司的起源故事：我们审视机器人学习的现状，当时对我们来说很显然，机器人智能、机器人软件就是那个瓶颈，对吧？六十年来，机器人一直是这种无法规模化的定制工程机器。是的，所以我们带着这样一个论点创立了公司：只要我们能开始部署机器人，我们就能建立起带标注的数据集，帮助训练这些模型，把这份价值交付给美国消费者。我们看过大量机器人路演，其中 99% 都从某种关于模型的洞见出发——你要怎么构建模型，或者对数据形状的某种洞见——却很少有人思考这一切的尽头机器人到底该做什么。你是反着来的，而且你有一个洞见：你其实可以「辅导」（tutor）机器人，这也是你公司的名字。这个洞见从哪来？它来自一个思想实验。我们意识到：地球上有八十亿人，他们都拥有我们希望机器人具备的那种基础物理智能。那么想象一下，你有一个无限强大的软件，能把地球上任何一个人连接到任何一台机器人上，让那个人把智能赋予这台机器人。这有点像机械降神（Deus Ex Machina）——突然间你就拥有了无限的机器人智能。假设这一点是白给你的，然后呢？对吧。问题其实并没有解决。你仍然需要真正地规模化制造机器人。而在今天的美国，机器人还是个家庭作坊式产业。

[240s - 360s] 即使在制造、物流这些自动化程度最高的行业里，这也只是个舍入误差。所以这些都不存在。软件当然不存在，客户同样不存在——因为机器人干不了客户想让机器人干的任何活。所以我认为，机器人不是技术栈里的某一层。它不只是智能，不只是硬件，也不只是客户，而是要想办法把这三块拼到一起；在我看来，正是三者的组合，才能让你在其中任何一层上建起成功的平台。是的，我们内部常说，机器人是「终极的系统公司」，而不是只做其中某一层的公司。不过有意思的是——这也是我一直想问你的——你有很深的技术背景，MIT 机器人学，这么说对吧？但每次跟你聊天，我都感觉是在跟一个搞运营的人聊：你们聊单位经济、运营和供应链，甚至还聊 go-to-market。所以可以说你已经完成了这种转变。那么这是个开放问题：这里更难的难题，会不会其实是你说的那些东西？不，但你必须全都做，这才是关键。在我还是个机器人研究者、只做这个问题技术侧的时候，我认为我作为机器人学者的优势恰恰在于看问题非常全栈：我愿意去想制造、想硬件，愿意去想 FPGA 和 ASIC。如果你像我一样把机器人看成终极的系统问题，那前进的唯一方式，就是在尽可能多的栈层之间找到聪明的取舍。而如果你扔掉其中一层、说「这个我们不碰」，你就消灭了 90% 用来取得进展的取舍空间。

[360s - 480s] 有意思。全栈不只是模型、硬件、软件，而是模型、硬件、软件、供应链、客户、产品，对吧？正是如此。同理你问：为什么机器人行业一直没成？出现过多得可以的机器人创业公司，但我认为没有一家取得过真正巨大的经济成功。为什么？我认为一是你缺这个核心原语——智能，它本可以让你把系统的单位经济拉到某个水平；二是这是一个高度分散的市场，玩家无穷多，而且他们全都在试图互相合作，而机器人行业真正需要的是能够跨全栈做取舍。有意思——这也是我们的一个观察，这么表述出来很有意思：大多数机器人公司从水平切入，最后全都走向垂直，无一例外——他们一开始说「我们有个模型」之类的，然后你猜怎么着，他们就开始买 3D 打印机了。你觉得这是因为市场太不成熟，就像早期造车那样不得不做，还是说这其实是这个领域的固有属性——哪怕二十年后你仍然必须垂直化，就因为移动的经济性之类的什么原因？我认为是市场不成熟，你在早期计算机市场里也能看到类似的现象。于是问题就变成：怎么建出那种通用平台，让人能在上面建起「不是机器人业务」的企业？今天我们有这么多出色的软件公司，而经营者根本不用去想芯片架构。怎么在物理世界里用机器人做到这一点？怎么着手去建这样一个平台？我同意你的看法。人们一直试图通过只碰技术栈里很窄的层来解决这个问题，而在我们看来，现实是你需要建一个全栈平台；而确认这个全栈平台真正有效的唯一方式，就是让它上面真的跑着成功的业务单元。

