基于 Metix AI 8.6 亿+ 全球人才库，对 13 家具身智能与人形机器人主体（纯具身创业 + 大厂内部机器人组）做全量画像：这个估值爆炸、人才池却极薄的赛道，人到底从哪来、有多少、谁可触达。Drawing on Metix AI's 860M+ global talent pool, we profile in full 13 embodied-AI and humanoid-robot players (pure-play embodied startups + in-house robotics groups at the tech giants): in a sector where valuations have exploded but the talent pool stays razor-thin, where do the people actually come from, how many are there, and who is reachable.
以下数字为 Metix AI 数据库口径（数据截至 2026 年上半年），统计对象 = 当前在职于 13 家具身主体、坐标美国为主（含 1X 挪威）的人才。The figures below follow Metix AI database methodology (data through H1 2026); the population = talent currently employed at 13 embodied-AI players, based mainly in the US (incl. 1X in Norway).
赛道估值口径已到 Figure $39B、Skild $14B、Physical Intelligence $5.6B（在谈 $11B），但可见技术人才合计仅 1,044 人，分散在 13 家主体。最受追捧的纯 AI 人形/基础模型团队尤其精干（Physical Intelligence、Skild、The Bot Company 各只有数十名可见技术人）。资本厚度与人才薄度的落差，是这个市场所有招聘动作的底层约束。Sector valuations have reached Figure $39B, Skild $14B, and Physical Intelligence $5.6B (with $11B reportedly under discussion), yet visible technical talent totals only 1,044 people, spread across 13 players. The most sought-after pure-AI humanoid / foundation-model teams are especially lean (Physical Intelligence, Skild, and The Bot Company each show only a few dozen visible technical staff). The gap between deep capital and thin talent is the underlying constraint on every hiring move in this market.
技术池里 26.4% 是硬件/机电（执行器、机械、电子、嵌入式），是仅次于软件的第二大职能；叠加运动控制等，造「身体」的工程师占近三成。这是具身与自动驾驶、AI Lab 最根本的不同，后两者硬件岗近乎为零。造身体的人和造大脑的人来自完全不同的供给池，招聘策略必须分轨。26.4% of the technical pool is hardware/mechatronics (actuators, mechanical, electronics, embedded) — the second-largest function after software; add motion control and the like, and the engineers who build the "body" account for nearly 30%. This is the most fundamental difference between embodied AI and both autonomous driving and AI labs, where hardware roles are close to zero. The people who build the body and the people who build the brain come from entirely different supply pools, so recruiting strategy has to run on separate tracks.
整体 8.2% 有 AV 背景，但高度集中在软件/基础模型团队：Bedrock Robotics 46.0%、The Bot Company 37.9%、Tesla Optimus 30.0% 的技术人来自自动驾驶。来源以 Cruise、Zoox、Waymo 为前三，都是自动驾驶领域近年人才释放的主力。硬件型公司（Boston Dynamics、Apptronik）AV 占比则个位数。Overall, 8.2% have an AV background, but it is highly concentrated in the software / foundation-model teams: 46.0% of technical staff at Bedrock Robotics, 37.9% at The Bot Company, and 30.0% at Tesla Optimus come from autonomous driving. The top three feeders are Cruise, Zoox, and Waymo — all major sources of AV talent released in recent years. At hardware-centric companies (Boston Dynamics, Apptronik), the AV share is in the single digits.
华人占技术池 14.1%，但在基础模型/研究层高度集中：NVIDIA GEAR 28.8%、Tesla Optimus 30.0%、DeepMind Robotics 22.0%、1X 20.8%、Figure 17.0%；硬件型团队华人占比明显更低。华人是这条赛道的「造脑」主力。Chinese talent makes up 14.1% of the technical pool, but is highly concentrated in the foundation-model / research layer: 28.8% at NVIDIA GEAR, 30.0% at Tesla Optimus, 22.0% at DeepMind Robotics, 20.8% at 1X, and 17.0% at Figure; the Chinese share at hardware-centric teams is markedly lower. Chinese engineers are the core brain-builders of this sector.
