星动纪元 / 加速进化 / 逐际动力整合研究

1. 介绍

本文将三家国产人形机器人新势力整合为一篇：星动纪元（RobotEra）、加速进化（Booster Robotics）、逐际动力（LimX Dynamics）。三家几乎同时成立（2022—2023），创始人高度学术化（清华陈建宇、清华赵明国团队、南科大张巍），但选择了三条可辨识的技术与商业路线——这正是本篇的组内差异主线：新势力三种技术路线。

1.1. 整合篇说明与三家画像

公司成立创始团队一句话画像2026 资本状态
星动纪元2023-08陈建宇（清华交叉信息研究院 / 伯克利博士）软硬件全栈自研 + 端到端 VLA，物流为商业化第一站2026 年两月融资近 25 亿元，估值破百亿；港股 IPO 为传闻
加速进化2023-06赵明国（清华自动化系仿人机器人团队），CEO 程昊面向开发者与赛事的平台型本体，「以赛代练、以赛促研」2026-04 B 轮约 10 亿元，进入独角兽行列
逐际动力2022张巍（南科大长聘教授）System 0/1/2 三层架构 + COSA 智能体操作系统，聚焦工业与特种半年融资 4 亿美元，Pre-IPO 投后估值 150 亿元

三家均已传出港股 IPO 动向（逐际完成股份制改造并传保密递表、星动被传冲刺港股），截至信息截止日均为报道 / 传闻口径（D 级），本文不作为事实写入。

1.2. 三种技术路线总览

  1. 星动纪元——全栈自研 + 世界模型路线：本体（L7/Q5/M7）与模型（ERA-42）全栈自研，主张「VLA 非终局，具身大脑进入世界模型时代」（陈建宇，WRC 2026）。
  2. 加速进化——开发者平台路线：把本体做成标准化的赛事与科研平台（RoboCup 类赛事常用机型），以社区与比赛驱动迭代，不与大厂争夺产业场景。
  3. 逐际动力——分层架构路线：以 System 0（全身运动基础模型）/ System 1（VLA/WAM）/ System 2（COSA 智能体操作系统）三层抽象组织全部产品与研发，先攻工业与特种的高价值场景。

2. 名词解释

术语英文 / 缩写释义
ERA-42—星动纪元自研端到端 VLA 大模型（官方称国内首个），具备世界模型能力（从人类视频学技能）
分频 VLAFrequency-split VLA星动提出的快慢分层概念：低频层负责规划、高频层负责实时动作
星动 L7—星动全尺寸双足旗舰（171cm / 55 自由度 / 膝峰值扭矩 400N·m）
XHAND 1—星动全直驱灵巧手（12 自由度、10 次每秒），有 Lite / PRO 变体
双芯片架构Dual-chip星动 L7 的 80 + 275 TOPS 双算力芯片设计，分别服务实时控制与模型推理
Booster T1 / T2—加速进化开发者 / 赛事平台人形；T2 标配英伟达 Thor 芯片（2070 TFLOPS、最高 75 自由度）
以赛代练Competition-driven Iteration以机器人赛事（RoboCup 类）作为训练、验证与社区运营载体的路线
System 0/1/2—逐际三层体系：全身运动基础模型 / 人形 VLA 与 WAM / 具身智能体操作系统
COSA—逐际以大语言模型 + 世界模型为引擎的具身智能体操作系统（System 2）
VGM—逐际「具身智能训练新范式」算法体系（官网在列）
DreamActor—逐际具身机器人操作算法体系（官网在列）
TRON—逐际多形态双足机器人系列（TRON 1 / TRON 2），科研与强化学习常用平台
全直驱Direct Drive灵巧手驱动方式，XHAND 1 采用全直驱以提升响应与力控精度
RoboCup—国际机器人足球及机器人学赛事体系，加速进化 T1 的主要应用场景之一

3. 功能说明与产品线

3.1. 星动纪元：全尺寸人形 + ERA-42

型号形态关键参数
星动 L7全尺寸双足171cm / 65kg / 55 自由度 / 膝峰值扭矩 400N·m·25rad/s / 跑速 4m/s / 双臂负载 20kg / 2.1m 直径操作空间 / 双芯片 80+275 TOPS
星动 Q5服务人形面向服务场景
星动 M7半身 + 立柱2026Q2 启动千台级批量交付
XHAND 1 / Lite / PRO灵巧手全直驱 12 自由度，10 次每秒动作

