Manifold AI
具身 AI 世界模型——战略溢价值得给,但高溢价入场还不到时候
观察——具身 AI 的战略期权价值真实存在,但公开证据仍落后于独角兽叙事
封面要素
公司概况
Manifold AI(流形空间)是一家北京具身 AI 初创公司,2025-05-22 由前 SenseTime 高管吴伟博士与清华大学 FIB Lab 相关团队共同创立。公司把自己定位为物理 AI 的世界模型基础设施提供商:WorldScape 承担实时空间世界模型,WorldScape Policy 把该栈延伸到动作和控制,WorldArena 则提供基准层,强化公司对功能型具身智能的定义。公开证据显示,公司在技术信号和融资上推进异常迅速——到 2026 年 6 月,六轮 Pre-A 累计融资接近 RMB1B,并在 WorldScore、WorldArena、RoboTwin 等基准中反复宣称排名第一;但运营指标和客户耐久性披露仍有限。
- 成立时间
- 2025-05-22
- 创始人
- Dr. Wu Wei, Tsinghua FIB Lab-linked founding team
- 创立地点
- Beijing, China
- 总部
- Beijing, China
- 产品
- WorldScape 世界模型、WorldScape Policy 世界动作层,以及配套机器人 / 具身 AI 基础设施,覆盖基准工具、数据采集闭环,并为物流、3C 制造、汽车制造和相邻物理 AI 场景提供部署支持。
- 客户
- 企业买方集中在电商物流、3C 制造、汽车制造,以及寻求具身 AI 控制、评估和部署能力的机器人 OEM / 集成商渠道。
- 商业模式
- 可能由软件许可、集成 / 部署服务、合作伙伴解决方案,以及潜在的基准 / 评估变现构成;但公开资料没有披露收入或价目表。
- 阶段
- Pre-A / Early Commercialization
- 融资情况
- 公司在大约一年内完成六轮融资;到 2026 年 6 月,Pre-A 累计融资接近 RMB 1 billion,把公司推入独角兽层级,但收入、现金跑道和优先股堆叠细节仍未披露。
执行摘要
主要优势
- 在具身 AI 世界模型层定位稀缺,WorldScore、WorldArena、RoboTwin 等方向都有公开领先主张。
- 融资速度快,Pre-A 累计近 RMB1B 资本提供可信资金底座,支撑继续研发和部署迭代。
- 产品叙事覆盖模型、动作、数据和基准层,而不是单点方案;如果商业化跟上,可能形成基础设施式权力。
- 物流和 3C 制造部署主张反复出现,再加上 UBTECH 这一具名商业化渠道,早期工业相关性已经可见。
- 创始人-市场匹配度强:Wu Wei 和 Tsinghua 系团队同时具备基准可信度和世界模型部署经验。
主要风险
- 收入、ARR、毛利率、客户数和现金跑道均未披露,仅靠公开证据无法按后期公司口径承保估值。
- 客户证明更多来自渠道和场景,而不是具体账户;公开记录中可点名的终端客户生产部署仍然稀少。
- 相比产品会影响物理世界这一事实,信任、安全、隐私和治理材料偏薄。
- 商业化似乎依赖少数渠道和战略生态,尤其是 UBTECH 以及有产业连接的支持方。
- 公司可能已按品类领导者定价,但公开证据尚未证明可复制、接近软件的部署经济性。
未决问题
- 需要具名生产客户清单,列明部署阶段、站点数量和可引用 KPI 结果。
- 需要收入、订单、毛利、烧钱速度和现金跑道数据,以区分软件经济性与服务密集型试点。
- 需要快速六轮融资之后的完整股权结构表、清算优先权和 side-letter 经济条款。
- 需要 UBTECH 及其他 OEM / 集成商渠道的合作经济性和商业所有权细节。
- 需要安全、隐私、事件响应和治理材料,验证工业部署准备度。
目录
01公司概况
1.1 身份、使命与产品范围
Manifold AI,也以中文品牌「流形空间」出现,公开起点可追溯到 2025-05-22:公司网站上线,公开资料里的法律主体时间线也从这里开始。所审阅来源一致把公司放在北京,并描述为较早用自研世界模型作为具身智能底层的玩家。官方定位很宽,不只是狭义的机器人软件:Manifold 称自己在为机器人和 XR 设备等 AI 硬件应用打造下一代世界模型。2026 年 6 月的独立报道进一步收窄并强化了叙事,称公司是中国首家把自有世界模型作为具身 AI 基础模型的初创公司。 核心产品叙事围绕 WorldScape 和 WorldScape Policy 展开。WorldScape 反复被描述为实时世界模型,可在同一交互闭环中支持移动和操作;这点重要,因为公司不再停留在更好看的视频生成,而是走向机器人可用的状态预测。公开架构描述强调混合专家设计和强几何基础。WorldScape Policy 把该栈延伸到动作执行,公司称其闭环表现强于现有 VLA 基线。合在一起,产品主张不只是“更好的仿真”,而是一套预训练和控制栈,目标是帮助机器人在混乱物理环境中感知、推理并行动。[CO001, CO002, CO003, CO004, CO005, CO006]
| 指标 | 数值 / 状态 | 日期 | 置信度 | 缺口 / 注意事项 |
|---|---|---|---|---|
| 成立时间 | 2025-05-22 | 2025-05-22 | 高 | 由官网里程碑和公开公司资料支撑 |
| 总部 | 中国北京 | 2026-06 | 中 | 公开资料多次重复;本次会话未能从注册信息独立确认 |
| 创始人 / CEO | 吴维博士 | 2026-06 | 高 | 多个媒体资料中的职务一致 |
| 当前阶段 | Pre-A 后期 / 独角兽阶段私营创业公司 | 2026-06 | 中 | 估值层级比经审计运营规模更清楚 |
| 公开证据最充分的融资额 | 累计 Pre-A 近 RMB 1 billion | 2026-06 | 高 | 多家 2026 年 6 月来源均指向“近10亿元” |
| 估值状态 | 十亿美元级独角兽 | 2026-06 | 高 | 叙事确认较强,具体估值方法未公开 |
| 核心产品栈 | WorldScape + WorldScape Policy | 2026-06 | 高 | 产品主张公开;商业化深度仍在浮现 |
| 公开部署证据 | 电商物流和 3C 制造 | 2026-07 | 中 | 已公开具名垂直场景,但客户名称和规模未公开 |
| 收入 / ARR | null | 2026-08-03 | 高 | 已审阅公开来源没有披露收入、ARR 或运行率 |
| 员工数 | null | 2026-08-03 | 高 | 已审阅来源没有披露精确员工数 |
| 客户数 | null | 2026-08-03 | 高 | 公开材料描述场景,未给出可辩护的客户总数 |
null 表示截至运行日,已审阅材料无法公开支撑该指标,而不是底层业务指标等于零。
[CO001, CO014, CO021, CO026, CO027, CO028]Manifold AI 如何串起研究血统、世界模型产品、基准测试、资本和工业部署路径。