[480s - 600s] 我的意思是，你能不能接受这样一种可能：物理世界跟数字世界就是不一样，我们之所以能把计算水平化，是因为存在可以普遍适用的干净抽象，而对于在物理空间里导航这类事，情况并非如此。我认为人形机器人之美恰恰在于：它是一样的，它就是那个干净的抽象。你想想 CPU 是什么——它是通用计算机器，对每个用例都不是最优的，但足够通用，能跑绝大多数计算负载。哦，所以人形就是你的水平组件。但比如说，造人形所用的同一个大脑或子组件，会跟造无人机或其他形态因子所用的一样吗？不一样，但你在计算领域也能做同样的类比，对吧？我们有 CPU、GPU，还有 ASIC。所以我认为你可以跟踪市场的成熟度：你从专用、定制机器的世界起步，因为那是市场唯一撑得起的东西；然后走向通用目的，在机器人领域，这就会长成人形和类似的形态；最终当市场足够大，你会突破到再专用化。这个东西为什么叫 Tutor？所以 Tutor 的创立带着这样一个论点：存在一整个人类智能的互联网，它是联网的——在社交媒体上发帖、写邮件——而这个人类智能网络以自然排放的方式，为 LLM 和其他数字世界模型造出了美丽的数据集，让我们能用来训练通用模型。问题是我们物理世界里的工作并没有以同样的方式联网。比如我端起我的咖啡杯，没人知道，对吧？没人记录过我指尖的全部压力、它们所在的位置和姿态，所以我们没有这份资产。Tutor 背后的核心想法是：与其试图找到某种算法、从零数据出发就能凭空造出智能——我们觉得那是一场对我们不利的、被操纵的比赛，计算机视觉不是这么赢的，NLP 也不是——不如问……

[600s - 720s] ……我们能不能在真实世界里建起一个机器人网络，用人类遥操作者——我们称之为「导师」（tutors）——把机器人连接起来，教它们做新任务；就像互联网以自然排放的方式造出了训练这些前沿模型的数据集，我们也能用机器人做同样的事，建起这个机器人训练数据的互联网。而这有一系列美妙的涌现属性。我们会发现：突然间这不再是一个巨型资本项目；同时你在做的过程中还在创造巨大的经济价值——你在部署这些机器人、收集这些数据的同时，让所有人的商品变得更便宜。所以我认为它在一种非常独特的意义上是正和的。于是当我们思考该给这门生意起什么名字的时候——它不是严格意义上的人工智能；里面有一些人工智能的成分，但归根结底，我们造的是这种连接并控制机器人的「导师智能」。主持人：这是个了不起的洞见。有意思的是，我们投资前跟你的客户聊过，大多数客户好像不知道、甚至不太在乎有导师存在，也不知道什么时候是导师在操作、什么时候是你的学习模型在操作——这也归功于你们把两者做得非常非常无缝。 Gruenstein：在我看来，这正是机器人足够好的一个信号。你想要的反馈是「我以规模生产的洗发水，现在比原来便宜了五分钱」。如果反馈是「这机器人挺好看的」——那大概是一台很差的机器人。主持人：目标是不是未来不再需要导师，还是说这无所谓？Gruenstein：我们在非常务实的层面看这件事：导师的运营复杂度是多少？导师的经济影响是多少？想想用 H200 提供模型服务的成本是多少，你可以把它算出来；H200 的占空比和利用率是多少？对人类智能，你也可以用非常相似的方式去思考……

[720s - 840s] ……你真正想要、而你的模型暂时还给不了的那种异构智能画像是什么？于是这就变成一个单位经济学问题：为了从你的模型里抠出最后那部分成色，你要做多少工程，对比直接用导师？而且——人们很容易想说，发生在生成式 AI 身上的经济奇迹也能发生在这里。生成式 AI 的奇迹是：你用某个分布的数据训练这些模型，结果基本上同一个分布就是经济上可行的。就像你用图像训练，人们就想要图像；你用文本训练，人们就想要文本——分布进，分布出。而对机器人，这一点是否成立并不清楚：你可以在某个数据分布上训练，但你真正需要的东西可能刚好在那个分布之外一点点。所以——我不知道，你大概对这件事怎么扩展、会不会收敛、能不能真的去掉导师有最好的能见度。你的直觉是什么？Gruenstein：我认为这是最大的问题之一，而它最终归结为：数据是什么、你从哪里得到它。对吧。互联网的魔力在于：如果你是一家公司，你为软件工程师提供加速软件开发的应用层工具，你就天然处在所有这些极其、极其有经济价值的工作的流动之中，对吧？如果你作为一门生意能赚钱，那你多半也在作为一门生意产出数据——那些经济激励是对齐的。对机器人也一样：如果你说「好，我们给所有人头上绑好相机，派他们跑遍印尼去采集做 X、Y、Z 的数据」，你没有任何保证，对吧？你可以猜，但不存在一只看不见的市场之手来引导你，确保那些数据真的具有经济价值。但如果它是这样一种……