ERA-42 具备世界模型能力（看人类视频学技能）；2025-08 世界人形机器人运动会上 L7 以原地跳高 95.641cm 夺冠。

3.2. 加速进化：Booster T1 / T2 开发者平台

Booster T1 为开发者与赛事平台人形，是多支高校 RoboCup 类赛事常用平台（口径待补）。Booster T2 于 2026-07-10 发布：标配英伟达 Thor 芯片（2070 TFLOPS）、全身最高 75 自由度、持续行走续航超 2 小时。T1 具体价格与出货量缺失，标注 [待填写]。

3.3. 逐际动力：TRON / Luna / oli

TRON 1（多形态双足，强化学习入门平台）、TRON 2（多形态具身）、Luna（全尺寸交互人形，2026-05-25 发布）、oli（全尺寸通用）。官网另列 VGM（具身智能训练新范式）与 DreamActor（具身机器人操作算法）两大算法体系。逐际具名客户案例缺失，标注 [待填写]。

4. 平台架构

4.1. 星动纪元：分频 VLA 与世界模型路线

星动的架构叙事围绕 ERA-42 展开：端到端 VLA 之上叠加世界模型能力（从人类视频学习技能），并以「分频 VLA」解决实时性——低频层做任务规划与语义决策，高频层做实时动作生成，双芯片（80 + 275 TOPS）分别承载。该架构与傅利叶双路认知（见 05-fourier.md）、银河 WAM（见 06-galbot.md）同属「快慢分层」家族，但星动的分层落在 VLA 模型内部。

4.2. 加速进化：以赛代练的工程闭环

加速进化的架构重心不在大模型叙事，而在标准化本体 + 可复现训练环境：T1/T2 作为赛事平台，规则固定、场地标准、任务可量化，天然构成一个公开的回归测试集——比赛成绩即评测分数。T2 标配 Thor（2070 TFLOPS）把模型推理算力提升到可跑大模型的档位，暗示其正在从「赛事平台」向「开发者 AI 平台」升级。

4.3. 逐际动力：System 0/1/2 三层体系

逐际的三层体系是本组与 Harness 六层模型映射最工整的架构：System 0 全身运动基础模型对应 L2 执行；System 1 人形 VLA / WAM 对应 L1 感知-语言-动作；System 2 COSA（LLM + 世界模型引擎的具身智能体操作系统）对应 L1/L3 的任务编排与智能体运行时。COSA 0.5 已于 2026-07-15 官网更新，显示该体系进入快速迭代期。

4.4. 三家架构对比图

图 7-1｜三家技术路线与架构分层对比

新势力三种技术路线（信息截止 2026-09-12） 星动纪元 · 全栈自研 ERA-42 端到端 VLA 世界模型 · 看视频学技能 分频 VLA 快慢分层 双芯片 80+275 TOPS 本体：L7 / Q5 / M7 场景：物流 / 零售 / 科研 2025Q3 交付 400+ 台 海外收入占比超 50% 加速进化 · 开发者平台 以赛代练 · 以赛促研 赛事 = 公开回归测试集 T2：Thor 2070 TFLOPS 最高 75 自由度 本体：T1 / T2 场景：教育 / 赛事 / 科研 出货量 [待填写] B 轮 10 亿元（2026-04） 逐际动力 · 三层架构 System 2 · COSA 操作系统 LLM + 世界模型引擎 System 1 · VLA / WAM System 0 · 全身运动基础模型 本体：TRON 1/2 · Luna · oli 场景：工业 / 特种（推进中） 具名客户 [待填写] 半年融资 4 亿美元 共同点：学者创业（清华 / 南科大）、2022—2023 成立、2026 年均传出港股 IPO（报道口径，D 级） 差异点：数据策略（真机交付反哺 / 赛事环境 / 实验室算法）与商业化第一场景各不相同

示意：基于本文分析绘制；各机型参数与融资额来自三家公司官网与公开报道，信息截止 2026-09-12。

5. Harness 设计

5.1. L1 上下文工程层

星动：ERA-42 的世界模型能力使视觉上下文可直接转化为技能先验，「看人类视频学技能」是把人类视频作为上下文源的路线。逐际：System 1 的 VLA / WAM 承载感知-语言-动作上下文，COSA 在其上做任务级语义编排。加速进化：赛事规则与场地信息构成标准化的任务上下文，由开发者自定义注入。

5.2. L2 工具与执行层

三家执行层各有侧重：星动 L7 的 55 自由度 + 400N·m 膝扭矩 + 全直驱灵巧手面向高动态作业；加速进化 T2 的 75 自由度（本组自由度最高披露值之一）面向全肢开发者实验；逐际 TRON 的多形态双足（可换装）面向科研可配置性，Luna / oli 补全交互与通用形态。