[CO003, CO005, CO009, CO011, CO021, CO031]截至运行日期,Manifold AI 成熟度指标一览。
融资和估值做了取整,因为公开报道使用的是“近10亿元”“10亿美元级”等叙事标签,而非审计数字。
[CO021, CO026, CO027, CO028, CO031, CO032]1.2 创始人、研究谱系与治理可见度
对一家中国民营 AI 初创公司而言,公开创始人归属异常一致:多方来源都把吴伟博士列为创始人兼 CEO,并称其曾是 SenseTime 高管,拥有直接的世界模型经验。几篇文章还重复了更强的可信度信号——吴伟曾带队连续两次在 Waymo SimAgents Challenge 获得第一——这说明创立故事扎根在仿真和智能体行为建模,而不是临时转向“物理 AI”。 更大的团队故事试图把学术背书和落地可信度捏在一起。QQ、盖世汽车和百度百科的报道称,公司集合了清华 FIB Lab 研究人员、清华教授 / 长江学者,以及前大厂世界模型或自动驾驶负责人。公开材料还称团队累计发表 200 多篇顶会论文、获得超过 100,000 次引用。对一家重评估、基准驱动的公司来说,这是很强的创始人市场匹配信号。弱点在治理透明度:所审阅材料没有披露董事会、独立监督或更完整的高管梯队。按尽调口径,Manifold 目前仍像一家由创始人和研究驱动的公司,关键人集中在吴伟身上,公开证据也有限,无法判断资本如何转化为正式治理控制。[CO014, CO015, CO016, CO017, CO018, CO019]
| 人物 / 节点 | 角色 | 背景 / 证明点 | 创始人与市场匹配 / 覆盖 | 关键人物依赖 |
|---|---|---|---|---|
| 吴维 | 创始人兼 CEO | 前 SenseTime 高管;Waymo SimAgents 冠军说法被多次提及 | 世界模型研发和仿真到创业组建之间有直接连续性 | 极高 |
| 清华 FIB Lab 关联联合创始人 | 学术 / 研究领导 | FIB Lab 被描述为中国最早的世界模型实验室,并创建 WorldArena | 增强基准可信度和研究招聘能力 | 高 |
| 前自动驾驶与大科技公司世界模型负责人 | 应用 AI 与基础设施覆盖 | 公开文章提到来自 Momenta、XPeng、Microsoft 等大公司的校友 | 提升从研究栈转向物理部署的能力 | 中 |
| 董事会 / 更广泛高管团队 | 未公开披露 | 已审阅来源未发布独立董事、董事会席位、CFO 或 COO 细节 | 治理覆盖仍是未解决尽调缺口 | 高 |
该表有意保持不完整,因为公开材料对创始人叙事很丰富,但对正式治理披露很稀疏。
[CO014, CO015, CO016, CO017, CO018, CO019]1.3 融资历程、投资方基础与独角兽状态
最清晰的资本事实出现在 2026 年 6 月。AIbase、盖世汽车、腾讯新闻和新浪都称,Manifold AI 在约一年内完成六轮融资,Pre-A 累计融资接近 RMB 1 billion。该轮新进投资方包括国新基金、Temasek 旗下怡峰资本、北汽产业投资和新能创投,组合里既有国资信用、主权相关资本,也有产业渠道价值。同一批来源还回溯到君联资本和 Huawei Hubble 等早期投资方,说明在 2026 年 6 月融资前,股权结构表已经汇集了一线 VC 和战略投资方。 6 月之前的时间线足以说明动能,尽管并不完全干净。2026 年春季报道包括一轮近 RMB-200 million 的 Pre-A,以及随后数亿元人民币的 Pre-A+。公开记录尚不能支持精确完整的股权结构表、任何债务工具或任何老股交易披露。不过,在当前披露水平下,估值信号已经异常强:多个 2026 年 6 月来源明确称 Manifold 已进入十亿美元“独角兽”行列。估值方向因此比收入质量清楚得多。所审阅公开来源没有披露 ARR、收入运行率或收入,资本市场热情目前跑在运营披露前面。[CO020, CO021, CO022, CO023, CO024, CO025]
| 利益相关方 / 投资人 | 角色 | 控制权或经济重要性 | 当前证据状态 | 尽调要求 |
|---|---|---|---|---|
| 吴维 / 创始团队 | 运营控制中心 | 创始叙事和技术方向似乎集中在创始人领导的团队周围 | 控制权未公开披露 | 索取章程、投票权和当前董事会构成 |
| Guoxin Fund | 2026 年 6 月新投资人 | 增加国资信誉和战略信号 | 多篇 2026 年 6 月报道具名 | 确认投资金额及任何治理权利 |
| Yifeng Capital(Temasek 关联) | 2026 年 6 月新投资人 | 增加主权资本关联和潜在区域网络价值 | 多篇 2026 年 6 月报道具名 | 确认实体名称、金额和董事会观察员身份 |
| BAIC 产业投资 / BAIC Capital | 战略产业投资人 | 潜在汽车与制造渠道伙伴 | 2026 年 6 月报道具名 | 厘清投资是否也包含商业合作条款 |
| Xinneng Venture Capital | 2026 年 6 月新投资人 | 最新超额认购融资的一部分 | 2026 年 6 月报道具名 | 确认规模及任何财团角色 |
| Legend Capital / Huawei Hubble / 既有股东 | 早期财务与战略支持方 | 在独角兽跃升轮之前传递持续支持信号 | 公开具名,但当前确切持股未披露 | 区分历史轮次参与和当前股权结构所有权 |
该表区分已识别投资人名称与未经验证的持股比例或权利,后者公开不可得。
[CO020, CO021, CO022, CO023, CO024, CO025]1.4 里程碑、基准证明与早期商业化信号
Manifold 的公开里程碑故事压缩且连贯。官网记录了公司在 2025 年 5 月启动、2025 年 7 月推出 RoboScape、2025 年 11 月推出 AirScape。到 2026 年初,叙事从领域模型转向通用化栈:WorldScape 与 WorldScape Policy、WorldScore 公开第一名主张,以及 WorldArena 的发布。WorldArena 作为基准,试图衡量模型是否真正能支撑具身任务,而不只是生成逼真的视频。WorldArena 论文和网站用 16 项指标、六个维度的基准给叙事补上实质,也提出更广泛的观点:视觉质量和具身可用性常常并不一致。 商业证明仍比基准证明更早期。所审阅来源称,Manifold 已经在电商物流和 3C 制造中落地用例,汽车生产和混合现实娱乐更多被放在下一阶段。因此,2026 年 7 月与 UBTECH 的合作很重要,因为它把基准领先转译为拥有制造触达能力的产业化伙伴。同时,周边行业仍有风险:MERICS 认为,中国具身 AI 栈仍依赖 Nvidia 工具,距离充分自主、规模化部署还很远。再叠加未披露的客户数、员工数和治理细节,当前里程碑故事很有力量,但还不能等同于商业化风险已经充分出清。[CO007, CO011, CO012, CO013, CO031, CO032]