以下基于 2025-2026 公开信源（完整清单见研究底稿），只保留影响人才判断的事实。估值为报道/确认口径，未核实处已标注。The following draws on public sources from 2025-2026 (full list in the research file), keeping only facts that bear on talent judgments. Valuations follow reported / confirmed figures, with unverified items flagged.
2025 全年机器人融资约 $400 亿级。确认口径：Figure $39B（2025-09）、Skild $14B（2026-01，七个月三倍）、Apptronik $5B（2026-02）、Physical Intelligence $5.6B（2025-11，在谈 $11B）。NVIDIA 几乎投了所有头部具身公司，同时自做 GR00T 基础模型与 Isaac 仿真平台，是「卖铲子 + 造模型」双线。Robotics funding in 2025 ran on the order of $40B for the year. Confirmed figures: Figure $39B (2025-09), Skild $14B (2026-01 — a 3x jump in seven months), Apptronik $5B (2026-02), Physical Intelligence $5.6B (2025-11; $11B reportedly under discussion). NVIDIA has invested in nearly every leading embodied-AI company while building its own GR00T foundation model and Isaac simulation platform — a dual play of "selling shovels + building models."
Figure 02 在 BMW 南卡工厂累计运行 1,250+ 小时；Agility Digit、Apptronik Apollo 在 GXO/仓储测试；Boston Dynamics 电动 Atlas 2026 量产机队已被现代预订。对照之下 Tesla Optimus 截至 2026 年初仍无外部客户、处 R&D 与数据采集阶段。落地差异直接决定各家对不同职能（量产工程 vs 研究）的需求结构。Figure 02 has logged 1,250+ cumulative hours at BMW's South Carolina plant; Agility's Digit and Apptronik's Apollo are in testing at GXO / warehousing; Boston Dynamics' electric Atlas production fleet for 2026 has been pre-ordered by Hyundai. By contrast, Tesla Optimus still had no external customers as of early 2026 and remains in the R&D and data-collection stage. These deployment differences directly shape each company's demand structure across functions (production engineering vs. research).
主流是视觉-语言-动作（VLA）模型 + 模仿学习 + 遥操作数据采集，世界模型与「真实经验强化学习」正在上升（Physical Intelligence π0.6 引入真实环境 RL）。基础模型路线（PI / Skild / NVIDIA GR00T / Gemini Robotics）与自造本体路线（Figure / Tesla / Apptronik / 1X）并行。The mainstream is vision-language-action (VLA) models + imitation learning + teleoperated data collection, with world models and "reinforcement learning from real-world experience" on the rise (Physical Intelligence's π0.6 introduces real-environment RL). The foundation-model path (PI / Skild / NVIDIA GR00T / Gemini Robotics) runs in parallel with the build-your-own-hardware path (Figure / Tesla / Apptronik / 1X).
四条进水管：自动驾驶（感知 / 规划 / 数据闭环 / sim2real 技能直接迁移，最成建制）、学术界（CMU / Stanford / Berkeley / UW 机器人实验室）、大厂 ML/AI、传统机器人（造硬件的那一半）。具身是少数同时大量需要「软件大脑」与「硬件身体」两类人的赛道。Four feeders: autonomous driving (perception / planning / data flywheel / sim2real skills transfer directly; the most fully-formed), academia (robotics labs at CMU / Stanford / Berkeley / UW), tech-giant ML/AI, and traditional robotics (the half that builds the hardware). Embodied AI is one of the few sectors that needs large numbers of both "software brain" and "hardware body" people at the same time.
统计对象 = 1,044 名技术人才（研究 / 基础模型 / 操作 / 运动控制 / 感知 / 仿真 / 硬件 / 机器人软件）。Population = 1,044 technical staff (research / foundation models / manipulation / motion control / perception / simulation / hardware / robotics software).
读数：Boston Dynamics 与 Apptronik 的可见技术池最大（团队规模大、含大量硬件岗）；而估值最高的纯 AI 人形/基础模型公司（Physical Intelligence、Skild、The Bot Company、Dexterity）可见技术人各只有数十名。「估值与人头倒挂」在这张图上一目了然，也解释了为什么这条赛道的每一个资深人选都被反复争夺。Read: Boston Dynamics and Apptronik have the largest visible technical pools (big teams with many hardware roles), while the highest-valued pure-AI humanoid / foundation-model companies (Physical Intelligence, Skild, The Bot Company, Dexterity) each show only a few dozen visible technical staff. The "valuation-vs-headcount inversion" is plain to see in this chart, and it explains why every senior candidate in this sector is fought over again and again.