5.3. L3 编排与控制层

逐际的 COSA 是三家唯一把「智能体操作系统」作为产品明确发布的 L3 级组件（COSA 0.5，2026-07-15）；星动以分频 VLA 隐式承载编排；加速进化依赖开发者自建（其平台定位本就是把 L3 留给用户）。

5.4. L4 记忆与状态层

三家均无运行时记忆机制的公开披露。星动的世界模型（从历史视频学习）与逐际 COSA 的世界模型引擎具备 L4 的理论基础，但落地细节缺失，标注 [待填写]。

5.5. L5 评估与观测层

加速进化最具结构性优势：赛事成绩是天然的可比、可复现评估。星动有运动会跳高冠军（95.641cm）单点成绩与交付数据（400+ 台）；逐际以实验室级算法演示为主，公开量化评估最少。

5.6. L6 治理与安全层

三家公开信息均薄弱，标注 [待填写]。相对可见的只有：逐际聚焦的工业 / 特种场景隐含更高的安全准入要求（其 COSA 架构将治理内化为操作系统职责的设计取向）；加速进化的教育赛事场景风险面较小。

5.7. 三家六层对比小结

层星动纪元加速进化逐际动力
L1 上下文工程★★ 世界模型先验★★ 赛事标准上下文★★★ System 1 + COSA
L2 工具与执行★★★ 55 DOF 高动态★★★ 75 DOF 开放平台★★ 多形态双足
L3 编排与控制★★ 分频隐式承载★ 留给开发者★★★ COSA 操作系统
L4 记忆与状态★★ 理论具备★★★ 理论具备
L5 评估与观测★★ 交付 + 单点成绩★★★ 赛事即评测★ 公开最少
L6 治理与安全★★★ 架构取向可见

6. 实际案例

6.1. 星动纪元：物流第一站与千台级交付

星动把物流作为商业化第一站（官方 PMF 表述）：与北自科技合作物流场景、海尔智家零售场景、MIT 与字节跳动实验室科研场景。2025Q3 累计交付 400+ 台、总订单额超 5 亿元；海外收入占比自 2025 年初以来超 50%；2026Q2 启动 M7 千台级批量交付。官方宣传口径称「全球 TOP10 市值科技巨头中 9 家为其客户」（可信度 C 级，本文标注后引用）。融资节奏密集：2025-11 近 10 亿元 A+（吉利资本领投）→ 2026-03 10 亿元战略轮（三星 / 中金，估值破百亿）→ 2026-07-06 10 亿元（诚通基金领投 + 多家国资），两个月近 25 亿元。

6.2. 加速进化：赛事与开发者生态

加速进化的「案例」以赛事存在感为形态：T1 是多支高校 RoboCup 类赛事常用平台（口径待补），「以赛代练、以赛促研」（浦银国际语）概括其路线。2026-04 完成 B 轮约 10 亿元融资进入独角兽行列后，2026-07-10 发布 T2（Thor 芯片 / 2070 TFLOPS / 最高 75 自由度），完成从赛事平台到高算力开发者平台的跳档。

6.3. 逐际动力：工业特种推进与 Pre-IPO

逐际聚焦工业与特种场景，「加速推动在汽车制造、仓储物流的规模化落地」（进展口径，无具名客户，标注 [待填写]）。官网创新故事展示了双足博物馆导览、山地光伏无人巡检等场景验证。资本侧是其最活跃的维度：2026-02 B 轮 2 亿美元（京东 / 上汽 / 蔚来 / IDG / 联想创投等）→ 2026-07-14 Pre-IPO 近 2 亿美元（IDG / 蓝思科技 / 华山资本 / 阿联酋磊石资本追投），投后估值 150 亿元；半年累计 4 亿美元，约 70% 来自海外机构；2026-03 完成股份制改造。注意：胡润 2026 全球独角兽榜对其估值为 68 亿元，与 Pre-IPO 投后 150 亿元两口径并列，未择一。

7. 总结

7.1. 优势与局限

星动：优势在交付数据（400+ 台）与全栈自研的模型迭代速度；局限在物流场景之外的第二曲线尚在验证。加速进化：优势在开发者生态与赛事评测的结构性优势、B 轮资本到位；局限在离产业场景最远、出货与定价不透明。逐际：优势在三层架构的方法论完整度与半年 4 亿美元的融资效率；局限在缺具名客户案例，落地叙事弱于融资叙事。

7.2. 适用边界

星动适合物流、零售与需要全栈交付的场景客户；加速进化适合高校、赛事组织方与需要开放本体做强化学习研究的开发者；逐际适合关注分层架构方法论、愿意共同定义工业 / 特种场景方案的早期合作者。