| 日期 | 事件 | 类型 | 金额 / 估值 / 状态 | 参与方 | 含义 |
|---|---|---|---|---|---|
| 2025-05-22 | 公司成立 / 官网上线 | 创立 | 公开露出开始 | Manifold AI | 官方公开时间线起点 |
| 2025-07-03 | RoboScape 发布 | 产品 | 面向机器人的世界模型 | Manifold AI / 清华关联作者 | 具身世界模型研究的早期产品化 |
| 2025-11-06 | AirScape 发布 | 产品 | 面向无人机的世界模型 | Manifold AI | 显示出超越单一机器人形态的多域野心 |
| 2026-02 | WorldArena 论文发布 | 产品 | 引入 14 模型基准 | Tsinghua FIB Lab、Manifold AI、合作者 | 基准领先成为公司护城河的一部分 |
| 2026-03 to 2026-04 | 近 RMB200M Pre-A 披露 | 融资 | Pre-A 轮 | Huakong Fund、Xichuangtou、Datai、既有投资人 | 资本支持持续模型与基础设施建设 |
| 2026-04-03 | 公开资料称 WorldScape 登顶 WorldScore | 产品 | #1 排名主张公开 | Manifold AI、WorldScore 生态 | 提升相对全球同行的技术可见度 |
| 2026-04-12 | CVPR 2026 WorldArena Challenge 公布 | 合作 | 挑战赛启动 | AMap CV Lab、Manifold AI、Tsinghua 和合作机构 | 推动基准从论文走向社区竞赛 |
| 2026-06-18 | 6 月融资使 Pre-A 累计接近 RMB1B | 融资 | 独角兽级私营估值叙事 | Guoxin、Yifeng、BAIC、Xinneng、既有股东 | 资本和叙事在一年内跃升 |
| 2026-06-18 | WorldScape / Policy 被公开称为 WorldScore、WorldArena、RoboTwin #1 | 规模 | 重复基准领先主张 | Manifold AI | 技术声誉成为主要公开牵引信号 |
| 2026-07-13 | UBTECH 战略合作公布 | 合作 | 商业化合作 | Manifold AI 与 UBTECH | 从基准胜出通往工业部署的潜在桥梁 |
日期反映公开披露时间。部分内部产品或融资交割日期可能早于这里使用的发布日期。
[CO020, CO021, CO024, CO025, CO033, CO034]从 2025 年 5 月创立到 2026 年 7 月与 UBTECH 建立商业化伙伴关系的压缩时间线。
[CO020, CO021, CO033, CO034, CO035, CO041]1.5 证据摘录
02市场分析
2.1 市场边界:具身 AI 赋能层,而不是整个机器人经济
最重要的框定选择,是不要把“具身 AI”、“人形机器人”、“自主移动”和“世界模型”混为一谈。对 Manifold AI 来说,真正相关的市场是赋能层:世界模型、动作模型、数据管线和部署软件,帮助机器在真实物理工作流中感知并行动。这个层面触达更广的具身智能经济,但并不等同于完整机器人硬件市场,也不是通用视频生成 AI。BCG 的能力框架在这里有用,因为它把物理 AI 视为一组能力栈——感知、操作、规划和推理——而不是单一机器人形态。 这个区分直接影响估值和进入市场路径。面向中国具身智能经济的万亿元市场叙事,方向上可能正确,但 Manifold 只能变现其中一部分支出:投向模型基础设施、数据、控制软件,以及部署到物理推理能带来可测价值的场景中的那一部分。相邻玩家有助于厘清边界。Physical Intelligence 讲的是“适用于任何机器人”的模型故事,Wayve 把具身 AI 用于自动驾驶,AGIBOT、Unitree、X Square、ROBOTERA、Fourier 和 UBTECH 等公司则以不同比例混合硬件、软件和场景所有权。Manifold 在这个更大领域里最接近软件与模型层。[CM001, CM002, CM006, CM016, CM017, CM018]
| 细分 / 品类 | 计入支出 | 不计入支出 | 买方 / 付款方 | 与 Manifold 的相关性 |
|---|---|---|---|---|
| 具身 AI 模型赋能 | 世界模型、动作模型、仿真、数据管线、部署软件 | 纯硬件 BOM、与物理任务无关的通用云 AI | 机器人 OEM、集成商、制造商、物流运营商 | 核心目标市场 |
| 人形 / 机器人硬件 | 机器人本体、执行器、传感器、末端执行器、制造产能 | 未捆绑时的纯模型软件层 | 机器人制造商、工业买家 | 相邻渠道和伙伴层 |
| 自动驾驶出行具身 AI | 驾驶模型、自动驾驶栈、网约车或 OEM 部署软件 | 除非共享模型 IP 适用,否则不含仓储 / 工厂机器人 | 汽车 OEM、AV 运营商、出行平台 | 世界模型可在另一垂直领域变现的相邻证明 |
| EAI 数据基础设施 | 远程操作、合成数据、数据存储、云数据商城、评估系统 | 不聚焦机器人任务的通用企业数据工具 | 模型团队、机器人实验室、集成商 | Manifold 世界模型训练与迭代的关键子板块 |
| 混合现实 / 娱乐物理 AI | 与 XR 相关的物理仿真和具身交互 | 与物理 AI 无关的纯消费者游戏 | 工作室、场馆运营商、设备制造商 | 相邻领域,不是当前核心 |
| 传统固定自动化 | 稳定环境中的可编程工业机器人 | 基于学习的感知和推理层 | 制造业资本开支负责人 | 现状替代方案,不是直接核心市场 |
该表把完整具身智能经济,与更贴合 Manifold AI 公开定位的较窄软件和数据楔子分开。
[CM001, CM002, CM017, CM019, CM029, CM038]从宽口径具身智能经济,到与 Manifold AI 更相关、更窄的软件和数据切口的分层视图。
底层刻意保持定性,因为公开来源没有披露足够的定价、部署或客户转化数据,无法支撑 Manifold SOM 的数字口径。
[CM003, CM004, CM005, CM029, CM035, CM037]2.2 规模测算视角:TAM 很大,眼前的软件机会小得多
