Foundation Robotics Deploys Humanoids in Ukraine as It Seeks $60K Worker Replacements

Foundation Robotics在乌克兰部署人形机器人，瞄准替代6万美元年薪工人 CEO Sankaet Pathak认为，武装人形机器人和边缘物理模型将在动荡的15年转型期内取代蓝领劳动力，同时消除对合成训练数据的依赖。 Sankaet Pathak分享了他对Foundation Robotics的愿景，详细阐述了基于边缘的世界模型（“Cortex”）如何让人形机器人模拟物理未来以实现零样本任务执行。他讨论了在汽车组装、港口物流和军事侦察领域的实际部署，包括在乌克兰的实地测试。Pathak还批评了美国硬件供应链的短板和人造“cosplay”式训练数据，同时探讨了国防机器人的地缘政治伦理以及未来不可避免的劳动力市场冲击。 Sankaet Pathak 离开金融科技领域创立 Foundation Robotics，旨在推动人类向 I 型文明迈进，其终极目标是在美国建造首个完全由机器人建造和管理的城市，并通过工业与国防领域的实际部署为研发提供资金。 (0:04) 在首年小规模部署中，Foundation Robotics 的机器人已协助制造了超过 24,000 辆汽车，并应用于港口集装箱物流；同时团队正利用 3D Gaussian splatting 开发室内侦察系统，以降低士兵在未知建筑中的作战风险。 (2:37) 针对具身智能架构，Pathak 强调机器人世界模型不同于追求视觉画面的视频生成器，其自研端侧系统 Cortex 能在前额叶层面预测数秒至数分钟的物理真实动态，实现零样本任务泛化。 (6:58) Pathak 指出美国硬件供应链存在严峻瓶颈，铜线圈成本高出中国百倍且稀土磁铁供应困难；此外他坚决摒弃实验室摆拍的“Cosplay 数据”，强调 Cortex 预训练需要 5 万到 50 万小时真实的工业运营数据。 (10:06) Pathak 确认重型人形机器人 Phantom 已在乌克兰进行物流与侦察实地测试，并表示如果美国政府提出需求，公司将毫不犹豫地为机器人装载打击能力以实现威慑与精准打击。 (14:13)

Foundation Robotics founder Sankaet Pathak argues that commercializing humanoid robotics requires deploying edge-based physical world models in heavy industry to generate revenue, while confronting severe Western hardware supply chain vulnerabilities and significant impending labor disruption. Industrial Deployments and Near-Term Labor Economics (2:37) Foundation Robotics focuses its initial deployments on industrial materials handling and autonomous navigation to build cash flow for long-term objectives, including Pathak's ultimate ambition of constructing robot-built cities. In its first small-scale deployment, the company's robots assisted in manufacturing more than 24,000 vehicles by handling material transport, picking, and precision racking. Pathak identifies severe labor shortages as the primary commercial driver, noting that a fully loaded U.S. factory worker costs approximately $60,000 annually per shift, whereas robots can operate across three continuous shifts. Edge World Models Enabling Physical Predictive Imagination (6:58) Pathak distinguishes robotic world models from generative video models like Runway, noting that while video systems prioritize visual frame persistence, robotics demands grounded semantic and physical intuition. Foundation's AI architecture, Cortex, is designed to run locally on the robot's edge hardware rather than relying on cloud diffusion. Operating within the low-latency control feedback loop, Cortex provides a predictive 'physics imagination' that forecasts physical state changes 30 seconds to several minutes into the future, enabling zero-shot task execution from natural language instructions without task-specific prior training. Hardware Bottlenecks and Rejection of Cosplay Data (10:06) Pathak identifies the American hardware supply chain as a critical vulnerability, pointing out that domestic copper coils cost 100 to 1,000 times more than in China, while domestic rare earth magnets and lithium refining remain critically constrained. Concurrently, Pathak criticizes venture-backed 'data factories' that collect staged human demonstrations in studio settings, dismissing these outputs as 'cosplay data' that fails to reflect heavy industrial realities. He states that model architectures are largely solved, making the primary bottleneck the collection of 50,000 to 500,000 hours of authentic, real-world operational data required to pre-train Cortex for complex planning and replanning. Defense Deployments and Willingness to Arm Robots (14:13) Foundation has deployed its heavy humanoid robot, Phantom, for field testing in Ukraine, currently limited to logistics, cargo loading, and interior reconnaissance using 3D Gaussian splatting to map hazardous buildings for soldier safety. Unlike competitors that avoid military collaboration, Pathak embraces defense contracting and explicitly commits to arming Foundation's robots if requested by the U.S. government. Pathak argues that deterrence maintains peace and contends that precision robotic targeting is morally superior to indiscriminate aerial bombardment. Navigating Transitional Labor Disruption and Societal Risk (18:17) Pathak acknowledges that humanoid robots will inevitably make traditional blue-collar physical labor obsolete, forecasting a disruptive economic transition over the next 10 to 15 years. Drawing parallels to the First Industrial Revolution, he notes that while mechanization ultimately lowered consumer costs and created larger overall industries, failure by policymakers to manage the near-term displacement could trigger severe societal unrest or political backlash from workers fearing permanent displacement. He advises incoming workers to cultivate adaptability and systems orchestration rather than focusing on automatable manual tasks. Pathak's roadmap hinges on two contentious premises: that Western robotics can overcome profound domestic manufacturing cost disadvantages against foreign supply chains, and that society can navigate a 10-to-15-year window of severe blue-collar displacement without triggering widespread unrest or backlash against weaponized and autonomous systems.