7.3. 选型建议

  1. 科研与赛事采购首选加速进化（评测可复现）或逐际 TRON（多形态可配置）；星动 L7 适合需要整机作业能力的研究。
  2. 产业客户应要求三家提供具名客户与现场运行数据——三家均已传出 IPO（D 级），宣传口径需与交付证据分开核验。
  3. 关注 COSA 与 ERA-42 的后续版本：分层操作系统与世界模型是两条可能改变行业架构共识的路线。
  4. 估值信息一律标注口径与日期，逐际 150 亿 vs 68 亿元、星动「估值破百亿」均为融资节点口径，非市场定价。

信息缺口声明

  1. 加速进化 T1 价格与出货量缺失，标注 [待填写]；「RoboCup 类赛事常用平台」为待补口径。
  2. 逐际动力具名客户案例缺失，标注 [待填写]。
  3. 估值两口径冲突：逐际投后 150 亿元（Pre-IPO 轮）vs 胡润榜 68 亿元，并列呈现。
  4. 港股 IPO（逐际保密递表、星动被传冲刺）均为 D 级传闻，仅作状态说明，不作为事实。
  5. 三家 L6 治理与安全公开信息普遍缺失。
  6. 星动「TOP10 市值科技巨头中 9 家为客户」为官方宣传口径（C 级），已标注引用。

8. 参考资料

  1. 星动纪元官网（产品：L7 / Q5 / M7 / XHAND；WRC 2026 新闻）— 星动纪元，2026。https://www.robotera.com/
  2. 星动 L7 产品页（全参数表）— 星动纪元官网，2025。https://www.robotera.com/robot/l7.html
  3. 星动 L7 词条（交付 / 融资 / 运动会成绩）— 百度百科，2026。https://baike.baidu.com/item/%E6%98%9F%E5%8A%A8L7/66239282
  4. 性能比肩特斯拉 Optimus，星动纪元发布星动 L7 — 搜狐科技，2025-07。https://it.sohu.com/a/916431213_362042
  5. 逐际动力 LimX Dynamics 官方网站（TRON / Luna / oli / COSA / VGM / DreamActor）— 逐际动力，2026。https://www.limxdynamics.com
  6. 逐际动力人形大脑系统 COSA 0.5 更新 — 逐际动力官网新闻，2026-07-15。https://www.limxdynamics.com/zh/news/BK000065
  7. 逐际动力完成 Pre-IPO 轮融资 — 逐际动力官网新闻，2026-07-14。https://www.limxdynamics.com/zh/news/BK000063
  8. 985 教授们集体创业做具身智能（星动 / 加速进化 / 逐际 / 穹彻等）— JQman，2026。https://www.jqman.com/ai/521998.html
  9. 2026 年具身智能进入「上市大年」（整机派 / 全栈派盘点）— 搜狐科技，2026。https://it.sohu.com/a/997432229_116132
  10. 【IPO 速递】逐际动力、星动纪元被传冲刺港股 — Mitrade 财华社，2026-08-20。https://www.mitrade.com/zh/insights/news/live-news/article-8-2015872-20260820
  11. AI 机器人周报（Booster T2 发布 / 逐际 Pre-IPO / 星动 10 亿元）— 网易新闻，2026-07。https://c.m.163.com/news/a/L1VPHI9505118I96.html
  12. 星动 L7 发布 — 环球网，2025-07。https://look.huanqiu.com/article/4Nb9s9pnmvh

RobotEra / Booster Robotics / LimX Dynamics: Integrated Research

1. Introduction

This article consolidates three emerging domestic humanoid-robotics players into a single report: RobotEra, Booster Robotics, and LimX Dynamics. All three were founded almost simultaneously (2022—2023) with highly academic founders (Chen Jianyu of Tsinghua University, the Zhao Mingguo team at Tsinghua University, and Zhang Wei of SUSTech), yet they chose three identifiable technical and commercial routes — which is precisely the intra-group differentiation theme of this article: the three technology routes of the emerging players.

1.1. Integrated Report Notes and the Three Company Profiles

CompanyFoundedFounding TeamOne-line Profile2026 Capital Status
RobotEra2023-08Chen Jianyu (Tsinghua University Institute for Interdisciplinary Information Sciences / UC Berkeley PhD)Full-stack in-house software and hardware + end-to-end VLA, with logistics as the first commercialization stopRaised nearly RMB 2.5 billion in two months of 2026, valuation broke RMB 10 billion; HK IPO is rumored
Booster Robotics2023-06Zhao Mingguo (Tsinghua University Department of Automation humanoid-robot team), CEO Cheng HaoPlatform-type embodiment for developers and competitions, "compete as practice, let competitions drive research"Series B of about RMB 1 billion in 2026-04, entering the ranks of unicorns
LimX Dynamics2022Zhang Wei (tenured professor at SUSTech)System 0/1/2 three-layer architecture + COSA agentic operating system, focused on industrial and special-purposeRaised USD 400 million in half a year, post-money valuation of RMB 15 billion in Pre-IPO

All three companies have been reported to be moving toward a HK stocks IPO (LimX Dynamics completed its joint-stock restructuring and is rumored to have filed confidentially; RobotEra is rumored to be sprinting for the HK market); as of the information cutoff date, all are report/rumor-level claims (Level D), which this article does not write in as fact.