[840s - 960s] ……排放副产物：你的机器人在真实的客户现场，客户正在为机器人干那些活付费——那你就确知这些数据具有经济价值。主持人：但这不也取决于解集合的维度吗？完全可能有这种情况：经济上值钱的部分维度太高，你能解掉其中的九成，但最后那一成，在一段时间里就是需要人来干，因为要采到那个数据集需要太多导师、太长时间。Gruenstein：是的，而那可能是没关系的——你也许愿意在经济上接受它。那也可能是其他形态的数据发挥作用的 places。我认为我们不是纯粹主义者。再说一遍，我们愿意接受任何取舍，对吧？我们热爱用模拟来拓宽这些数据分布，也热爱用互联网级的数据来拓宽这些分布。所以除了真实世界的机器人数据，我们工具箱里还有很多很多工具。主持人：多讲讲模拟和其他工具，以及 Astra 这类模型正在对它产生的影响。那似乎是机器人世界里另一个让人兴奋的家庭作坊。你们怎么用它？Gruenstein：这是一场海变。有段时间我们活在这样一个世界里：人人都知道这些模型里烤着关于这个世界的大量通用知识，但它们没法真正表达出来，对吧？它们缺了空间推理的最后一公里，没法把一个 GPT 模型知道的东西和一个动作扩散 transformer 头能知道的东西连起来。而现在这变了：各家实验室正在用 CAD、三维建模这类三维空间数据给他们的模型做后训练——顺便说一句，他们这么做是因为存在直接的经济激励，因为那里有一个真实的市场，他们可以立刻变现——而那恰好就是我们的模型需要的空间推理。所以我认为我们被向前弹射了很多年，因为我们开始能够把大型前沿模型里的所有这些通用智能用到机器人上。

[960s - 1080s] 主持人：而且因为这是一个系统问题，对你的净效应是加速性的。而不是拖慢你，或者说什么人会来吃掉你的午餐。Gruenstein：对，这种事我们已经见过几次了。我们 2021 年开始做这门生意，最初用的是卷积神经网络，对吧？所以我们经历了每一次海变，而永远有价值的东西是数据流。数据流永远是那个只有你能做的事；至于那上百万种架构技巧——如果它们能帮你加速飞轮，就去做，对吧？如果有助于你部署更多机器人、收集更多数据、更大规模地制造、做更多任务，那它们就值得做。但归根结底，我们认为一门生意里真正具有长期、跨世代价值的，是飞轮和循环本身。主持人：能不能讲讲你说你们「造机器人」是什么意思？这个连续谱的一端是：对，你买一台 Unitree，然后给它写提示词；另一端是你自己在机加工零件。你们在哪？Gruenstein：先给大家一个背景：2025 年，美国部署了三万八千个工业机器人工作单元，跟大约一百万相比微不足道——这还没算那些老的 Fanuc。但事实就是这样：没有人把任何 Unitree 部署进美国的任何生产环境。如果有的话，它们会体现在这些数字里，但大概是零。所以，现存的机器人供应链既太小、又如你所说「不对」，对吧？Fanuc、ABB 造的那些机器人，大部分不是在美国造的；而就算它们在美国造，它们造的也是为系统集成商设计的物件、为更快的重复动作设计的——不是那种你会拿来与前沿模型协同设计的取舍组合。所以我们在硬件方面，几乎是从零开始。