读数：软件/AI 综合最大，但硬件/机电以 26.4% 成为第二大职能，叠加运动控制、机电系统，造身体的工程师占近三成（后两者硬件岗近乎为零，这是具身的结构指纹）。明确从事基础模型/VLA、机器人学习/RL、操作的人极少（各数人到数十人），是全市场最稀缺、最难寻的尖端画像。Read: software/AI (general) is the largest, but hardware/mechatronics is the second-largest function at 26.4%; add motion control and mechatronic systems, and the engineers who build the body account for nearly 30% (against the near-zero hardware roles at the latter two — AV and AI labs — this is embodied AI's structural fingerprint). The people explicitly working on foundation models/VLA, robot learning/RL, and manipulation are very few (from a handful to a few dozen each), the scarcest and hardest-to-find frontier profiles in the entire market.
读数：基础模型/研究型主体（NVIDIA GEAR、DeepMind Robotics、Tesla Optimus）普遍在华人浓度均值线以上，硬件型主体（Boston Dynamics、Apptronik、Agility）在均值线下。华人在「大脑」层显著富集，在「身体」层偏少。Read: foundation-model / research players (NVIDIA GEAR, DeepMind Robotics, Tesla Optimus) generally sit above the Chinese-density average line, while hardware-centric players (Boston Dynamics, Apptronik, Agility) sit below it. Chinese talent is markedly enriched in the "brain" layer and thinner in the "body" layer.
读数：这张图把「家底」拆到职能格。硬件/机电的深色集中在 Boston Dynamics、Apptronik、Agility、1X 这类自造本体的公司；研究科学家与基础模型的人手则集中在 NVIDIA GEAR、DeepMind Robotics、Physical Intelligence。换句话说，想挖「造身体」的人要去硬件本体公司，想挖「造大脑」的人要去大厂研究组与基础模型独角兽，两个池子几乎不重叠。Read: this chart breaks the "assets" down to the function-cell level. The dark hardware/mechatronics cells cluster at build-your-own-hardware companies like Boston Dynamics, Apptronik, Agility, and 1X; the research scientists and foundation-model people cluster at NVIDIA GEAR, DeepMind Robotics, and Physical Intelligence. In other words, to poach the "body-builders" you go to the hardware-platform companies, and to poach the "brain-builders" you go to the tech-giant research groups and foundation-model unicorns — the two pools barely overlap.
读数：右上方的 Boston Dynamics、NVIDIA GEAR（中位 22 个月、超四分之一已过 4 年 cliff）是「成熟团队」，人才进入主动看机会的窗口更多；左下方的 The Bot Company、Bedrock、1X（中位 9-10 个月、几乎无人过 cliff）仍处组建蜜月期，团队黏性高。成立时间与团队成熟度，直接决定了一家公司「现在好不好谈人」。Read: top-right, Boston Dynamics and NVIDIA GEAR (median 22 months, more than a quarter past the 4-year cliff) are "mature teams," where more talent has entered the window of actively looking; bottom-left, The Bot Company, Bedrock, and 1X (median 9-10 months, almost no one past the cliff) are still in the team-building honeymoon, with high retention. Founding date and team maturity directly determine how easy it is to talk to a company's people right now.
具身的人才主要从四处来：自动驾驶、学术界、大厂 ML 与传统机器人。其中自动驾驶最成建制、技能最对口，也是这条赛道与自动驾驶人才市场连续的地方。Embodied-AI talent comes mainly from four places: autonomous driving, academia, tech-giant ML, and traditional robotics. Of these, autonomous driving is the most fully-formed and the best skills match, and it is where this sector connects seamlessly with the AV talent market.