1.2. Overview of the Three Technology Routes

  1. RobotEra — full-stack in-house + world-model route: both the embodiment (L7/Q5/M7) and the model (ERA-42) are developed in-house across the full stack, advocating that "VLA is not the endgame; the embodied brain is entering the era of world models" (Chen Jianyu, WRC 2026).
  2. Booster Robotics — developer-platform route: turning the embodiment into a standardized competition and research platform (a common model in RoboCup-style competitions), driving iteration through the community and competitions, and not competing with the majors over industrial scenarios.
  3. LimX Dynamics — layered-architecture route: organizing all products and R&D around three layers of abstraction — System 0 (whole-body motion foundation model) / System 1 (VLA/WAM) / System 2 (COSA agentic operating system) — and first tackling the high-value scenarios of industrial and special-purpose.

2. Glossary

TermEnglish / AbbreviationDefinition
ERA-42—RobotEra's in-house end-to-end VLA large model (officially described as the first in China), with world-model capability (learning skills from human video)
Frequency-split VLAFrequency-split VLARobotEra's fast/slow layering concept: the low-frequency layer handles planning, the high-frequency layer handles real-time action
RobotEra L7—RobotEra's full-size bipedal flagship (171cm / 55 DOF / peak knee torque 400N·m)
XHAND 1—RobotEra's fully direct-drive dexterous hand (12 DOF, 10 actions per second), with Lite / PRO variants
Dual-chip architectureDual-chipRobotEra L7's 80 + 275 TOPS dual-compute chip design, serving real-time control and model inference respectively
Booster T1 / T2—Booster Robotics' developer / competition-platform humanoids; T2 ships with the NVIDIA Thor chip by default (2070 TFLOPS, up to 75 DOF)
Competition-driven iterationCompetition-driven IterationThe route that uses robotics competitions (RoboCup-style) as the vehicle for training, validation, and community operation
System 0/1/2—LimX Dynamics' three-layer system: whole-body motion foundation model / humanoid VLA and WAM / embodied agentic operating system
COSA—LimX Dynamics' embodied agentic operating system (System 2) powered by a large language model + world model
VGM—LimX Dynamics' "new paradigm for embodied intelligence training" algorithm system (listed on the official site)
DreamActor—LimX Dynamics' embodied robot manipulation algorithm system (listed on the official site)
TRON—LimX Dynamics' multi-form bipedal robot series (TRON 1 / TRON 2), a common platform for research and reinforcement learning
Direct driveDirect DriveA dexterous-hand driving method; XHAND 1 adopts full direct drive to improve response and force-control precision
RoboCup—The international robot soccer and robotics competition system, one of the main application scenarios of Booster T1

3. Feature Description and Product Line

3.1. RobotEra: Full-size Humanoid + ERA-42

ModelForm FactorKey Specs
RobotEra L7Full-size bipedal171cm / 65kg / 55 DOF / peak knee torque 400N·m·25rad/s / running speed 4m/s / dual-arm payload 20kg / 2.1m-diameter workspace / dual chips 80+275 TOPS
RobotEra Q5Service humanoidFor service scenarios
RobotEra M7Half body + columnStarted thousand-unit-scale mass delivery in 2026Q2
XHAND 1 / Lite / PRODexterous handFull direct drive, 12 DOF, 10 actions per second

ERA-42 has world-model capability (learning skills from human video); at the 2025-08 World Humanoid Robot Games, L7 won the standing high jump with 95.641cm.

3.2. Booster Robotics: Booster T1 / T2 Developer Platform

Booster T1 is a developer and competition-platform humanoid, and is a commonly used platform for RoboCup-style competitions at multiple universities (wording to be supplemented). Booster T2 was released on 2026-07-10: it ships with the NVIDIA Thor chip by default (2070 TFLOPS), up to 75 DOF across the full body, and over 2 hours of continuous walking range. T1's specific price and shipment volume are missing, marked [To be filled].