[1080s - 1200s] 为了规模化部署机器人，Tutor 实际上被迫自己设计和制造我们的硬件。对我们今天来说，我们不造执行器，对吧？我们也不造相机。有一整层东西我们从 OEM 采购，他们的制造规模比我们大得多。但机器人本体——就是那个会四处跑、装着机械臂、能干所有任务的机器——我们确实在马萨诸塞州 Watertown 以规模化生产。主持人：讲讲机器人的形态？是叫 morphology 吗？形态学、构造学？你想用哪个词都行。它们长什么样？我们造两款。一款——看下网站吧，Martin。好，请讲。Cassie 有点像一头蛮牛：它是为重活设计的，单臂、大量吸盘，由大约两亿到四亿参数量级的基础模型驱动，直接出自 Tutor 2021 年起家的那个世界观——基础模型能为机器人做什么？它们能预测可供性和抓取，然后你去执行。所以 Cassie 是一种协同设计：你能规模化制造的硬件、你的模型真正能做的事情、加上一个真实的大市场需求——在美国的实体经济里，任何东西都必须装在托盘上或纸箱里运输，于是存在巨量的、徒手搬运托盘和箱子的活。Cassie 就是这个协同设计的产物，我们以非常大的规模制造它，并部署到制造、物流里极其多样的企业。我们的第二款机器人叫 Sonny，它是一种工业半人形：就好像你看着一台 Unitree 说「我需要这台机器人在工厂里以人类的速率和可靠性运行三到十年」，于是你会得出一组略微不同的设计决策。

[1200s - 1320s] Sonny 是轮式的、不是腿式的，因为大多数室内工业环境的地面是抛光的；它的载荷能力高得多、移动速度快得多，并且满足让它能在人周围作业的工业协作安全标准。我们有一百台 Sonny 在波士顿的「数据工厂」里运行，为内部用途采集数据、推进我们的研究议程。同时，我们现在也有一支在真实世界里增长的 Sonny 机队，做更灵巧的双手操作。主持人：你们看到这条不同形态的产品线会铺多宽吗？三款就能吃下有用市场的一大块，还是需要三十款？Gruenstein：取决于你怎么定义有用市场。你想去哪里？如果你只想待在制造和物流里面，Sonny 和 Cassie 今天大概能干其中 80% 到 90% 的活。这些机型就是为这些场景切的。主持人：那先描述一下这些活是什么？因为我觉得它可能没有被充分理解。Gruenstein：大多数人对美国工业基础或制造物流的想象，来自电视节目《How It’s Made》。你们俩肯定都看过。你会看到那些鲁布·戈德堡式的连杆机器，画面里看不到任何人，对吧？有人造了一台「机器即工厂」的东西，而那台机器就是工厂。这种东西确实存在，而且如果它存在，他们多半已经拍下来放进《How It’s Made》了，因为那是美国制造物流的前万分之零点一。那剩下的呢？你走进那些百万平方英尺的设施，会看到一大群人站在传送带或折叠桌旁徒手干活——对任何一个不来自这个行业和这个社群的人来说，这都是相当震撼的。于是你看 Cassie 这样的机器人：巨量的工作是那种重体力活，招人非常难，三个月的流失率，干这个活的平均年龄五十五岁、马上要退休——好，那我们就可以用一台 Cassie 去干那种重体力活。Sonny 同理：基本上任何你买过的、装在盒子里的东西，都是某个人把它装进盒子的——对大多数人来说这是个惊人的启示。任何需要某种程度装配的东西都是这样。连 iPhone——你会以为包装是机器人装的——不，是人，哪怕在最高的规模化程度上也是人。

[1320s - 1440s] 主持人：回到他的问题：你有带吸盘的单臂，有双臂的人形——它们干不了什么？下一步往哪走？Gruenstein：它们出不了门，而「门外」非常非常有趣。主持人：那只是移动能力的问题吗？Gruenstein：不是移动能力，是硬件，对吧？这些机器人有轮子，有一点抓地力；它们有那种铰接式的后腿，但开不过碎石路。而机器人在建筑业、在垃圾填埋场回收物料方面有巨大的机会——那类事情我们需要腿式机器人，所以我认为那最终也会成为我们具身形态万神殿的一部分。同样还有家庭里的机器人——你大概率不想在家里放一台 Cassie 或 Sonny，这设备对家庭来说太工业了。主持人：对，它们看着吓人。聊聊客户采用这一侧吧。情况怎么样？显然不是「我没有人手，所以我有什么要什么」那种。到他们真的把机器人接进装配线或其他什么之后，采用情况在心态上如何？Gruenstein：这是个有意思的挑战，因为我们卖机器人的对象，几乎每一个人都从没买过机器人，而这正好来自那些统计：如果美国一年只部署三万八千台机器人，如果十分之九的美国工厂和仓库没有任何机器人，那几乎整个市场都是绿色地带。这些人了解老一代自动化长什么样，他们假设机器人对自己不适用。而我们打破这个假设的方式，其实……