公开市场估算最好被当作多组镜头,而不是一个可完全调和的答案。在最宽口径下,36Kr Research Institute 估计中国具身智能市场 2025 年达到 RMB915 billion,2026 年可能超过 RMB1 trillion。更窄口径下,ResearchInChina 估计中国 EAI 数据市场 2025 年为 RMB500 million,同比增长 203%,约占全球市场的 40%。第三个镜头来自融资流而不是终端市场支出:Embodied Global 的 2026 年上半年统计显示,322 笔交易合计 RMB93.5 billion,且呈现强杠铃结构,大部分资本流向少数旗舰公司。 这些估算并不彼此矛盾,只是在衡量不同对象。宽口径数字统计的是一个包含硬件、集成、服务和下游部署的经济体。窄口径数字聚焦数据基础设施和模型开发输入。融资流本身不是收入,但能显示投资人在成熟商业指标出现之前多么激进地承保未来类别领导者。对 Manifold 来说,正确解读应更保守:头部 TAM 证明方向和政策优先级,但眼前可服务市场更接近较小、软件和数据权重更高的楔子,而不是完整的万亿元工业总量。[CM003, CM004, CM005, CM014, CM015, CM029]
| 发布方 / 视角 | 年份 | 地理 / 范围 | 数值 | 方法 / 计入口径 | 置信度 / 限制 |
|---|---|---|---|---|---|
| 36Kr Research Institute | 2025 | 中国具身智能经济 | RMB915B | 覆盖上游至下游商业化的宽口径产业经济估算 | 中;对 Manifold 特定 SAM 来说过宽 |
| 36Kr Research Institute | 2026E | 中国具身智能经济 | >RMB1T | 延续同一宽口径市场定义的前瞻估算 | 中;标题口径 TAM,不是可服务市场 |
| ResearchInChina | 2025 | 中国 EAI 数据市场 | RMB500M | 更窄的具身 AI 数据基础设施市场 | 中;更接近 Manifold 软件楔子 |
| ResearchInChina | 2030E | 全球 EAI 数据市场 | US$5.25B | 基于 2025 年基准、由 CAGR 推动的全球数据市场预测 | 中;基础设施切片,不是完整机器人市场 |
| Embodied Global / Ebrun 统计 | H1 2026 | 中国具身 AI 融资 | 322 笔交易合计 RMB93.5B | 融资流视角,而不是终端市场支出 | 中;不是收入,可能夸大已兑现需求 |
| China Economic Net / IDC 引用 | 2025 | 中国用户在具身智能机器人上的支出 | >US$1.4B | 聚焦机器人的当前支出视角,不只是软件 | 中;相邻但可用作当前需求锚 |
| China Economic Net / IDC 引用 | 2030E | 中国用户在具身智能机器人上的支出 | US$77B | 引用文章中 94% CAGR 的前瞻支出预测 | 中低;预测敏感度高 |
这些估算有意保留范围不匹配。宽口径经济数字、较窄数据市场数字和融资流视角,不应直接相加或互换。
[CM003, CM004, CM005, CM035, CM037]相互矛盾的市场口径保留为低 / 基准 / 高区间,而不是强行压成单点估计。
部分条目是固定来源点,而非真正区间。只有引用预测明确呈方向性时,才使用可变上下界。各行单位不同,并已在各标签中标明。
[CM003, CM004, CM005, CM035]2.3 买方、用户、付款方与采用路径
近期具身 AI 需求更多来自企业工作流痛点,而不是消费者机器人幻想。在制造和物流中,用户通常是一线工人、仓库运营人员或机器人工程师。买方在运营、自动化、制造或平台工程部门。付款方是资本开支、自动化或运营预算负责人,他们关心吞吐、换线灵活性、用工短缺和正常运行时间,而不是新奇感。因此,这个市场更像 B2B 基础设施,而不是消费硬件。 全行业的采用路径也越来越清晰。基准或实验室证明仍然重要——这也是 WorldArena 和类似测试床受到关注的原因——但买方不会只凭基准分数就扩大部署。路径通常从技术证明走向受控试点,再到线体单元或站点部署,只有在运营稳定后才进入更广泛铺开。公开案例支持这一判断:UBTECH 2025 年报显示工业人形机器人收入开始形成有意义规模;AGIBOT 把 Longcheer 消费电子部署包装为生产突破;Manifold 与 UBTECH 也明确把物流定位为商业化楔子。对 Manifold 来说,这意味着在任何大众化软件分发模式具备可信度之前,销售很可能先由少数设计伙伴关系驱动。[CM020, CM021, CM022, CM023, CM024, CM025]
| 细分 | 买方 | 用户 | 付款方 / 预算负责人 | 工作流 / 采用触发因素 | 对 Manifold 的意义 |
|---|---|---|---|---|---|
| 3C 制造 | 工厂自动化负责人 / COO | 产线操作员、QA、机器人工程师 | 制造业资本开支或自动化预算 | 柔性装配、换线、劳动力短缺 | 已在公开部署证据中具名 |
| 电商物流 / 仓储 | 运营 VP / 履约负责人 | 拣选员、分拣员、机器人操作员 | 运营预算或租赁自动化支出 | 非规则 SKU 处理、7×24 吞吐压力 | 被具名为早期商业化楔子 |
| 汽车生产 | 智能工厂负责人 / 工业工程 | 装配与物料搬运团队 | 工厂资本开支 | 复杂多步骤搬运与检测 | 可借 BAIC 等战略投资人切入的大型邻近切口 |
| 机器人 OEM / 平台伙伴 | CTO / 产品负责人 | 模型、控制与数据团队 | R&D 平台预算 | 需要世界模型预训练和评测基础设施 | 直接的软件客户原型 |
| 研究实验室 / 基准参与者 | PI / 实验室负责人 / 首席科学家 | 研究人员和开发者 | 资助金或 R&D 预算 | 需要评测、合成数据和实验工具 | 推动基准采用和生态心智 |
| MR / 场馆运营方 | 创新负责人 / 制片人 | 体验设计师与运营人员 | 项目预算 | 混合现实体验中的具身交互 | 邻近方向,但当前证据弱于工业场景 |
买方与付款方往往不是同一人。多数近期行业里,使用者是操作员或工程师,但采购决定由企业运营或自动化负责人拍板。
[CM022, CM024, CM025, CM027, CM038]将六个目标细分市场映射到用户、付款方、预算权和采用触发因素。
[CM016, CM024, CM025, CM027, CM038]从技术承诺到工业具身 AI 大规模铺开的示意性采用路径。
该漏斗是方向性示意,并非来源自带图。它展示了头部 TAM 经过工作流匹配、ROI、部署就绪度和披露质量等连续关口后,能留下多少。
[CM026, CM031, CM036, CM037]2.4 增长动力与约束