读数：来源高度分散，印证这是一个「从四面八方抽人」的新赛道。自动驾驶是其中最成建制、技能最对口的一条，它供给的不是数量最大，而是质量最高、整队迁移的「大脑」层人才（见 4.2）。学术界直供与大厂 ML 紧随，传统机器人则主要供给硬件层。Read: the sources are highly dispersed, confirming this is a new sector "pulling people in from all directions." Autonomous driving is the most fully-formed and best-matched of them; what it supplies is not the largest in number but the highest-quality, team-level migrating "brain" talent (see 4.2). Direct supply from academia and tech-giant ML follow closely, while traditional robotics mainly supplies the hardware layer.
读数：这张图是整份报告最有信息量的一张。Bedrock Robotics（46.0%）、The Bot Company（37.9%）、Tesla Optimus（30.0%）几乎「长在自动驾驶基因上」：Bedrock 是前 Waymo 卡车团队、The Bot Company 是前 Cruise 班底、Optimus 直接复用 Tesla FSD 的 AI 栈。而 Boston Dynamics、Apptronik、DeepMind Robotics 的 AV 占比仅个位数，它们或是硬件传统、或是大厂研究血统。Read: this is the most information-rich chart in the whole report. Bedrock Robotics (46.0%), The Bot Company (37.9%), and Tesla Optimus (30.0%) are almost "built on autonomous-driving DNA": Bedrock is the former Waymo trucking team, The Bot Company is the former Cruise crew, and Optimus directly reuses Tesla FSD's AI stack. By contrast, Boston Dynamics, Apptronik, and DeepMind Robotics have single-digit AV shares — they are either hardware-heritage or tech-giant-research lineage.结论：自动驾驶喂的是具身的「软件大脑」，不是「硬件身体」。Conclusion: autonomous driving feeds embodied AI's "software brain," not its "hardware body."
读数：Cruise、Zoox、Waymo 是前三大输送方，正是自动驾驶近年「出清/收缩」的几家（Cruise 关停、Zoox 商业化迟缓）。那批被释放或主动求变的自动驾驶人才，相当一部分的下一站就是具身：智驾的「流出」就是具身的「流入」。Read: Cruise, Zoox, and Waymo are the top three feeders — precisely the AV companies that have been "clearing out / contracting" in recent years (Cruise shut down, Zoox slow to commercialize). A sizable share of that released or change-seeking AV talent has embodied AI as its next stop: autonomous driving's "outflow" is embodied AI's "inflow."
读数：这是自动驾驶到具身的人才流「接线图」。最粗的一条是 Waymo → Bedrock Robotics（前 Waymo 卡车团队整建制创业），其次 Cruise → The Bot Company（前 Cruise 班底）。Cruise 一家就同时向 The Bot Company、Agility、Bedrock、NVIDIA GEAR 多处输送，是名副其实的「具身人才摇篮」。这些具名流向，把抽象的「8.2% 有 AV 背景」还原成了可逐条追溯的真实团队迁移。Read: this is the "wiring diagram" of talent flowing from autonomous driving into embodied AI. The thickest line is Waymo → Bedrock Robotics (the former Waymo trucking team spun out as a whole), followed by Cruise → The Bot Company (the former Cruise crew). Cruise alone feeds The Bot Company, Agility, Bedrock, and NVIDIA GEAR at once — a true "cradle of embodied-AI talent." These named flows turn the abstract "8.2% with an AV background" into real, traceable team migrations.
读数：现存技术人才里，2025 年入职数是 2023 年的 3.4 倍，这个赛道的团队主体几乎都是最近两年搭起来的。自动驾驶来源（紫色）正是在 2024-2025 这波里显著加厚，与 Cruise 关停、Zoox/Waymo 收缩的时间线吻合：具身的招兵买马，踩着自动驾驶出清的节奏。Read: among today's technical staff, 2025 hires are 3.4x those of 2023 — the bulk of this sector's teams were built almost entirely in the last two years. The autonomous-driving source (purple) thickens markedly in this 2024-2025 wave, matching the timeline of Cruise's shutdown and Zoox/Waymo's contraction: embodied AI's hiring spree is keeping step with the AV clear-out.