[1440s - 1560s] ……不是靠聊 AI、模型、导师或任何这类东西。我们打电话过去说：「嘿，你考虑过 16 美元一小时的机器人吗？」他们没有考虑过，对吧？这对他们是一个全新的命题。于是你能约到会议，然后可以向他们解释我们在做的这件事，把人吸引进来、参与进来。而通过创造这种商业模式——在机器人厂商和客户之间存在真正的激励对齐——我们能够以比传统工业自动化行业快多个数量级的速度推进机器人的分销。从第一通电话到签署合同，平均 40 到 60 天；从签约到机器人真正进入客户现场干活，平均约 40 天；机器人到场后一天内就在生产中站起来了，通常是在到场后的头几个小时。我们能跑得非常非常快，不仅因为这台机器人是通用智能、通用能力的机器——我们不做定制，直接从工厂发货到客户——也因为我们不给客户压任何经济风险，所以能让他们顺利上车。主持人：部署之后，他们对「劳动力 vs 机器」的态度变了吗？Gruenstein：运营这个群体非常非常棒，因为他们看到一个东西能干活，然后就是：「行了，它能干。」对吧？就是这样。「我能看出它显然干得了这个活。我知道所有需要干这个活的地方——走，全都装上。」所以我们扩张得非常非常快，既跟小企业，也跟供应链里一些最大的公司：从一个站点的一台机器人，到多个站点的十几台以上。主持人：有意思的是，生成式 AI 里有一种社会性元素——人们对自己的模型非常依恋，换模型很难。你现在大概比地球上活着的大多数人都有更多部署机器人的经验——你看到的是同样的现象吗？它们被当成设备看待，还是开始被当成生命和同事？

[1560s - 1680s] Gruenstein：我看过一条很好笑的推文：「机器人就是机器人，除非它能干活——一旦它能干活，它就是一台机器。」主持人：出处我还得查一下。但这是真的吗？我猜是真的。它就是这样吗？一旦能干，就被纯粹当成一件设备？Gruenstein：至少在工厂经理这个层面，肯定是。但有意思的是，在机器人行业，你的买家不是你的用户，对吧？你的买家是 COO 或总经理，而用户是工厂工人，是那个每天和机器人一起工作的人。而在那里，你确实需要经营那些工人与机器人之间的关系，因为我们并不把一座工厂 100% 全自动化，我们自动化的是一座工厂或仓库里的一部分岗位。主持人：你对这件事有方法论吗？你试图把它们做得有趣？让它们更像同事还是更不像同事？让它们说话？你从哲学上怎么处理？Gruenstein：非常幸运的是，有一整个学术研究流派叫 HRI，人机交互。主持人：没听说过。Gruenstein：给你补上了。很多人花了大量时间思考这个问题，包括不少来自我学术背景的人，而且有一整套技巧。一个技巧是：每台机器人都有名字，这让我们能够真正与机器人建立关系。主持人：所以是你们命名，不是——这不像狗：你领一只新狗回家，可以给它改名。Gruenstein：所以当客户与 Tutor 扩大合作、下重复订单时，他们就获得给机器人命名的权利。主持人：哦，你不应该给船改名，那不吉利；你得保留船名，但你可以给狗改名。Gruenstein：对，所以新机器人可以被命名。这就解决了问题，而且我们由此看到客户有了更强的拥有感——他们给它命了名——我们收到非常强烈的信号：有人用自己孩子的名字给机器人命名。

[1680s - 1800s] 有人用最喜欢的足球运动员、最喜欢的流行歌星命名。人们会非常依恋，因为机器人归根结底是在帮忙，对吧？它让每个人的工作都更轻松。如果你是设施里的产线工人，你就不用再干那么多重体力活，你能被提升到更高的岗位——何况这些地方一个都招不到人。而如果你是工厂经理，你的日子就好过一千倍。主持人：你们用什么命名体系？Gruenstein：我们一开始用飓风的命名表。主持人：你们真是没读 HRI 的书啊。Gruenstein：怎么说呢——客户又不知道 Brittany 其实是 NMA-26 之类的东西。不过后来我们停了，因为飓风表绕了太多圈，记不住了。主持人：那现在用什么？Gruenstein：我觉得现在就是技师们想叫什么叫什么、开心就好。我们非常幸运，也很荣幸能在美国雇用一大批帮我们造这些机器人的人，命名的事也归他们。主持人：太好了。另一件我们在投资前做功课时感到惊讶的事：大多数时候，我们聊的机器人公司追的都是那些庞然大物，对吧？那些财富五十强级别、看起来有预算做各种实验的大公司。而你们相当多数的客户，几乎都是本地企业。这是怎么回事？这里面有什么信息，还是有意识的战略？Gruenstein：这有很多原因。第一，那就是市场的大多数：美国制造物流里的绝大多数企业是中小企业或中型市场企业。所以你要服务这个市场，就得这么干。我认为还有一点：我们创业的时期，正看到这些非常学院派的、围绕供应链做定制的机器人创业公司纷纷起步——