Foundation Robotics 创始人 Sankaet Pathak 指出，人形机器人商业化必须以重工业制造和国防后勤为突破口，并通过本地边缘部署的物理世界模型实现任务自适应；然而，美国本土关键硬件供应链的严重短缺、虚假合成数据的泛滥，以及未来十至十五年的就业替代阵痛，构成了行业推进的核心制约。 重工物流与国防测绘的近远期部署 (2:37) Foundation Robotics 的商业化重点聚焦于重型物料搬运与自主导航。Pathak 透露，在其首年小规模部署中，机器人已进入汽车制造工厂承担物料搬运、拣选与精密零部件上架，协助完成了超过 24,000 辆汽车的生产；港口场景中，机器人则与运营方合作进行集装箱转运及零部件分配。在劳动力经济性方面，美国制造业单班次工人年综合成本约为 60,000 美元，而机器人可连续运行三个班次，有效填补了制造业空缺。未来 5 至 10 年，公司计划拓展至建筑施工、全城物流、数据中心基建和农业作业，其远期目标是建设并运营美国首座完全由机器人建造的城市。 在非工业场景中，公司正为国防研发端到端侦察系统。该系统针对士兵进入未知建筑易遭伏击的痛点，由机器人先行进入室内并攀爬楼梯，输出 RGB 地图、标准网格地图以及模拟人眼空间视觉的 3D 高斯泼溅（3D Gaussian splatting）高精地图，在人员进入前提供战术数据支持。 Cortex 架构与边缘端物理世界模型 (6:58) 针对驱动机器人的核心 AI 架构 Cortex，Pathak 强调了具身智能模型与 Runway 等纯视频生成工具的本质区别。视频生成模型的核心目标在于维持视觉画面的跨帧一致性（例如视线移开后白板笔迹不变），而具身机器人模型必须具备底层的物理动态与空间语义直觉，类似于儿童通过物理碰撞认知真实世界规律。 Cortex 模仿人类前额叶皮层，被设计为必须在机器人本体边缘硬件上运行，而非依赖云端扩散模型。这种本地闭环设计允许机器人在执行连续动作时，实时预测未来 30 秒至数分钟内的物理状态演变，并将预测结果融入控制反馈机制中。通过结合自然语言与语义理解，Cortex 具备零样本（Zero-shot）执行能力，使机器人无需针对特定动作进行显式预训练即可直接解析并执行全新指令。 硬件供应链断层与非道具实景数据瓶颈 (10:06) Pathak 直言美国人形机器人产业正受制于两项深层瓶颈：硬件供应链失衡与错误的数据训练路径。在硬件端，美国关键加工制造链严重滞后，导致本土铜线圈成本达到中国的 100 至 1,000 倍，稀土磁铁在本土难以稳定获取，同时风投资本普遍回避锂电池精炼等基础制造业投资，更倾向于采购海外制造的执行器。 在算法数据端，Pathak 强烈抨击了行业内建设人工演示“数据工厂”的做法，将其贬斥为脱离真实的“角色扮演数据”（Cosplay data），并比喻为让美剧演员去进行实际外科手术。他表示，在模型架构问题已基本解决的当下，公司的关键挑战是获取 50,000 至 500,000 小时来自真实生产环境的非摆拍作业数据，以此来预训练 Actor 基座模型，确保其任务规划、重规划与控制精度。 乌克兰实地测试与防务伦理考量 (14:13) Foundation Robotics 的重型人形机器人 Phantom 已被部署至乌克兰进行严苛环境下的实地测试。目前机器人在战区主要执行非战斗减险任务，包括货物装卸、物资后勤补给与战术勘测制图。Pathak 称该公司是西方极少数愿意与防务机构直接合作的人形机器人企业之一，并明确表示若美国政府提出要求，公司将毫不犹豫地为机器人加装动能武器系统。 针对武器化带来的伦理质疑，Pathak 结合自身成长经历主张和平必须以实力为后盾。在道德层面上，他认为高精度的机器人精确打击远比摧毁整座街区的无差别空中轰炸更具人道主义价值。同时他强调，如果西方企业主动放弃军用机器人的研发，威权对手必然会抢先掌握该技术，从而使西方社会陷入单向暴露的危险境地。 劳动力市场重构与过渡期社会风险 (18:17) 针对自动化带来的冲击，Pathak 坦承人形机器人将不可逆转地淘汰大量重复性蓝领体力工作。他援引第一次工业革命期间纺织机与蒸汽机普及的案例指出，虽然特定工种在初期被消灭，但生产成本的大幅下降最终引爆了市场需求，使得纺织业雇佣的整体规模远超以往。他评估，社会吸纳颠覆性技术并转化为生产力所需的时间已缩短至过去的约三分之二。 然而，Pathak 预测未来 10 至 15 年将是转型阵痛最为剧烈的摩擦期。他警告，如果政策制定者不能妥善平抑转型冲击，公众因担忧沦为“永久底层阶级”而产生的恐慌可能引发严重的社会动荡或抵制反抗。对于年轻一代，他建议应避免专精于机器可替代的单纯体力劳动，转而聚焦于系统编排、技术素养与应对复杂环境的高维度自适应能力。 Pathak 认为具身智能终将通过大幅压低生产成本来提升全球生活水准，并推动人类向 I 型文明跨越；但该愿景能否兑现，短期内依然取决于企业能否突破高保真物理交互数据的规模化采集，以及西方能否在关键硬件供应链上摆脱外部依赖并渡过未来十余年的社会就业摩擦。