13 家主体技术池中有华人 147 人（高置信 137），占 14.1%，且高度集中在基础模型与研究层。The technical pools of the 13 players include 147 Chinese (137 high-confidence), 14.1% of the total, and highly concentrated in the foundation-model and research layers.
读数：华人浓度最高的全是「大脑」型主体，NVIDIA GEAR、Tesla Optimus、DeepMind Robotics、1X、Figure。基础模型与机器人学习层的华人密度，与自动驾驶、前沿 AI Lab 两个领域的格局一致：华人是 AI 三条赛道（智驾 / Lab / 具身）共同的造脑主力。Read: the highest Chinese-density players are all "brain"-type — NVIDIA GEAR, Tesla Optimus, DeepMind Robotics, 1X, Figure. The Chinese density in the foundation-model and robot-learning layers mirrors the pattern in both autonomous driving and frontier AI labs: Chinese engineers are the shared core brain-builders across all three AI sectors (autonomous driving / labs / embodied).
读数：华人池 PhD 占比 20.4%（全池 15.9%）。上交、台大、浙大、清华是主要本科来源，研究生集中于 Stanford / Berkeley / CMU / UCSD 等机器人重镇。其中 14 名华人具身人才（占华人池 9.5%）有自动驾驶履历，是 AV→具身迁移在华人群体中的镜像。Read: the Chinese pool's PhD share is 20.4% (full pool 15.9%). SJTU, NTU (Taiwan), Zhejiang, and Tsinghua are the main undergraduate feeders, with graduate study concentrated at robotics strongholds like Stanford / Berkeley / CMU / UCSD. Of these, 14 Chinese embodied-AI staff (9.5% of the Chinese pool) have an autonomous-driving track record — a mirror of the AV→embodied migration within the Chinese cohort.
北美造脑层North American brain-building layer：Jim Fan（范麟熙，NVIDIA GEAR 联合负责人）、Yuke Zhu（朱玉可，UT Austin + NVIDIA GEAR）是具身基础模型方向声量最高的华人；Stanford 吴佳俊、UCSD 王小龙（其创业公司 ARI 2026-05 被 Meta 收购、本人任 Meta 超智实验室具身研究总监）、UCSD 苏昊（2026-04 全职回国加盟复旦、筹建通用物理智能研究院）是学术枢纽。: Jim Fan (Linxi Fan, co-lead of NVIDIA GEAR) and Yuke Zhu (UT Austin + NVIDIA GEAR) are the highest-profile Chinese voices in embodied foundation models; Jiajun Wu (Stanford), Xiaolong Wang (UCSD — whose startup ARI was acquired by Meta in 2026-05 — with Wang now Director of Embodied Research at Meta's Superintelligence Lab), and Hao Su (UCSD — who returned to China full-time in 2026-04 to join Fudan and build the Institute for General Physical Intelligence) are the academic hubs.
AV→具身的华人迁移The AV→embodied Chinese migration：与本报告数据呼应的中国侧标志案例，李力耘（小鹏自动驾驶一号位 → 众擎机器人 CTO，2026-04）、星海图「清华 + Waymo 天团"（4 位创始人 3 人有 Waymo/Momenta 履历）、余轶南（地平线 → 维他动力）。北美与中国两侧，自动驾驶华人正同步流入具身。: China-side landmark cases echoing this report's data — Liyun Li (head of XPeng autonomous driving → CTO of EngineAI, 2026-04), Galaxea AI's "Tsinghua + Waymo dream team" (3 of its 4 founders have Waymo/Momenta track records), and Yinan Yu (Horizon Robotics → Vita Power). On both the North American and Chinese sides, AV-trained Chinese engineers are flowing into embodied AI in tandem.
从 1,930 人画像中按级别、方向稀缺度与履历强度精选三组共 20 人。档案来自 Metix AI 数据库；标「公开核实」者已对照 2025-2026 公开信源确认现职。「AV 背景」标签标示有自动驾驶履历者。公开版人名默认模糊。From 1,930 profiles, three groups totaling 20 people were hand-picked by seniority, scarcity of specialty, and strength of track record. Profiles come from the Metix AI database; those tagged "publicly verified" have had their current role confirmed against 2025-2026 public sources. The "AV background" tag marks those with an autonomous-driving track record. Names are masked by default in the public version.