[1800s - 1920s] 然后并没有真正成功，而我们有机会复盘：好，这里哪里出了错？我认为机器人公司犯的最大错误之一，是他们其实不是真正的产品公司。他们的剧本是：某个非常大的公司有一个非常具体的问题，他们愿意给我们开一张闪亮的纸，说他们以后会扩张采购；我们花两年时间造一套高度定制的系统去解决他们的问题；我们把它部署上去。而多半做不成——因为任何做过产品的人都知道，产品不可能一次成型，做产品需要迭代。而就算做成了，谁敢说那个客户还有那个问题？然后，砰，公司就没了。这种事我眼看连续发生了五到十次，看到后来我想：天哪，这要是发生在我们身上，那就太难堪了。所以我认为，与多样的客户群合作是一个美丽的强制函数，逼你做出正确形状的产品——对我们来说，就是通用机器人。我们不想造一台对大工业客户 A 或 B 特别好用的鲁布·戈德堡机器；我们要造一个平台：通用能力的机器人本体，跑着通用能力的机器人模型，到场立刻就能干活。而如果你能跟一家中型客户两周签约、机器人随后立刻跑起来，并且这与你其他客户用的是完全同一类型的产品和生意——那你就是做成了。主持人：你们持续成功的话，这个产业在极限状态下会是什么样？是同样的活变得更便宜、更快，还是质量上升，还是它们能干更多的事？经济上的转化是什么？

[1920s - 2040s] Gruenstein：我认为一切都会变好。有一大类商品，只要更便宜，我们就会买更多；而如果你相信这些商品是有价值的，那这对所有人都好。世界上还有一系列不连续点。前几天我跟一个人聊，他在一家住房工厂的产线上工作过——我想是在威斯康星——他描述说，他们工厂造的房子，比传统方式在现场施工大约贵 5%；但如果这些房子哪怕只比同等质量的传统现场建造便宜 1%，来自开发商的全部需求就会立刻转向「在工厂里造房子」这种新形态。同样的话也可以用在从垃圾填埋场回收物料，以及一大堆其他市场上。所以我们的期望是：如果 Tutor 真的非常成功，我们能提供这种机器人基底——一支其他企业可以使用的劳动力——去改造实体经济，让这个世界最终变成一个住起来更便宜、也更美好的地方。主持人：你们会进家庭吗？Gruenstein：我们对这件事没有立场，哪儿都会去。主持人：但我本来会赌那是个难得多的技术问题——从任务复杂度、安全性，还有经济性上讲，对吧？一台 Tutor 机器人在工厂里创造的经济价值是多少？有多少消费者愿意在家里为它付费？这笔账对不太上。但再说一次，这又回到我们的总体哲学：在尽可能多的栈层之间做取舍，先拿到立足点，开始建平台；而随着平台建起来，你随时间需要做的取舍会越来越少。