## Background 该内容由 YouTube 频道 Ai4 于 2026 年 8 月 22 日发布，主要内容为 CNN 资深记者 Matt Egan 与 Foundation Robotics 首席执行官 Sankaet Pathak 的对谈。视频说明指出，这场演讲深入探讨了基于物理学的推理与想象在推理阶段的应用，介绍了一种结合高层规划与底层全身控制的层级化架构，展现了具身人工智能领域的全新发展路径。 ## TL;DR 对谈围绕 Foundation Robotics 的创业愿景及其技术路线展开。创始人 Sankaet Pathak 分享了离开金融科技领域、致力于通过具身智能推进人类迈向“1 型文明”的初衷，以及未来建造新城的宏大构想。同时，内容重点阐述了如何借助物理隐空间世界模型实现无须外部仿真器的实时“想象”与推理，克服传统模仿学习的局限，从而大幅降低数据需求并提升机器人的泛化与适应能力。 ## Atmosphere & style 对谈充满了前瞻性的科幻色彩与极客式的创业热情，主持人与嘉宾之间的交流自然而富有深度。Pathak 的分享既有对人类文明未来的宏大憧憬，又落脚于严谨扎实的技术架构细节，呈现出一种充满野心且富有感染力的思想碰撞氛围。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=Rr9orQ5nuiE&t=4s) Sankaet Pathak 表示自己离开金融科技行业是为了专注于能加速人类迈向“1 型文明”的伟大事业，创立 Foundation Robotics 的长远愿景是在美国从零开始建起第一座全新城市。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=Rr9orQ5nuiE&t=4s) 创业愿景：从 1 型文明到构建第一座新城 访谈由主持人 Matt Egan 对 Foundation Robotics 首席执行官 Sankaet Pathak 的问候拉开序幕。主持人提及 Pathak 过去曾表示离开金融科技（Fintech）行业的原因在于只希望专注于能够加速人类迈向“1 型文明”（Type-One Civilization）的工作，并询问人形机器人如何契合这一宏大目标。Pathak 坦言创业极为繁重且充满挑战，与其在华尔街赚取更多财富，不如选择真正能激发自身热情、将科幻概念落地变为现实的伟大使命。 Pathak 重点讲述了他在 2023 年 12 月圣诞节期间的“发日间梦”经历。当时他在与家人聚会时完全无法集中注意力听人交流，脑海中不断构思着机器人未来构建、管理与维护基础设施的图景。这种执念最终促成他在几月后创立了 Foundation Robotics。公司的长远愿景是在美国从零### 创业初衷与“一型文明”愿景 在对谈开头，主持人 Matt Egan 提及 Sankaet Pathak 此前放弃 Fintech（金融科技）领域的经历。Pathak 解释道，创办公司极 Sorry, something went wrong. Please try your request again. ## Notable quotes > [00:04](https://www.youtube.com/watch?v=Rr9orQ5nuiE&t=4s) 只希望专注于能够加速人类迈向“1 型文明”（Type-One Civilization）的工作 — Sankaet Pathak ## Mentioned Sankaet Pathak, Matt Egan, Foundation Robotics, CNN, Ai4 ## Video description Sankaet Pathak — CEO, Foundation Robotics Matt Egan — Senior Reporter, CNN What if humanoid robots could imagine possible futures before acting—just like humans do? The talk will include a deep dive into why today’s dominant approaches, from imitation learning to model-free reinforcement learning, struggle to scale toward general-purpose embodied intelligence. It will illustrate a new hierarchical architecture in which a physics-grounded latent world model bridges high-level planning and low-level whole-body control, enabling true model-based reasoning directly on the robot at control frequency. The session will explain how this system learns physics and task rewards from human videos, performs rapid imaginative rollouts in latent space, and selects actions through internal forward simulation—without external simulators or rote trajectory memorization. Attendees will gain a clear