具身赛道薪酬分层极端。数字为 2026-06 检索的报道/levels.fyi/H1B 口径，非公司官方。Pay tiers in embodied AI are extreme. Figures follow reported / levels.fyi / H1B sources retrieved 2026-06; not official company data.
| 主体Player | 工程师参考 TCEngineer reference TC | 顶尖研究岗Top research roles | 股权特征Equity profile | 备注Notes |
|---|---|---|---|---|
| NVIDIA GEAR | 机器人岗 ~$333KRobotics roles ~$333K | 基础模型研究 $400K+Foundation-model research $400K+ | NVIDIA 上市股票NVIDIA public stock | 造脑层，对标 AI LabBrain-building layer, benchmarked to AI labs |
| Physical Intelligence | 研究员 $300-475K+Researchers $300-475K+ | 早期股权杠杆大Heavy early-stage equity leverage | 独角兽期权（$5.6B）Unicorn options ($5.6B) | 团队精干、人均稀缺Lean team, scarce per head |
| Figure AI | H1B 中位 ~$230KH1B median ~$230K | Helix 研究更高Helix research higher | $39B 估值期权 + $100M 员工回购$39B-valuation options + $100M employee buyback | 用回购留人Retaining people via buyback |
| Tesla Optimus | 复用 Tesla 体系Reuses the Tesla system | n/a | Tesla 股票Tesla stock | 低 base + 股票杠杆Low base + stock leverage |
| Apptronik / Agility | ~$122-135K | n/a | 私司期权Private-company options | 硬件岗 + 地域，显著低于造脑层Hardware roles + location, markedly below the brain-building layer |
| 遥操作 / 数据采集Teleoperation / data collection | $25-35/小时$25-35/hour | n/a | 无None | 赛道薪酬另一极The other pole of the sector's pay |
顶端机器人基础模型研究员的总包（$300-475K+ 股权）与底端遥操作数据采集员（时薪 $25-35）相差一个数量级以上；硬件型公司的工程岗（约 $122-135K）也明显低于造脑层。同一份「具身人才」报告里，存在三个几乎不相干的薪酬市场。The total package for a top robotics foundation-model researcher ($300-475K+ in equity) differs from a bottom-tier teleoperated data collector ($25-35/hour) by more than an order of magnitude; engineering roles at hardware-centric companies (around $122-135K) are also clearly below the brain-building layer. Within a single "embodied-AI talent" report there are three almost unrelated pay markets.
造脑层（基础模型 / 机器人学习 / 感知）与自动驾驶、AI Lab 抢同一批人，薪酬被抬到接近 AI Lab 水平；硬件 / 机电层供给相对独立（来自传统机器人、车企、航空航天），竞价压力小得多。招聘预算应按这两个市场分别定价。The brain-building layer (foundation models / robot learning / perception) competes with autonomous driving and AI labs for the same people, pushing pay close to AI-lab levels; the hardware / mechatronics layer has relatively independent supply (from traditional robotics, automakers, and aerospace), with far less bidding pressure. Recruiting budgets should be priced separately for these two markets.
同一组数据，对猎头、机器人公司 HR、VC 各有不同含义。The same dataset means different things to headhunters, robotics-company HR, and VCs.
① 这是个「薄而贵」的市场，精准远胜广撒：造脑层（基础模型 / 机器人学习 / 操作）总量仅数十到百余人，值得逐人经营；② 自动驾驶是最对口的相邻池，自动驾驶「出清 / 收缩」公司（Cruise / Zoox / Motional 系）的感知 / 规划 / sim2real 人才，是具身造脑岗的现成来源；③ 硬件 / 机电岗要去传统机器人、车企、航空航天找，与造脑岗完全分轨；④ 华人造脑人才循校友与院校脉络（上交 / 清华 / 浙大 + Stanford / Berkeley / UCSD / CMU）最易触达。① This is a "thin but expensive" market where precision beats broad outreach: the brain-building layer (foundation models / robot learning / manipulation) totals only a few dozen to just over a hundred people, worth cultivating one by one; ② autonomous driving is the best-matched adjacent pool — the perception / planning / sim2real talent from "clearing-out / contracting" AV companies (the Cruise / Zoox / Motional camp) is a ready-made source for embodied brain-building roles; ③ hardware / mechatronics roles must be sourced from traditional robotics, automakers, and aerospace, on a completely separate track from brain-building roles; ④ Chinese brain-building talent is most reachable along alumni and university lines (SJTU / Tsinghua / Zhejiang + Stanford / Berkeley / UCSD / CMU).