[2040s - 2160s] 主持人：机器人行业的产品创新是怎么运作的？多数公司还只在 v1，而你有两个版本的机器人，而且持续快速迭代。它跟数字产品创新或其他品类有什么不同？Gruenstein：我们正设法让它尽可能相似。我们自己制造机器人，这让我们能够拥有制造过程，并弄清楚怎么每两周向制造流程推送一次变更。我们给机器人发代码，要么每天发，要么每两周一次大版本发布，而且我们建了巨量的基础设施来支撑这种迭代水平。我们还建了一个我认为形状相当独特的产品职能：我们组织里的产品职能，就坐在客户成功和实施职能旁边，实际上作为一个大团队运转——他们离客户近到不能再近，我认为这对我们非常成功。然后我们还有一个研究团队，它一方面能深深嵌在客户问题空间里，能拉动各种杠杆从客户那里拿到特定类型的数据；另一方面又能打那些登月式的赌注。如果我们在建模侧有了突破、让机器人能干一种全新形态的事，我们就能围绕它协同设计一款产品——我认为大多数公司做不到这一点——并更快地把它推到现场。主持人：如果我听着这些、我有一个仓库，我想要一台机器人，我是直接去你们网站填个邮箱就行吗？这是怎么运作的？Gruenstein：你可以上我们的网站，留一个邮箱。我们还有很酷的定价计算器：你输入你的仓库或工厂今天如何运营的一些运营数据，就能得到一个估算——嘿，初始部署两台机器人，我们每年可能帮你省十万美元、一百万美元、一亿美元。而这不是「好吧，这是五年定制工程的路线图」——而是：不，两个月后的今天，你就能每年省下一百万美元，而且不用压任何资本开支。主持人：你发现外面有需求吗？这毕竟是个科幻色彩很重的东西。真有需求吗？Gruenstein：需求多得很，而且它甚至不是对机器人的需求——因为人们不知道这种形态的机器人存在，不知道它能干活，也不知道它能干这个岗位。所以归根结底——我们有这样一个框架：机器人的价值，等于它能做多少工作，除以它的成本，就这么多，对吧？其他一切都是评论。

[2160s - 2280s] 所以，如果我们能在遍布美国实体经济的这些高度流动的劳动力市场里竞争，能提供以更低成本交付同等能力的东西，那这个市场就是整个市场——是全部的体力劳动。主持人：就此问一个较真的小问题：你说「成本」时，指的是摊销前期投入，还是持续运营的成本？Gruenstein：都是，而这非常有趣、也很不同，对吧？传统上，我们把机器人当作 ROI 来想：一大笔资本开支，然后花十年把它赚回来；而现在的现实是，越来越多的成本进入了持续的推理和遥操作。主持人：那这也包括维护、故障修复这类东西吗？主要的成本是什么，还是大致相当？Gruenstein：我们把它们全部随时间往下压。主持人：不，我只是说，从你们的立场，你们摊开自己的销货成本算一算，前期投入占多少、持续成本占多少？Gruenstein：现阶段是持续成本占比更大，而这让很多人意外。主持人：有意思。哦，为什么？其实对我来说不意外——你真正在竞争的，是一种你喂它米饭它就干活的东西，而且效率高得惊人，对吧？就是人体。

[2280s - 2400s] 主持人：那客户来电，是因为黑色星期五之类要来了、需要短期内干完一大批活，还是这已经进入了他们真正的劳动力规划周期？Gruenstein：不，是战略性的。我认为这些企业面临一个非常艰难的选择：他们知道自己员工队伍的年龄，知道需求正因这些长期趋势而上行，而完成这些工作的成本基础也在上升，于是他们问：我们这门生意的账还算得过来吗？因为这些不是那种拥有巨大毛利、可以压缩一点也无所谓软件生意——不是的，这门生意要么成立，要么不成立。所以我认为，我们的客户最终寻找的是一个能够长期负责「机器人这件事」的伙伴。如果我们的一家客户在做除臭剂，他们做的就是除臭剂，对吧？他们会成为除臭剂的运营专家，或者蜡笔的运营专家，或者他们制造的任何东西的运营专家——而他们需要一个在机器人方面是专家的伙伴。主持人：我有个关于工厂本质的问题：你可以随时把机器人发货给任何人——今天还不行，因为你的产能不足，但未来某个时候可以——而如果你的机器人随时间越来越通用，这对工厂的选址、人们围绕工厂的组织方式、那种「一厂镇」意味着什么？Gruenstein：对，这一切现在都从根本上改变了。我认为这里有很多地缘政治。一个大判断是：这将成为美国与其他国家——对不起，是与中国这样的其他国家——之间的一个伟大均衡器。中国投资了上一代工业自动化，这在某些方面让他们非常成功，而在另一些方面，事后来看，那可能是一个巨大的错误：因为在美国，如我们刚才所说，我们正在用通用能力的机器人做工业产能的绿色地带建设；而在中国，他们将不得不拆掉重建。在美国，我们不需要以同样的方式做这件事。我认为另一件会改变、而且会非常有意思的事是：我们将拥有多得多的工厂，就像我们拥有多得多的企业一样——因为云和所有软件工具让人人都能创办新企业。如果我们有了这种无限灵活的劳动基底，任何人都将能够在物理世界创办企业，在物理世界规模化企业。