understanding of how physics-based reasoning unlocks faster adaptation, stronger generalization, and dramatically reduced real-world data requirements. Join this talk to explore a fundamentally different path toward scalable, embodied AI. [Source](https://www.youtube.com/watch?v=Rr9orQ5nuiE)

[0:04] Sankaet Pathak: 只希望专注于能够加速人类迈向“1 型文明”（Type-One Civilization）的工作

[0:04] Sankaet Pathak: making sci-fi capabilities a real-world reality to dramatically accelerate human progress. [0:04] Sankaet Pathak: build and manage the first entire city in America constructed by robots. [2:37] Sankaet Pathak: helped build over 24,000 cars in automotive manufacturing plants by handling material transport, picking, and precision parts racking. [2:37] Sankaet Pathak: A fully loaded U.S. factory worker costs ~$60,000 annually per shift. A robot working three continuous shifts provides massive cost efficiency [6:58] Sankaet Pathak: predicts physical state changes 30 seconds to several minutes ahead during continuous action execution. [10:06] Sankaet Pathak: Copper coils can cost 100 to 1,000 times more in America than in China [10:06] Sankaet Pathak: cosplay data [10:06] Sankaet Pathak: expecting a actor on Grey's Anatomy to perform real surgical procedures. [14:13] Sankaet Pathak: if requested by the U.S. government to equip robots with kinetic capabilities, Foundation would comply without hesitation. [14:13] Sankaet Pathak: precise robotic target engagement is far more humane than indiscriminate bombing that destroys entire city blocks. [14:13] Sankaet Pathak: if Western defense-oriented robotics companies do not build these capabilities, totalitarian adversaries will, leaving Western nations exposed. [18:17] Sankaet Pathak: humanoid robotics will make significant portions of traditional blue-collar jobs obsolete. [18:17] Sankaet Pathak: avoid optimizing for repetitive physical tasks that machines can automate. Instead, they should cultivate adaptative problem-solving skills [19:26] Sankaet Pathak: true friction will occur over the next 10 to 15 years, admitting that it would be disingenuous to pretend this transition won't be messy. [19:26] Sankaet Pathak: failing to do so could lead to severe societal unrest or "a mutiny" driven by people's legitimate fear of becoming a permanent underclass.