最强需求驱动都很务实。36Kr 指向老龄化、用工短缺,以及僵硬的传统自动化在柔性制造中的边界。ResearchInChina 补充说,硬件一旦越过最早原型阶段,真正瓶颈会转向可扩展、物理上逼真的多模态数据,以及把这些数据变成持续改进模型的系统。BCG 从经济性角度强化了同一点:机器人 TCO 主要被设置和再工程吃掉,所以软件定义的灵活性即使在通用自主尚未出现前,也能释放不成比例的价值。 约束同样重要。MERICS 认为,中国具身 AI 推进仍高度依赖 Nvidia 生态,而精度、灵巧度和成本在许多部署中仍未解决。BCG 在概念上更直接:今天的市场可以围绕感知和有边界的操作创造价值,但不确定性下的真正推理仍是前沿。监管是另一个慢变量。具身 AI 一旦扩展到移动、公共服务和劳动力管理,一些系统会碰到关键基础设施或就业决策等高风险类别。结果是一个动能极强、但 ROI 边界仍不确定的市场——正是在这种环境里,Manifold 这样的基准领先者可以早早吸引资本,但仍需客户证明,实际 SOM 才会清楚。[CM007, CM008, CM009, CM012, CM013, CM031]
| 驱动 / 约束 | 方向 | 时间 | 影响 | 尽调问题 |
|---|---|---|---|---|
| 劳动力短缺与柔性制造需求 | 正向 | 当前 | 固定自动化失灵时,企业更愿意测试具身 AI | 向潜在客户确认替代人工与提升吞吐量的决策规则 |
| 瓶颈从硬件转向数据 | 利好模型供应商 | 当前至 2028 年 | 抬高世界模型、合成数据和评测系统的战略价值 | 量化 Manifold 自有数据资产优势 |
| 资本向头部平台集中 | 利好领先者,利空跟随者 | 当前 | 可加速基准战与招聘上的赢家拿大头态势 | 判断 Manifold 是否进入资本集中圈层 |
| 依赖 Nvidia 生态 | 负向 / 约束 | 当前 | 带来供应链与地缘政治平台风险 | 测试能否迁移到国产或替代算力栈 |
| 推理仍是前沿能力 | 负向 / 约束 | 中期 | 限制近期通用自主能力主张 | 把可落地的 Level-2/3 任务与愿景式 Level-5 叙事拆开 |
| 监管扩展到高风险领域 | 负向 / 选择性 | 中期 | 可能拖慢出行、公共服务或劳动力相关部署 | 按 AI Act / 本地安全规则分类拟定用例 |
| ROI 与定价数据不透明 | 负向 / 卡口 | 当前 | 公开证据不足,SAM 与客户转化建模困难 | 在 NDA 下索取部署经济性、定价模型和转化漏斗 |
同一因素是驱动还是约束,取决于 Manifold 能否把技术深度转成可衡量的部署 ROI。
[CM009, CM012, CM013, CM031, CM032, CM034]2.5 证据摘录
03竞争格局
3.1 竞争格局概览
Manifold AI 面对的不是一组干净同行。格局至少分为四层。第一层是软件优先或模型优先的具身 AI 公司,例如 Physical Intelligence,以及中国语境下的 X Square,它们强调通用学习系统、数据闭环和基础模型宽度。第二层是全栈机器人平台厂商,例如 AGIBOT、UBTECH、Unitree 和 ROBOTERA,它们拥有或紧密控制硬件、软件和部署。第三层是自动驾驶中的相邻具身 AI 公司,主要是 Wayve 和 Waymo;它们证明世界模型和端到端学习思想可以变成商业系统,即便终端市场不同于 Manifold。第四层是非初创替代品:固定工业自动化,以及机器人 OEM 或大型工业集团内部自建项目。 这种分层很重要,因为客户购买 Manifold 时,不只在比较研究排名。他们还会比较供应商是否控制硬件、能否交付工业正常运行时间、是否具备监管可信度、能否生成自有数据,以及能否借伙伴扩张。因此,Manifold 在模型层的直接竞争很窄,但在预算层的竞争要宽得多。世界模型初创公司不只会输给另一家世界模型初创公司,也可能输给垂直整合的机器人厂商,或输给决定围绕更窄内部栈自建的客户。[CP001, CP002, CP003, CP004, CP017, CP021]
| 竞争对手 | 类别 | 规模 / 融资信号 | 目标客群 | 差异化 | 局限 |
|---|---|---|---|---|---|
| Physical Intelligence | 软件优先的机器人基础模型 | 历史融资 $1B+;openpi 分发 | 机器人 OEM 与开发者 | 通用任意机器人模型叙事;开源 VLA 栈 | 开源和融资可见度高,商业化深度公开较少 |
| Wayve | 面向自动驾驶出行的具身 AI | 官网 About 页称总融资 $2.8B | 汽车 OEM、出行平台 | 面向驾驶的世界模型;资本与伙伴基础庞大 | 终端市场不同于工厂 / 物流机器人 |
| AGIBOT | 全栈人形机器人与数据集平台 | 制造业部署证据;产品节奏快 | 制造、商业、研究 | 拥有硬件,加上数据集生态和部署证据 | 仍偏中国市场;除有限审计指标外,公关色彩重 |
| UBTECH | 规模化工业人形机器人平台 | HKEX 申报:2025 年收入 RMB2.0B | 工业制造与物流 | 公开财务、参与标准制定、规模可信度 | 纯软件灵活性弱于模型原生供应商叙事 |
| X Square Robot | 全栈具身 AI 基础模型平台 | 据 2026 年 7 月新闻稿,估值 >US$2.8B | 家庭、工业、物流 | 场景广,叠加基础模型和数据管线 | 非公开运营指标仍薄 |
| ROBOTERA | 全栈具身 AI 物流与工业玩家 | 2026 年 7 月融资 >$200M;声称交付千台级 | 物流、汽车、电子 | 快速物流 PMF 叙事和强工业投资人阵容 | 公开证据集中在公司关联披露 |
| Unitree | 低成本足式与人形硬件供应商 | 声称在四足机器人全球销量领先 | 巡检、研究、商业娱乐 | 性价比与商业化速度 | 模型层差异化弱于世界模型原生同业 |
| Fourier | 医疗 / 康复加人形机器人邻近方向 | 具身 AI 与康复在机构中有布局 | 医疗、康复、人形机器人 | 差异化领域和护理用例 | 与 Manifold 的工业世界模型叙事可比性较弱 |
| Waymo | 运营式自动驾驶服务标杆 | 商业乘车服务已上线 | 自动驾驶出行服务 | 信任与公共部署可信度 | 并不直接向工厂授权或销售模型 |
本表混合直接同业与战略邻近公司,因为买方和投资人比较具身 AI 叙事时,常跨邻近类别,不只看完全相同的产品。
[CP005, CP007, CP009, CP010, CP013, CP016]按技术世界模型深度和当前商业化证据比较相对位置。
坐标轴是基于证据的顺序评分,由公开披露综合得出,并非经审计的量化排名。
[CP005, CP007, CP011, CP013, CP016, CP020]3.2 直接和相邻技术同行