① 对照本报告定位自家在规模 × 浓度图上的位置；② 造脑岗与 AV / AI Lab 同台竞价，差异化靠「算法直接驱动真实身体」的兑现感与早期股权；③ 硬件岗供给独立、竞争缓和，是更可控的扩张面；④ 注意赛道高频流动（「其他 / 创业」是最大来源），留人需要清晰的技术路线与里程碑。① Use this report to locate where you sit on the scale × density chart; ② brain-building roles bid head-to-head against AV / AI labs, and your differentiation rests on the tangible payoff of "algorithms directly driving a real body" plus early-stage equity; ③ hardware roles have independent supply and gentler competition, a more controllable surface to expand on; ④ mind the sector's high churn ("other / startups" is the largest source) — retention needs a clear technical roadmap and milestones.
① 团队的自动驾驶基因是一个可量化的尽调信号：Bedrock / The Bot Company / Optimus 的高 AV 占比对应「成建制、有数据闭环经验「的团队；② 估值与人才厚度的倒挂提示风险，数十人团队支撑数十亿估值，关键人离职冲击大；③ 学术创始团队（CMU / Stanford / Berkeley 系）与产业团队（AV 系）是两类不同的下注，前者强研究、后者强工程落地。① A team's autonomous-driving DNA is a quantifiable due-diligence signal: the high AV shares at Bedrock / The Bot Company / Optimus correspond to teams that are "fully formed, with data-flywheel experience"; ② the inversion of valuation against talent depth flags risk — a few-dozen-person team supporting a multibillion-dollar valuation takes a heavy hit if a key person leaves; ③ academic founding teams (the CMU / Stanford / Berkeley camp) and industry teams (the AV camp) are two different bets, the former strong on research, the latter strong on engineering deployment.
本报告的检索、画像与迁移分析全部由 Metix AI 完成。可按同样口径为任意公司生成定制人才地图：全量名单、来源迁移网络、邮箱解锁与多渠道联系，并按「只为合格面试付费」计费。No interview, no charge.The search, profiling, and migration analysis in this report were all done by Metix AI. We can build a custom talent map for any company on the same methodology: a full list, a source-migration network, email unlocking, and multi-channel outreach — billed on a "pay only for qualified interviews" basis. No interview, no charge.
8.6 亿+ 全球人才画像860M+ global talent profiles1,044 人具身技术池1,044-person embodied technical poolAV→具身迁移监测AV→embodied migration tracking只为合格面试付费Pay only for qualified interviews纯具身创业：Physical Intelligence、Figure AI、Skild AI、1X Technologies、Apptronik、Agility Robotics、Dexterity、The Bot Company、Bedrock Robotics、Boston Dynamics。大厂内部机器人组：Tesla Optimus、Google DeepMind Robotics、NVIDIA GEAR。后三者为按机器人职能关键词识别的子集，非整公司。地理以美国为主，含 1X 挪威等。Pure-play embodied startups: Physical Intelligence, Figure AI, Skild AI, 1X Technologies, Apptronik, Agility Robotics, Dexterity, The Bot Company, Bedrock Robotics, Boston Dynamics. In-house robotics groups at the tech giants: Tesla Optimus, Google DeepMind Robotics, NVIDIA GEAR. The latter three are subsets identified by robotics-function keywords, not whole companies. Geography is mainly the US, including 1X in Norway and others.