[0:04] Thanks for doing this. Thanks for uh  agreeing to to talk to all of us here. [0:08] Um I think everyone would love to hear more about  your company and your vision. Um you've said in [0:15] the past that you left fintech because you only  want to focus on work that accelerates humanity [0:22] to be a type one civilization. I'm wondering if  you can just sort of talk to us about how these [0:29] humanoid robots that your company is building  fit into this ambitious goal that you have. [0:37] Uh yeah, I mean generally what I learned  is that building companies is hard. So, um, [0:43] you can do what motivates you the most.  And actually, what gets me up is, uh, [0:48] getting very fired up about an exciting mission  and, uh, trying to accomplish that over, uh, [0:54] I don't know, maybe if I worked at  Wall Street making more money. Um, [0:58] so I decided that I wanted to spend the rest  of my life doing those kinds of things. And [1:04] I think now more than ever, it's finally  possible to be able to kind of make [1:10] sci-fi not sci-fi and make it actually happen.  And majority of those things I think will [1:15] greatly accelerate humanity. the area where I got  most excited about and so much so uh in December [1:21] 2023 um I was uh u at Christmas with family and I  would not pay attention to any conversation just [1:29] kept on daydreaming was this notion that one day  um uh robots would be able to build infrastructure [1:36] uh and also manage it and maintain it and do all  these different things and that obsession uh and [1:42] daydream continued for months and I decided that  maybe I should try to do something about about it [1:47] and um decided to start foundation and the goal  with foundation by and large is uh we are uh we're [1:54] building robots for heavy industrial work. So  humanoids are one form of robot that are extremely [2:00] important if you want to be able to build homes  and uh build and uh essentially manage cities and [2:05] things like farming and defense. Uh there are  other forms of robots that are also important [2:10] but uh from now till we get to the point where  we can actually build a first city in America [2:15] which is our big self-stated goal. Uh there are a  bunch of technologies that need to be created and [2:20] you have to be able to finance your way there  and the only way you do that is by creating [2:23] real utility and value. So today majority of our  focus ends up being uh defense and industrial work [2:30] and a lot of that is to be able to build the  robots and capabilities needed to be able to [2:34] build these very giant cities. Okay. So, so let's  get into to your business. Um, you mentioned [2:42] sci-fi. So, can you explain the use cases,  the specific use cases for these AI powered [2:49] robots that your company's building? I mean, h how  are they being deployed today and and what are the [2:55] use cases that might feel like science fiction,  but you believe will actually become a reality [3:00] in the next 10 years? I mean the things that we're  working on right now are um I think we're getting [3:05] to a place where our robots can by and large do  a couple of things uh almost anywhere. So one is [3:12] be able to move materials around which uh as it  turns out if you go to factories or even if you go [3:18] to logistic centers or you go uh into the defense  industrial complex a lot of the work that people [3:24] are doing is just moving and transporting things  from point A to point B. It sounds very mundane [3:29] but there is a lack of labor force there in itself  and you can get a lot done by just doing that. Um [3:34] which means today uh we we do a bunch of materials  handling work and also pick place racking type [3:41] work in auto manufacturing. So uh we helped build  over 24,000 cars last year which was the first [3:47] year of our very tiny deployment and this year  uh our goal is to ship more and more robots. Um, [3:54] and we also have contracts with port operators  to be able to pretty much do the same thing, [4:00] carry materials from point A to point B and uh,  in some cases some slightly more precise things [4:07] that you have to do around it, which is racking  and racking. You also to rack a part inside. Um, [4:12] so there are tons of jobs in just doing that.  Uh, the other kinds of use cases are now a [4:18] robots can uh, autonomously navigate and map an  environment. So that alone is a capability have [4:24] has has massive value. Um uh so currently we are  building an end toend reconnaissance system. So [4:31] today soldiers get heard uh frequently going  into buildings that are unknown. So we want [4:37] robots to be able to go in first be able to take  stairs and map the entire environment. um uh [4:44] RGB map, obviously your grid map, but then also  your Gaussians map, which looks similar to how [4:50] our eyes see the world. And with that, our goal is  give people give soldiers very precise information [4:56] about where things are. But those capabilities  are also very helpful. Again, as you go back to [5:01] operating a port, when you have to go from one  remote end of the port to another, you need a [5:05] robot that can navigate itself and understands the  map quite well. Um currently we're heads down on [5:11] those kinds of use cases. Uh we think those create  just tremendous value in the economy generally [5:16] uh uh solve for a massive need that exists across  different industries both