3.3. LimX Dynamics: TRON / Luna / oli

TRON 1 (multi-form bipedal, an entry-level platform for reinforcement learning), TRON 2 (multi-form embodied), Luna (full-size interactive humanoid, released 2026-05-25), and oli (full-size general-purpose). The official site also lists two major algorithm systems: VGM (a new paradigm for embodied intelligence training) and DreamActor (an embodied robot manipulation algorithm). LimX Dynamics' named customer cases are missing, marked [To be filled].

4. Platform Architecture

4.1. RobotEra: Frequency-split VLA and World Model Route

RobotEra's architectural narrative revolves around ERA-42: a world-model capability (learning skills from human video) is layered on top of the end-to-end VLA, and real-time performance is solved with the "frequency-split VLA" — the low-frequency layer handles task planning and semantic decisions, the high-frequency layer handles real-time action generation, and the dual chips (80 + 275 TOPS) carry them respectively. This architecture belongs to the same "fast/slow layering" family as Fourier's dual-path cognition (see 05-fourier.md) and Galbot's WAM (see 06-galbot.md), but RobotEra's layering falls inside the VLA model.

4.2. Booster Robotics: The Engineering Closed Loop of Competition-driven Iteration

Booster Robotics' architectural center of gravity is not the large-model narrative but the standardized embodiment + reproducible training environment: as a competition platform, T1/T2 have fixed rules, standardized venues, and quantifiable tasks, naturally forming a public regression test set — competition results are evaluation scores. T2's standard Thor (2070 TFLOPS) lifts model-inference compute into the tier that can run large models, suggesting it is upgrading from a "competition platform" to a "developer AI platform".

4.3. LimX Dynamics: The System 0/1/2 Three-layer System

LimX Dynamics' three-layer system is the most neatly mapped architecture of this group onto the Harness six-layer model: System 0's whole-body motion foundation model corresponds to L2 execution; System 1's humanoid VLA / WAM corresponds to L1 perception–language–action; System 2's COSA (an embodied agentic operating system powered by an LLM + world-model engine) corresponds to L1/L3 task orchestration and the agentic runtime. COSA 0.5 was updated on the official site on 2026-07-15, indicating that this system has entered a fast-iteration phase.

4.4. Architecture Comparison of the Three Companies

图 7-1｜三家技术路线与架构分层对比

新势力三种技术路线（信息截止 2026-09-12） 星动纪元 · 全栈自研 ERA-42 端到端 VLA 世界模型 · 看视频学技能 分频 VLA 快慢分层 双芯片 80+275 TOPS 本体：L7 / Q5 / M7 场景：物流 / 零售 / 科研 2025Q3 交付 400+ 台 海外收入占比超 50% 加速进化 · 开发者平台 以赛代练 · 以赛促研 赛事 = 公开回归测试集 T2：Thor 2070 TFLOPS 最高 75 自由度 本体：T1 / T2 场景：教育 / 赛事 / 科研 出货量 [待填写] B 轮 10 亿元（2026-04） 逐际动力 · 三层架构 System 2 · COSA 操作系统 LLM + 世界模型引擎 System 1 · VLA / WAM System 0 · 全身运动基础模型 本体：TRON 1/2 · Luna · oli 场景：工业 / 特种（推进中） 具名客户 [待填写] 半年融资 4 亿美元 共同点：学者创业（清华 / 南科大）、2022—2023 成立、2026 年均传出港股 IPO（报道口径，D 级） 差异点：数据策略（真机交付反哺 / 赛事环境 / 实验室算法）与商业化第一场景各不相同

示意：基于本文分析绘制；各机型参数与融资额来自三家公司官网与公开报道，信息截止 2026-09-12。

5. Harness Design

5.1. L1 Context Engineering Layer

RobotEra: ERA-42's world-model capability lets visual context be directly converted into skill priors; "learning skills from human video" is the route of treating human video as the context source. LimX Dynamics: System 1's VLA / WAM carries the perception–language–action context, and COSA performs task-level semantic orchestration on top of it. Booster Robotics: competition rules and venue information form a standardized task context, injected as customized by the developer.

5.2. L2 Tools and Execution Layer

The three companies' execution layers each have their emphasis: RobotEra L7's 55 DOF + 400N·m knee torque + fully direct-drive dexterous hand target high-dynamic work; Booster Robotics T2's 75 DOF (one of the highest disclosed DOF values in this group) target whole-limb developer experiments; LimX Dynamics TRON's multi-form bipedal (re-configurable outfitting) target research configurability, with Luna / oli completing the interactive and general-purpose forms.