Physical Intelligence 是最清晰的软件优先标杆。其官网营销一套可控制任何机器人完成任何任务的模型,openpi 仓库则把这一主张变成分发引擎,让开发者接触到多个 VLA 变体和微调路径,这些模型基于 10k-plus 小时机器人数据训练。Wayve 在商业上没那么直接,但技术上相邻:它明确把具身 AI 用于自主移动,并用 GAIA-1 展示生成式世界模型如何加速自动驾驶训练和评估。两家公司都重要,因为它们验证了一个观点:战略控制点可以在模型层,而不只在机器人本体。 在中国同行中,X Square 和 ROBOTERA 尤其相关,因为它们把全栈叙事与大量资本和场景宽度结合起来。X Square 称自己结合了基础模型、机器人硬件、自有数据管线,以及在家庭、工业和物流场景中的真实部署;公司还称估值超过 US$2.8 billion。ROBOTERA 把自己营销为全栈自研具身智能公司,并称已经在十多个物流中心部署,Q2 2026 开始千台级交付。这些同行压缩了 Manifold 的差异化空间:在这个类别里,“全栈”、“数据驱动”或“世界模型原生”已经不再是稀缺语言。[CP005, CP006, CP007, CP008, CP016, CP017]
| 购买标准 | Manifold AI | Physical Intelligence | Wayve | AGIBOT | UBTECH | X Square / ROBOTERA |
|---|---|---|---|---|---|---|
| 世界模型品牌与基准可见度 | 很强 | 强 | 自动驾驶领域强 | 中等至强 | 中等 | 强 |
| 自有机器人硬件平台 | 否 | 否 | 否 | 是 | 是 | 是 |
| 开源分发 | 公开证据有限 | 是(openpi) | 有限 | 选择性生态 | 有限 | 是 / 选择性 |
| 公开工业部署证据 | 初现 | 公开证据有限 | 不同细分市场 | 强 | 强 | 增长中 |
| 经审计或申报支撑的财务披露 | 否 | 否 | 否 | 否 | 是 | 否 |
| 物流 / 制造 / 家庭场景广度 | 叙事宽 | 通用机器人野心 | 以出行为中心 | 广 | 偏工业 | 广 |
| 战略投资人 / 伙伴渠道强度 | 增长中 | 强 | 很强 | 强 | 很强 | 强 |
单元格为定性判断,因为公开来源没有披露同业之间可直接比较的合同、定价或单位经济数据。
[CP006, CP008, CP010, CP013, CP016, CP018]概览:不同同行群体如何强调模型、硬件、部署和开放生态策略。
[CP006, CP010, CP013, CP016, CP019, CP021]3.3 商业化、分发与信任
Manifold 与最成熟竞争者之间最大的差距,是商业化可见度。UBTECH 是最强公开标杆,因为它有监管文件支持的收入、明确的工业应用重点、超过 6,000 台人形机器人的年化产能,以及清楚的标准制定角色。基于 Longcheer 制造公告和“Deployment Year One”的叙事,AGIBOT 在部署曲线上也比 Manifold 走得更远。Waymo One 证明的是另一点:运营型自动驾驶服务模式可以达到公开商业规模,即便它不是直接许可同行。 分发能力越来越来自生态,而不只是算法。战略投资方和产业伙伴重要,是因为它们提供客户入口、供应链、集成能力和可信度。X Square、ROBOTERA、UBTECH 和 Wayve 都在技术深度之外强调投资方和伙伴网络。NVIDIA 的 Jetson Thor 生态也提醒市场,基础算力正被全行业共享,这会削弱低层基础设施差异化,把竞争推向数据、部署和信任。对 Manifold 来说,真正的竞赛是能否在更大、伙伴更多的玩家锁定最佳场景前,把基准关注转化为耐久渠道。[CP011, CP012, CP013, CP014, CP015, CP022]
| 公司 | 公开价格 / 合同模式 | 包含能力 | 折扣 / 未知项可见度 | 影响 |
|---|---|---|---|---|
| Manifold AI | 未披露;可能是企业项目 / 软件加部署模式 | WorldScape、Policy、基准可信度、集成工作 | 定价、试点、续约高度未知 | 难以凭公开证据建模 SOM |
| Physical Intelligence | 未披露;开源加可能的企业合作 | VLA 模型、openpi、微调路径 | 商业合同条款未公开 | 开源降低采用摩擦,但定价仍不透明 |
| Wayve | 未披露的企业 / 伙伴模式 | 自动驾驶栈、投资人 / 伙伴生态 | 审阅来源未见公开合同价格 | 资本实力可抵消定价不透明 |
| AGIBOT / UBTECH | 主要是软硬件捆绑部署合同 | 机器人本体、软件栈、集成和服务 | 审阅来源基本未见标价 | 硬件捆绑可能让软件附加更难拆分 |
| X Square / ROBOTERA | 未披露的企业部署模式 | 基础模型、硬件、数据管线、场景落地 | 新闻稿强调规模和 PMF,而非合同条款 | 在没有公开价格发现的情况下竞争生产力叙事 |
该品类公开定价稀少;这本身就是竞争事实,因为不透明会抬高切换摩擦,也让标题式比较不可靠。
[CP024, CP026, CP030, CP035]3.4 护城河耐久性与替代风险
Manifold 当前的护城河真实存在,但很脆。WorldArena 和 WorldScore 认可可以帮公司打开工程师、研究者和早期设计伙伴的大门;与具身世界模型绑定,也给公司带来锋利的叙事优势。但同一组公开证据显示,这个市场的耐久优势正在模型、数据、硬件和部署渠道之间的整合闭环中形成。没有分发的基准领先只是可发现性,不是统治力。 反向证据强化了这一谨慎判断。BCG 认为行业常常过度解读亮眼演示,因为灵巧度和因果推理落后于感知。MERICS 也称,中国具身 AI 生态仍依赖 Nvidia,距离规模化充分自主还很远。整个同行集的公开定价、续约和客户集中度数据同样稀缺,这意味着许多竞争主张仍缺少硬运营验证。正确解读是,Manifold 拥有机会窗口,而不是已经稳固的护城河:它必须先把评估领先变成可重复的工业结果,否则模型层能力会更容易被复制,或被打包进更大的全栈平台。[CP024, CP025, CP028, CP029, CP031, CP032]
| 护城河主张 | 威胁 | 严重性 | 缓释措施 / 尽调问题 |
|---|---|---|---|
| 凭 WorldArena / WorldScore 拿到基准领先 | 若同业追平分数,或买方不再在意,可能变成营销同质化 | 高 | 把基准胜利绑定到付费部署结果和客户背书 |
| 模型原生软件叙事 | 全栈机器人供应商可能把类似能力与硬件和渠道捆绑 | 高 | 在硬件领导者内化模型层之前先建立合作 |
| 中国优先的具身 AI 定位 | 资金充足的对手可用 PR、试点和人才报价淹没市场 | 高 | 靠速度和技术可信度尽早锁定设计伙伴 |
| 硬件无关叙事 | 按具身形态微调仍可能为垂直整合对手创造隐性锁定优势 | 中 | 用第三方机器人展示跨具身形态证据 |
| 数据闭环优势 | 开源和共享算力生态可能压缩模型差异 | 中 | 证明自有数据质量和闭环改进速度 |
| 战略投资人信号 | 伙伴可转投或同时支持多个对手 | 中 | 记录排他性、渠道条款和部署义务 |
严重性衡量的是:如果基准领先不能在下一轮融资周期内转成运营和分销证据,Manifold 护城河会承受多大风险。