defense and industrial. [5:21] So auto manufacturing port logistics defense  logistics those are the focuses. Did you say [5:27] 24,000 cars were made? Well we didn't make the car  entirely but we helped make 24,000 cars. Okay. So [5:34] autos ports defense now. Yeah. What about the  use cases going forward the next 5 to 10 years? [5:45] Where where else is it going? Um I I mean  construction uh a lot of um city logistics and um [5:54] uh farming farming is challenging for a  very interesting reason which is today the [6:01] cost of labor over there is extremely low. So you  need robots that can uh uh be very economically um [6:11] uh feasible to be able to do that versus  if you go to a factory in the US a fully [6:15] loaded u cost of a one one factory worker is about  $60,000. So uh if if a robot can do three shifts [6:24] uh you are demonstrating not only you're adding  more capacity uh for these people to be able [6:28] to build more goods and services because they  already have a shortage of labor uh but then on [6:34] top of that you can do it at a cheaper price. So  I think it's going to take some time to be to the [6:39] equation to be reversed there but but I think that  would happen. uh the areas that like we didn't I [6:44] I didn't anticipate uh when I started the company  but a lot of the infrastructure projects around uh [6:50] data centers uh that would also become a big part  of the use cases uh in the near future. Okay. [6:58] Can we talk about the technology? Um can you  tell us about Cortex the AI architecture powering [7:05] your robots? Um and and can you just explain  to the audience how world models AI is [7:11] making these robotic advances possible? Yeah,  I mean um so ju just some like technical [7:18] differentiation different people consider um  there are different kinds of world models. So [7:25] some people call simulators world models. Some  people call video generation tools like uh [7:31] uh runway right before this uh a world model. Um  and uh it seems like now everyone awkwardly is [7:38] trying to nudge those as a use case into robotics  which uh you have slightly different problems. So [7:43] what what what Runby really needs to be able  to build for their original use case is a very [7:48] persistent layer of video generation that to their  point if you leave a room and come back what was [7:54] written on the whiteboard it's exactly the same.  It doesn't change. Um the problem set you want to [7:59] optimize for robotics is slightly different. You  want to be able to build a semantic intuition. So [8:03] very similar to how we as kids build by bumping  into objects and over time learning how the [8:09] physics of the world really works. Uh some kind of  intuition around it and then I can just read text [8:15] and I can do not everything like like an expert  but I can do majority of the stuff quite well. Um [8:23] so we focus our time and energy in building out  that. So the biggest question on the research side [8:29] that uh we've been obsessed with trying to answer  is how big of a world model do you need to be able [8:33] to build that gives you enough predictive ability  to be able to look a minute maybe 2 minutes [8:40] even 30 seconds out into the future so that  whatever is the length of the action you need [8:46] to perform that's next right next to whatever you  had just done you know how the world's going to [8:51] shape as you do that. Um so we more so call that  like uh uh physics imagination than anything else [8:58] which is a kind of a world model and our goal  is we want that world model because it needs to [9:03] be uh uh somewhat predictive and somewhat reactive  because that is the big differentiation between a [9:09] diffusion model that does video generation for  graphics and another one that does for robots. Um [9:16] you want that to sit on the edge. So you want  you really want that to sit on the robot. [9:20] If it cannot sit on the robot, you cannot  realistically perform actions because you really [9:24] need that in the feedback loop. It's not something  it generates and forgets. Uh so a lot of our focus [9:29] and energy on the cortex side is um pretty pretty  much building this like prefrontal cortex and pre [9:38] preoter cortex equivalent of what humans have.  um which by and large is being able to take [9:46] a language and have some kind of a semantic  layer that lets you take that language in [9:53] zeroot which means you have gotten no training  data for that task whatsoever and um still be [10:00] able to do that task reasonably well. So  as you're you're building these robots and [10:06] the world models, what are the biggest bottlenecks  and obstacles that you're facing as you continue [10:14] to build out this business? Uh I mean this is  a very complex business like everything's very [10:19] interdependent. Uh so if I we have 9 minutes I  could go for an hour about the bottlenecks uh that [10:26] we're facing on a regular basis. Um let me kind  of like focus on something that is very very [10:32] minute and then I'll focus on something like very  very big. Uh what keeps me up at night is uh uh [10:38] a coil of copper is about 100 times more expensive  sometimes thousand times more expensive in America [10:44] than it is in China. What keeps me up at night is  for the life of me I cannot find a good supplier [10:49] that can consistently give me rare earth magnets  in America. Uh what keeps me up at night is you [10:54] cannot get good lithium-ion refinery. uh to be set  up. I cannot get a single venture capitalist to [11:00] be excited about building a lithium refinery  in America, which I get it people like it's [11:07] not the most fancy thing, but they would fund an  actuator which is built by China 90% of the time, [11:13] but they would not fund what builds that  actuator, which is just crazy. And if like [11:18] if I were in fintech or I was in investment  banking, uh this to me this is the next SpaceX. [11:26] This is the there's no good reason for if you  go cost by cost and I have done that because [11:31] I'm an obsessive person of exactly what is it that  costs you to be able to get copper built in the US [11:38] versus China. There are some Chinese subsidies  but our labor isn't that expensive for it to be [11:42] 100 to a thousandx more expensive here than it is  there. um that is a big bottleneck because that [11:48] does make it so we are an all-American robot  that is also serving the defense industry but [11:55] there are parts we're slamming to get  that don't actually come from the US um [12:00] so that that in itself is a bottleneck now you  zoom all the way out uh leave that piece aside [12:05] for a second uh for last two years our biggest  bottleneck was what is the right data recipe [12:12] and the model architecture recipe on the AI side  that could lend itself well to be able to process [12:20] a lot of atcale data that doesn't have to be  teleyoperated data to get these capabilities. Uh [12:27] we finally cracked the code on that. So now  over there our big challenge is how do we get [12:35] 50,000 to 500,000 hours of non what I call  non-costplay data which is data that was not [12:43] created in a data factory which my peers are  very excited to announce with every single [12:47] venture round that they do that everyone's  now doing a data factory they're all fake jobs [12:52] like there's nothing that's happening in  there that is going to happen in a factory [12:56] you could have the best script writer Grey's  Anatomy cannot teach you how to be a doctor. [13:02] It's just that plain and simple. The same way it  cannot teach a robot how to build a car. There's [13:06] just no two ways to it. So right now my big  bottleneck is again supply chain on the data side. [13:11] What we really nailed down the last two years  is we've built a robot that is really good, [13:16] quite reliable, probably the best actuators.  It's waterproof, dustproof, it's very strong. [13:22] But those are two ends of the spectrum um that  I think are the big big bottlenecks. And then [13:26] the lazy answers are oh I need more compute blah  blah blah blah blah. The reality is today like [13:33] you don't have data deficit unless the data  you're talking about is this right like a [13:38] large portions of our AI layer can get quite  good just with the data on the internet. [13:44] But the pieces that really matter  which is planning, replanning, [13:48] um task precision, all of that which is  the actor layer of the model architecture, [13:53] all of that needs non-costplay data at the  foundation model pre-training level. Um and uh [14:00] that is definitely like another big bottleneck.  I think that would be solved before my um [14:05] my uh copper um steel magnet and lithium issue  would be solved. Okay. So, you mentioned that [14:13] you've built these robots that are quite good.  That's the Phantom that you're talking about. So, [14:20] I understand that you've sent units, these  Phantom units to Ukraine for for testing. [14:27] Just for testing. Just for testing. What can you  tell us about how these these robots are working [14:35] in the real world battlefield? And you  relatedly, it it seems as though your [14:40] company's taking a different approach from Boston  Dynamics and others in the field in terms of uh [14:47] using humanoid robots in these defense Yeah.  settings. Uh can you just talk about why [14:54] you're leaning into um the military defense use  cases and and what do you say to those who have [15:01] ethical concerns about Yeah. using robots  potentially lowering the threshold uh to going to [15:08] war. Well, so taking the sensationalist piece out  for a second, today our robots are not uh fighting [15:14] wars. So that's not what they're doing in the  defense complex in itself. There's a large part of [15:21] war that is just making sure the person  who's sitting in the foxhole has supplies. [15:27] Making sure they don't just leave  that foxhole and they get shot at. [15:31] making sure the fossil they're entering in  is safe. Uh all of that work needs to happen [15:36] around them. So a lot of our focus has been those  things which is being able to carry materials [15:41] outside to inside, inside to outside, load load a  truck, unload a truck, be able to reconnaissance, [15:47] the thing that I talked about, map a building,  map an environment that's unknown so that a [15:51] human doesn't have to go in first. Um so those  are the big things we're focused on. Now if [16:00] the US government came to us tomorrow and said  we need you to put real guns in hands of these [16:06] uh robots as well and put them out  in the battlefield. We will do it [16:11] and that is the point of friction to some people  but I can explain why we would do it and why [16:15] you shouldn't have a point of friction  around this. Every single time I say that, [16:19] people look at me as if