[2400s - 2520s] 主持人：你提到了中国——我们其实发布过一份关于供应链与机器人的报告，而中国在其中似乎非常核心。谈谈你对它的依赖、这件事如何随时间演变，以及中国对这条供应链的风险有多大。Gruenstein：我认为，无论谁怎么说，2026 年的今天，我们都有一个高度互联、相互依存的全球供应链，而这只是它的众多例证之一。美国公司和中国公司都在非常认真地思考：这一切的长期战略到底长什么样。中国公司在投资美国，美国公司在投资中国，而归根结底，所有人都在努力建设自己的工业产能。主持人：我注意到，据我所知你已经垂直化了更多——把更多环节收进自己手里，对吗？还会继续这样吗？Gruenstein：对，供应链里有件很棘手的事——再说一遍，因为机器人行业是如此的家庭作坊化：你可能从一个子组件全部来自 OEM 的世界起步，然后你的产量增长到某个点，它们就不再是 OEM 了——它们成了光荣的合同制造商，你占了它们产能的一个实质比例。到那个时候，纯粹是关系和动力学就变了，对吧？这是跟那类东西打交道的正确方式吗？我是想继续用它，还是想自己制造？主持人：我是说，Tutor 肯定是美国最大的「机器人雇主」之一，因此你们大概是这些 OEM 当中一些家的最大客户。我因此可以想象，随时间推移，你们会越来越多地自己来驱动这些组件。那意味着收进来，还是告诉他们怎么做？它们还是供应商吗？Gruenstein：我猜是某种组合。某种程度上是组合。我认为归根结底，Tutor 要成功，美国就需要有一条成熟的机器人供应链；而我们正处在这样一个位置上：因为我们是这条供应链需求的很大一部分，去非常、非常、非常重地投资建设美国的这份产能，既是我们的义务，也是一个机会——这就是我们在马萨诸塞州 Watertown 造机器人的原因，也是我们预期继续垂直化、在美国制造越来越多机器人组件的原因。

[2520s - 2640s] 主持人：再回到我们最初开始聊的时候——我记得大概是我们第二次会面，才真正开始谈研究，之前大部分时间都在谈生意，因为我们有这么好的一个生意——但研究也是你的一大强项，对吧？谈谈你怎么对待机器人学习这场赌局——关于哪种机器人学习模型会赢，有各种各样的理论。Gruenstein：在研究上，我们思考的是这件事在 LLM 里怎么演化的，哪些相同、哪些不同。我认为在 LLM 里显然非常重要的一点是：存在很多异构的数据来源，但数据金字塔的顶端，是「做你想让它做的那件事」的真实世界强化学习轨迹。所以我们把研究项目建成了这样一个形状：思考如何从真实世界的机器人经验中挤出最多的价值。为此我们在波士顿建了这支一百台机器人的机队，叫 Data Factory One。据我们所知，这是美国最大的半人形数据采集场。我们把它当作一个研究试验台：怎么用机器人做大规模的真实世界强化学习。要知道，强化学习这个技术家族在小批量下基本不奏效；而如果你在真实世界里有一支非常大的机队，你就可以开始玩这些方法——我认为无论学术界还是工业界，以前都没人玩得起。

[2640s - 2697s] 主持人：我猜为这件事建团队相当不一样，跟那些实验室不同。也许聊聊团队的构成，还有你们在招什么样的人。Gruenstein：我们在招所有人。所有人。我们在招研究者和工程师，横跨技术栈的每一层。哇。我们也在招造机器人的技师。顺便说，所谓每一层，你是说软件、硬件、材料、电子这些都要？对。在这件事上，没有任何一个地方可以推卸责任，对吧？不管你愿不愿意，你都会与每一层打交道；问题只在于你是否理解它。所以我们正在建这样一支团队，而我认为这是一种非常独特的团队形状：不仅因为它如此跨学科、在技术栈的所有这些层之间高度协作，还因为我认为它是这个宇宙里仅有的几个这样的地方之一：早上你还在白板前聊模型架构，下午你就在看一条工厂产线、看一台机器人干得怎么样。主持人：团队主要在马萨诸塞吗？Gruenstein：我们的团队遍布美国各地。尤其是 go-to-market 和技师这些角色，他们要在一线接触——我们的客户遍布全美。主持人：好，太棒了。谢谢你们。