[CP023, CP028, CP031, CP032, CP033, CP034]Manifold 相对更广泛同业的竞争耐久性指标,一眼看清。
[CP023, CP024, CP031, CP038]3.5 证据摘录
04财务情况
4.1 收入模型与定价可见度
公开证据足以推断 Manifold 收入模型的形状,但不足以判断规模。官网和 2026 年 6 月融资报道把公司定位在世界模型、具身控制,以及向机器人和 XR 相关物理 AI 场景的部署上。这意味着收入很可能来自软件许可、集成和部署服务、伙伴驱动解决方案,以及可能与 WorldArena 生态绑定的基准或评估服务。公开来源没有给出任何订单、确认收入、ARR 或定价的硬数字。 与相邻公司相比,这个缺口更明显。Wayve 公开把 AI Driver 描述为车辆无关的软件平台,具备经常性软件经济性;AGIBOT 的商店页面让硬件价格可见,但实现 ASP 和合同条款仍不透明。Physical Intelligence 和 Manifold 呈现同一模式:技术叙事具体、融资信号强,但没有公开变现数据。按尽调口径,正确姿态不是认为 Manifold 没有商业模式,而是承认其商业模式在公开市场里仍未被定价。[CI001, CI003, CI004, CI005, CI022, CI024]
| 收入流 | 机制 | 单位 | 当前价值 / 状态 | 质量 | 尽调问题 |
|---|---|---|---|---|---|
| 模型授权 / 平台费 | 世界模型 / 动作模型栈的软件许可 | 企业合同 | 未披露 | 合理但未验证 | 索取合同样例和定价依据 |
| 部署 / 集成服务 | 定制、集成和上线支持 | 项目费 / 里程碑费 | 未披露 | 若付费部署存在,可能是当前收入桥梁 | 索取服务收入结构和利用率数据 |
| 基准 / 评测服务 | WorldArena 相关评测、合成数据或模型评估工作 | 项目 / 订阅 | 未披露 | 战略上合理,但公开资料未显示已货币化 | 澄清是否有任何基准活动收费 |
| 伙伴解决方案收入 | 与机器人或工业伙伴联合部署 | 收入分成 / 捆绑合同 | 未披露 | 鉴于 UBTECH 与投资人关系,存在可能 | 索取伙伴商业条款 |
| MR / XR 应用工作 | 混合现实场景中的物理 AI 应用 | 项目费 | 仅探索 | 当前信心低 | 在收入证据出现前视为邻近方向 |
每个收入流都是基于公开产品和场景主张的假设;没有任何一项有公开收入披露支撑。
[CI003, CI004, CI005]| 参照项 | 价格 / 单位 / 合同 | 标价与实际成交价 | 包含能力 | 来源可见度 | 对 Manifold 的含义 |
|---|---|---|---|---|---|
| Manifold AI | 未披露的企业定价 | Unknown | 世界模型、策略模型、部署工作 | 无公开价格 | 无法凭公开数据建模 ARR |
| AGIBOT X2 | US$24,240 标价 | 仅标价 | 娱乐 / 商用人形机器人硬件 | 官方店铺 | 展示相邻市场可见的硬件定价 |
| AGIBOT A2 Lite | US$44,560 标价 | 仅标价 | 全尺寸性能型人形机器人硬件 | 官方店铺 | 说明硬件 ASP 区间,不代表软件 ASP |
| AGIBOT A2 Ultra | 无公开标价;企业询价模式 | Unknown | 带支持和质保条款的认证企业级人形机器人 | 官方店铺 / 销售主导 | 企业交易很可能打包服务和支持 |
| Wayve AI Driver | 授权模式未披露 | Unknown | 不绑定车型的 AI 软件平台 | 官方产品页 | 最接近软件主导型经常性经济模型的公开参照 |
相邻市场定价只用作参照。硬件标价不能直接换算成 Manifold 的软件经济模型。
[CI022, CI023, CI024, CI025]如果部署成熟,技术资产可能如何转成经济收入流。
[CI003, CI004, CI005, CI006]4.2 成本结构与单位经济代理指标
Manifold 的实体资产负债表可能比机器人制造商更轻,但默认并不是低成本纯软件业务。主要成本桶很可能包括训练和推理算力、多模态数据采集和标注、仿真和评估基础设施、部署工程,以及伙伴支持。BCG 关于传统机器人 TCO 主要由设置和再工程主导的发现很有用,因为它解释了以模型为中心的供应商在哪里创造价值:减少工程劳动、加快换线,而不只是产出“更聪明的演示”。ResearchInChina 也从另一侧强化了这一点,称瓶颈已转向可扩展的高质量数据和数据运营。 代理比较有利有弊。UBTECH 的监管文件显示,即便收入增长、毛利率改善,规模化工业人形机器人公司仍可能出现有意义亏损,这提醒市场不要天真乐观地看待具身 AI 经济性。但同一份文件也提示了软件优先供应商的上行空间:如果 Manifold 能避开大额硬件 BOM、库存和制造成本,长期利润率结构应好于硬件主导同行。问题在服务。如果每次部署都需要定制调优、重度现场支持或劳动密集型数据工作,软件式毛利率可能永远无法充分出现。[CI008, CI009, CI010, CI011, CI020, CI021]
| 指标 | 数值 / 参照 | 置信度 | 重要性 | 尽调索取项 |
|---|---|---|---|---|
| 毛利率特征 | 可能高于硬件同行,但尚未披露 | 低 | 判断软件主导叙事能否经得起部署现实 | 索取分板块毛利率或综合贡献利润率 |
| 主要成本中心 | 算力 + 数据 + 部署工程 | 中 | 判断扩张是轻资本,还是被服务交付拖重 | 拆分 GPU、数据运营和现场支持支出 |
| 营运资金压力 | 可能低于硬件厂商,但不是零 | 低 | 预付款、伙伴硬件套件或云承诺仍会吃掉现金 | 索取预付云服务 / 硬件 / 数据采集义务 |
| 实施强度 | 未知;早期试点很可能不轻 | 低 | 服务交付过重会稀释软件利润率 | 索取每个试点和每个规模化站点的实施工时 |
| 基准测试到收入的转化 | Unknown | 低 | 判断技术领先是否有经济价值 | 提供付费基准测试、评估或转化漏斗数据 |
本表刻意依赖参照项,因为 Manifold AI 没有公开单位经济模型披露。
[CI008, CI009, CI010, CI011, CI034]定性梳理软件优先的具身 AI 利润率在算力和部署压力下会在哪里扩张、又会在哪里流失。
[CI008, CI009, CI010, CI011, CI029, CI035]有来源支撑的具身 AI 相邻经济性参考点;仅作背景,不代表 Manifold 经营结果。
来自固定来源的数值按点区间绘制。Manifold 融资区间采用公开的「接近 RMB1B」说法,因此低位边界偏保守,并非公司披露。
[CI014, CI020, CI022, CI024]4.3 资本充足性与融资依赖