like, I don't know, I'm  Oenheimer or something and I'm building this [16:23] atomic bomb. And the reality is it's quite the  opposite. If your concerns around war are ethical, [16:30] would you rather want a drone dropping a bomb on a  building and taking out three buildings next to it [16:36] and everybody in that building or would you rather  want a robot that can walk into that building and [16:40] go to the target? I would much rather want the  ladder. I cannot stop wars. You cannot. Nobody [16:45] can. And this idealistic position is like very  privileged to be able to say, "Oh, I wish there to [16:52] be no wars." Yeah, no [ __ ] So, so do I. So does  everybody else. Like that's not really the issue. [16:57] There are wars. They're going to be wars. The  question is, can we make them more precise? And [17:01] to the credit of the Western government, doesn't  matter which party you really believe in. It's [17:07] it's the elephant or the donkey. It really  doesn't matter. Something in the middle. [17:11] The reality is we have gotten better at it.  We have become more precise. We have become [17:16] better and better and more just and more  empathetic in being able to execute things [17:20] that are really scary. There are real national  security concerns that do warrant us to defend [17:26] ourselves and go to war. I was born in a country  that was really bad at protecting itself and was [17:31] paged for generations again and again and again  and again. And if you read history, the lesson you [17:35] really learn is for you to have peace, you have  peace through strength. You do not have peace [17:41] with no weapons. That just doesn't happen.  So if the US government comes to us and says [17:47] we need you to arm these robots,  that means 99.9999% of the time, [17:53] what that implies is they want to take it  to the next level in precision. And we want [17:59] to be supportive of that. And the reality is if  foundation dies today, if we stopped existing as [18:04] a company, there's not a second humanoid company  in the entire western world that would do that. [18:11] And that is unethical. Me building robots that  can potentially carry weapons, I don't think is [18:17] okay. I think we have time for maybe one more  question. Um, you know, I've talked to people [18:21] at this conference and around the conference who  are nervous about the labor market implications [18:29] of AI and and robots and and some of them are also  not just concerned about their own professions, [18:34] but what their kids are going to do, what fields  they're going to get into. And I'm just wondering [18:39] whether or not the you believe that the  robots that you're building are going to make [18:45] a significant amount of bluecollar jobs obsolete  and what your advice is for young people as they [18:52] try to navigate this disruptive environment.  Um short answer is yes. A robots will make [18:59] some kind of blue collar jobs uh obsolete just  like the first industrial revolution did with [19:06] textile um and with engines  and the second one did um [19:11] or more things and the third one did with  more desk jobs all these different things [19:16] uh that will happen and I think there will  be a period that's highly disruptive uh [19:20] like again first industrial revolution I think  was 60 years second was about half that third [19:26] was about like 2/3 that um so we are getting  better as society to be able to take these very [19:33] disruptive technologies and then quickly point  them to a more productive outcome for humanity. [19:39] Uh but there's going to be a disruptive u part to  this. Uh and I have friends who ask me say well [19:46] uh you're in this field what do you think my kid  should do? Well the reality is if your kids's like [19:50] 5 years or younger you have nothing to worry  about because any disruption will be solved for [19:55] by that time. The issue is the next 10 to 15 years  is the one that is going to be a little bit messy. [20:01] And I'd be lying to you if like I I I think that  would not be the case. I think there's going to be [20:05] some movement there. I'm very optimistic. Uh I do  spend a very tiny portion of my time talking to uh [20:12] people who uh run this country as well. And um  it seems like everyone's been quite thoughtful [20:18] about it. Um and I'm very optimistic that we  can have a much smaller disruptive period than [20:24] we would otherwise. Uh, and I think that's like  very very needed because if that doesn't happen, [20:28] the alternative to that is pretty much a mutiny.  So, uh, because people are very afraid of just [20:34] being a permanent underclass. And I think  there's some real legitimate fears around that. [20:37] Uh, but from what I can tell, the short answer  is I think I I think we're going to do a good [20:42] job navigating it. But yes, it will eliminate some  some jobs just like how industrial revolution one [20:47] eliminated some jobs in textile. But guess what?  We employ significantly more amount of people in [20:52] textile today than we did back then. And that's  because the market got bigger, the goods got [20:56] cheaper, quality of life got great. Like that is  all going to happen. Majority of the people do not [21:01] have the quality of life that we have in America.  And guess what? That would change. Okay. Well, [21:06] thank you so much for uh making the time to talk  to us. Thank you for having me. Thanks for coming.