5.3. L3 Orchestration and Control Layer

LimX Dynamics' COSA is the only L3-level component among the three companies explicitly released as a product under the name "agentic operating system" (COSA 0.5, 2026-07-15); RobotEra implicitly carries orchestration via the frequency-split VLA; Booster Robotics relies on developer-built solutions (its platform positioning is precisely to leave L3 to the user).

5.4. L4 Memory and State Layer

None of the three companies has publicly disclosed a runtime memory mechanism. RobotEra's world model (learning from historical video) and LimX Dynamics' COSA world-model engine have the theoretical foundation for L4, but implementation details are missing, marked [To be filled].

5.5. L5 Evaluation and Observation Layer

Booster Robotics has the most structural advantage: competition results are a natural, comparable and reproducible evaluation. RobotEra has a single-point result — the games high-jump championship (95.641cm) — plus delivery data (400+ units); LimX Dynamics is mainly lab-level algorithm demonstrations, with the least public quantitative evaluation.

5.6. L6 Governance and Safety Layer

Public information is thin for all three companies, marked [To be filled]. The only relatively visible point is: the industrial / special-purpose scenarios LimX Dynamics focuses on imply higher safety-access requirements (its COSA architecture takes the design orientation of internalizing governance into an operating-system responsibility); Booster Robotics' education and competition scenarios carry a smaller risk surface.

5.7. Summary of the Three Companies' Six-layer Comparison

LayerRobotEraBooster RoboticsLimX Dynamics
L1 Context engineering★★ World-model priors★★ Competition-standard context★★★ System 1 + COSA
L2 Tools and execution★★★ 55 DOF high-dynamic★★★ 75 DOF open platform★★ Multi-form bipedal
L3 Orchestration and control★★ Frequency-split implicit carrying★ Left to developers★★★ COSA operating system
L4 Memory and state★★ Theoretically present★★★ Theoretically present
L5 Evaluation and observation★★ Delivery + single-point result★★★ Competition as evaluation★ Least public
L6 Governance and safety★★★ Architecture orientation visible

6. Actual Cases

6.1. RobotEra: Logistics as the First Stop and Thousand-unit-scale Delivery

RobotEra made logistics its first commercialization stop (official PMF wording): logistics scenarios with Beizi Technology, retail scenarios with Haier Smart Home, and research scenarios with MIT and ByteDance labs. As of 2025Q3 it had cumulatively delivered 400+ units with total orders exceeding RMB 500 million; overseas revenue has been over 50% of the total since early 2025; 2026Q2 launched thousand-unit-scale mass delivery of the M7. The official promotional wording claims that "9 of the world's TOP10 tech giants by market cap are its customers" (Credibility Level C, cited here with the label attached). The funding pace is dense: 2025-11, nearly RMB 1 billion Series A+ (led by Geely Capital) → 2026-03, RMB 1 billion strategic round (Samsung / CICC, valuation broke RMB 10 billion) → 2026-07-06, RMB 1 billion (led by Chengtong Fund + multiple state-owned funds), nearly RMB 2.5 billion in two months.

6.2. Booster Robotics: Competition and Developer Ecosystem

Booster Robotics' "cases" take the form of competition presence: T1 is a commonly used platform for RoboCup-style competitions at multiple universities (wording to be supplemented), and "compete as practice, let competitions drive research" (in the words of 浦银国际) summarizes its route. After completing the Series B of about RMB 1 billion in 2026-04 and entering the ranks of unicorns, it released T2 on 2026-07-10 (Thor chip / 2070 TFLOPS / up to 75 DOF), completing the jump from a competition platform to a high-compute developer platform.

6.3. LimX Dynamics: Industrial and Special-purpose Push and Pre-IPO

LimX Dynamics focuses on industrial and special-purpose scenarios: "accelerating scaled deployment in automotive manufacturing and warehouse logistics" (progress wording, no named customers, marked [To be filled]). The official site's innovation stories show scenario validation such as bipedal museum guiding and unmanned mountain photovoltaic inspection. The capital side is its most active dimension: 2026-02 Series B of USD 200 million (JD / SAIC / NIO / IDG / Lenovo Capital, etc.) → 2026-07-14 Pre-IPO of nearly USD 200 million (additional investment by IDG / Lens Technology / Huashan Capital / UAE Leishi Capital), post-money valuation of RMB 15 billion; half a year's cumulative total of USD 400 million, about 70% from overseas institutions; 2026-03 completed the joint-stock restructuring. Note: the Hurun 2026 Global Unicorn List values it at RMB 6.8 billion, standing side by side with the Pre-IPO post-money RMB 15 billion in two metrics, without choosing one.