技术人才池 = 基础模型/VLA、机器人学习/RL、操作、运动控制、感知、仿真、teleop数据、硬件/机电、机器人软件、研究科学家、软件AI综合、创始人/高管。华人识别采用五信号交叉验证（姓名族裔模型 / 汉字 / 中文 / 中国院校 / 多拼写姓氏库），分高/中置信，主口径 = 高 + 中。Technical talent pool = foundation models/VLA, robot learning/RL, manipulation, motion control, perception, simulation, teleop data, hardware/mechatronics, robotics software, research scientists, software/AI (general), founders/executives. Chinese-talent identification uses five cross-validated signals (name-ethnicity model / Chinese characters / Chinese text / Chinese universities / multi-romanization surname library), split into high / medium confidence, with the headline methodology = high + medium.
"有自动驾驶背景" = 历史雇主含 Waymo/Cruise/Zoox/Nuro/Aurora/Argo/Motional/Pony.ai/WeRide/TuSimple/Kodiak/Gatik/Torc/Wayve 等。Tesla 因兼具车与机器人业务，不计入严格 AV 判定，以免高估。"Has an autonomous-driving background" = past employers include Waymo/Cruise/Zoox/Nuro/Aurora/Argo/Motional/Pony.ai/WeRide/TuSimple/Kodiak/Gatik/Torc/Wayve and others. Tesla, given that it spans both automotive and robotics, is excluded from the strict AV determination to avoid overcounting.
数据截至 2026 年上半年；AV→具身迁移的源头数据基于同一人才库。重点人物已逐人对照公开信息核实。Data is current through H1 2026; the source data for AV→embodied migration is based on the same talent pool. Key figures have been verified one by one against public information.
| 主体Player | 在职画像Current profiles | 技术池Technical pool | 华人Chinese | 华人占比Chinese share | AV 背景占比AV-background share | PhD 率PhD rate |
|---|---|---|---|---|---|---|
| Boston Dynamics | 455 | 297 | 27 | 9.1% | 2.0% | 12.5% |
| Apptronik | 316 | 146 | 11 | 7.5% | 2.7% | 15.1% |
| Agility Robotics | 274 | 131 | 10 | 7.6% | 6.1% | 13.7% |
| 1X Technologies | 227 | 101 | 21 | 20.8% | 4.0% | 7.9% |
| Figure AI | 200 | 94 | 16 | 17.0% | 8.5% | 22.3% |
| NVIDIA GEAR | 90 | 73 | 21 | 28.8% | 12.3% | 21.9% |
| DeepMind Robotics | 86 | 59 | 13 | 22.0% | 1.7% | 35.6% |
| Bedrock Robotics | 82 | 50 | 4 | 8.0% | 46.0% | 14.0% |
| Tesla Optimus | 47 | 30 | 9 | 30.0% | 30.0% | 26.7% |
| The Bot Company | 81 | 29 | 9 | 31.0% | 37.9% | 3.4% |
| Skild AI | 27 | 15 | 4 | 26.7% | 6.7% | 20.0% |
| Physical Intelligence | 28 | 11 | 2 | 18.2% | 18.2% | 36.4% |
| Dexterity | 17 | 8 | 0 | 0.0% | 0.0% | 0.0% |
① 本报告基于公开职业档案聚合，规模型与维护 LinkedIn 充分的团队覆盖度更高；纯研究小团队（Physical Intelligence / Skild / Dexterity）样本偏小，其绝对数宜理解为下限。② 大厂机器人组为职能识别子集，与公司真实机器人编制存在偏差。③ 职能与层级基于 title/headline 关键词推断。④ 华人识别为概率判定，主口径 147 人 = 高置信 137 + 中置信 10。⑤ 各主体之间比较以占比与结构为主、绝对数为辅。① This report aggregates public professional profiles, so larger teams and those that maintain LinkedIn thoroughly have higher coverage; small pure-research teams (Physical Intelligence / Skild / Dexterity) have smaller samples, and their absolute counts should be read as lower bounds. ② The tech-giant robotics groups are function-identified subsets and diverge from the companies' true robotics org charts. ③ Function and level are inferred from title/headline keywords. ④ Chinese-talent identification is a probabilistic determination, with the headline methodology of 147 = 137 high-confidence + 10 medium-confidence. ⑤ Comparisons across players rely mainly on shares and structure, with absolute counts secondary.