7. Summary

7.1. Strengths and Limitations

RobotEra: the strengths are the delivery data (400+ units) and the model-iteration speed of full-stack in-house R&D; the limitation is that the second curve beyond logistics scenarios is still under validation. Booster Robotics: the strengths are the developer ecosystem and the structural advantage of competition-based evaluation, plus Series B capital in place; the limitation is being furthest from industrial scenarios, with opaque shipments and pricing. LimX Dynamics: the strengths are the methodological completeness of the three-layer architecture and the funding efficiency of USD 400 million in half a year; the limitation is the lack of named customer cases, with the deployment narrative weaker than the funding narrative.

7.2. Scope of Applicability

RobotEra suits scenario customers in logistics, retail, and those needing full-stack delivery; Booster Robotics suits universities, competition organizers, and developers who need an open embodiment for reinforcement-learning research; LimX Dynamics suits early partners who care about the layered-architecture methodology and are willing to co-define industrial / special-purpose scenario solutions.

7.3. Selection Recommendations

  1. For research and competition procurement, first choose Booster Robotics (reproducible evaluation) or LimX Dynamics TRON (multi-form configurable); RobotEra L7 suits research requiring whole-machine operational capability.
  2. Industrial customers should require all three companies to provide named customers and on-site operational data — all three have been rumored to have an IPO (Level D), and promotional wording needs to be verified separately from delivery evidence.
  3. Watch subsequent versions of COSA and ERA-42: the layered operating system and the world model are two routes that could change the industry's architectural consensus.
  4. Always label valuation information with its metric and date: LimX Dynamics' RMB 15 billion vs RMB 6.8 billion, and RobotEra's "valuation broke RMB 10 billion" are all funding-milestone metrics, not market pricing.

Information-Gap Statement

  1. Booster Robotics T1's price and shipment volume are missing, marked [To be filled]; "a commonly used platform for RoboCup-style competitions" is a wording to be supplemented.
  2. LimX Dynamics' named customer cases are missing, marked [To be filled].
  3. Conflict between two valuation metrics: LimX Dynamics' post-money RMB 15 billion (Pre-IPO round) vs RMB 6.8 billion on the Hurun list, presented in parallel.
  4. HK stocks IPO (LimX Dynamics' confidential filing, RobotEra rumored to be sprinting) are all Level D rumors, used only as status descriptions, not as fact.
  5. All three companies' L6 governance and safety public information is generally missing.
  6. RobotEra's "9 of the TOP10 tech giants by market cap are customers" is official promotional wording (Level C), already labeled when cited.

8. References

  1. RobotEra official website (products: L7 / Q5 / M7 / XHAND; WRC 2026 news) — RobotEra, 2026. https://www.robotera.com/
  2. RobotEra L7 product page (full spec table) — RobotEra official site, 2025. https://www.robotera.com/robot/l7.html
  3. RobotEra L7 entry (delivery / funding / games results) — Baidu Baike, 2026. https://baike.baidu.com/item/%E6%98%9F%E5%8A%A8L7/66239282
  4. Performance rivaling Tesla Optimus: RobotEra releases RobotEra L7 — Sohu Tech, 2025-07. https://it.sohu.com/a/916431213_362042
  5. LimX Dynamics official website (TRON / Luna / oli / COSA / VGM / DreamActor) — LimX Dynamics, 2026. https://www.limxdynamics.com
  6. LimX Dynamics humanoid brain system COSA 0.5 update — LimX Dynamics official site news, 2026-07-15. https://www.limxdynamics.com/zh/news/BK000065
  7. LimX Dynamics completes Pre-IPO round — LimX Dynamics official site news, 2026-07-14. https://www.limxdynamics.com/zh/news/BK000063
  8. Professors from 985 universities collectively found embodied-intelligence startups (RobotEra / Booster Robotics / LimX Dynamics / Noematrix, etc.) — JQman, 2026. https://www.jqman.com/ai/521998.html
  9. Embodied intelligence enters a "big year of listings" in 2026 (a roundup of the full-machine camp and the full-stack camp) — Sohu Tech, 2026. https://it.sohu.com/a/997432229_116132
  10. [IPO Express] LimX Dynamics and RobotEra rumored to be sprinting for the HK stocks — Mitrade 财华社, 2026-08-20. https://www.mitrade.com/zh/insights/news/live-news/article-8-2015872-20260820
  11. AI robotics weekly (Booster T2 release / LimX Pre-IPO / RobotEra RMB 1 billion) — NetEase News, 2026-07. https://c.m.163.com/news/a/L1VPHI9505118I96.html
  12. RobotEra L7 launch — Huanqiu, 2025-07. https://look.huanqiu.com/article/4Nb9s9pnmvh