融资故事很清楚,尽管运营故事还不清楚。Manifold 在大约一年内完成六轮融资,到 2026 年 6 月 Pre-A 累计融资接近 RMB1 billion;此前还有 2026 年春季近 RMB200 million 的 Pre-A 轮,以及随后更大规模融资。管理层称新资金将支持下一代模型研究、多模态基础设施,以及在物流、制造、汽车和 MR 相邻场景中的部署。这符合一家仍在为产品市场发现和基础设施建设投入资金、而不是收割成熟经常性收入的公司。 由于公开资料没有披露债务、信贷或项目融资义务,从外部看资本结构似乎由股权融资支撑。但资本充足性无法精确衡量,因为烧钱速度和现金跑道并不公开。实际情况是,下一轮融资触发点很可能更多取决于基准成功是否转化为可重复的工业部署和伙伴渠道,而不是会计里程碑。Embodied Global 的融资数据表明,资本仍会投向类别领导者;但这也意味着投资人会越来越要求运营证明,而不只是技术叙事。[CI013, CI014, CI015, CI016, CI017, CI018]
| 指标 | 数值 / 状态 | 置信度 | 重要性 | 尽调索取项 |
|---|---|---|---|---|
| 累计融资 | 截至 2026 年 6 月 Pre-A 融资近 RMB1B | 高 | 显示外部融资支持很强 | 核对各轮精确到账现金 |
| 最新资金用途 | 模型迭代、多模态基础设施、部署扩张 | 中 | 表明公司仍在建设期,不是收获期 | 梳理支出科目和时间节奏 |
| 债务 / 信贷额度 | 未见公开披露 | 中 | 没有数据室资料,资产负债表风险不能排除 | 索取债务明细表和债务契约摘要 |
| 资金续航月数 | Unknown | 低 | 没有现金消耗数据,就无法判断融资紧迫性 | 索取现金消耗趋势和账上现金状况 |
| 下一轮触发条件 | 可能是付费部署的可复制性 | 中 | 显示投资人下一步会要求哪类里程碑 | 索取董事会 KPI 材料和融资计划 |
融资总额让资本充足性部分可见,但现金消耗未披露,资金续航仍不透明。
[CI013, CI014, CI016, CI017, CI018, CI019]即便不做硬件制造,软件主导的具身 AI 创业公司仍可能依赖融资。
[CI016, CI018, CI019, CI031, CI036]4.4 财务结论与尽调阻断项
核心财务结论是,Manifold 更容易被承保为一项资本充足、押注具身 AI 领导地位的期权,而不是一个可量化的运营业务。资本基础可信,用途叙事连贯,竞争代理指标也显示,如果软件优先供应商能成为工业机器人里的模型和部署层,确实存在经济奖品。但完整财务视图所需的硬运营证据仍然缺席:价格实现、客户集中度、烧钱速度、毛利率、营运资金需求和续约行为。 这些缺失数据很重要,因为如果服务、集成和一次性试点主导收入结构,具身 AI 的收入质量会迅速变差。Manifold 最终可能证明自身经济性远好于硬件重同行,但公开来源尚未证明这个结果。在公司能展示付费部署可重复性之前,它应被视为高潜力但高不透明度的资本消耗者。因此,尽调优先级很简单:用价格、合同和单位经济证明,替代基准和融资证明。[CI002, CI006, CI007, CI032, CI033, CI034]
| 缺失的私有指标 | 对分析的影响 | 精确尽调路径 |
|---|---|---|
| 收入 / ARR | 无法评估估值或收入质量 | 索取月度收入、订单额和确认收入桥表 |
| 客户数量和集中度 | 无法判断管线耐久性 | 索取分阶段活跃客户数和头部客户占比 |
| 毛利率和贡献利润率 | 无法评估软件与服务的收入结构 | 索取按收入流划分的毛利率 |
| 现金消耗和资金续航 | 无法评估资本充足性 | 索取月度现金消耗和账上现金余额 |
| 定价和续约条款 | 无法分析 LTV/CAC 和回收期 | 索取合同样本和续约分群 |
| 部署利用率 / 可用时间 | 无法评估真实场景 ROI | 索取按站点的部署指标和支持负担 |
在把 Manifold 当作经营性公司、而不是技术期权来投资定价之前,至少需要补齐这些财务披露。
[CI001, CI002, CI007, CI017, CI033, CI036]4.5 证据摘录
05产品与技术
5.1 按工作流定义产品
Manifold 呈现的不是一张简单产品菜单,而是一套能力栈。按工作流看,公司出售的是一个系统,帮助机器人和其他物理 AI 智能体建模世界、预测结果,并在混乱真实环境中执行动作。公开描述一致把它放在物流、3C 制造、汽车制造和混合现实探索中。这意味着产品不是孤立的“给机器人用的视频生成”。它是客户工作流内部的赋能层,覆盖感知、规划、数据生成、策略评估和动作控制,而这些工作流都需要物理交互。 模块地图已经公开到足够可用。WorldScape 是实时空间世界模型。WorldScape Policy 是建立在该模型之上的动作和控制层。RoboScape 是机器人专用的模型 / 论文线。AirScape 把路线图延伸到无人机自主。WorldArena 是公开评估层,把公司偏好的功能性框定变成基准。这是一套连贯栈:基础模型、动作层、应用变体,以及一个基准外壳,在客户和研究社区面前强化公司的世界观。[CE001, CE002, CE003, CE011, CE012, CE013]
| 模块 / 资产 | 主要用户 | 状态 / 成熟度 | 差异化 | 尽调缺口 |
|---|---|---|---|---|
| WorldScape | 机器人 / 物理 AI 团队 | 已公开描述,生产成熟度未验证 | 覆盖移动 + 操作的实时世界模型 | 需要架构、部署和可靠性文档 |
| WorldScape Policy | 机器人控制 / 集成团队 | 已公开宣称,并出现在基准测试中 | 面向空间推理和控制的世界动作层 | 需要延迟、安全边界和动作失败处理资料 |
| RoboScape | 研究 + 产品化桥梁 | 论文和代码可见 | 融合物理先验的机器人世界模型 | 需要付费生产部署中的使用证据 |
| AirScape | 无人机 / 空中自主路线图 | 早期公开研究里程碑 | 将世界模型叙事延伸到空中领域 | 需要产品界面和商业化证据 |
| WorldArena | 研究者、模型厂商、企业评估方 | 公开网站 + GitHub + CVPR 挑战赛 | 与具身用例对齐的功能型基准 | 需要证明基准领先能预测现场结果 |
公开模块图把商业资产、研究线和评估基础设施混在一起,因为公司将其作为同一生态来营销。
[CE001, CE002, CE003, CE008, CE021]| 用户任务 | 当前工作流 | Manifold 方案 | 可衡量收益 | 局限 |
|---|---|---|---|---|
| 训练具身策略 | 采集稀缺机器人数据 + 手动仿真 | 用世界模型充当合成数据和策略评估引擎 | 策略迭代可能更快 | 无公开转化指标 |
| 在噪声场景中控制机器人 | VLA 或规则系统难以应对光照 / 背景变化 | 用 WorldScape Policy 做空间推理和动作生成 | 宣称对视觉噪声更稳健 | 无公开安全率或失败率数据 |
| 部署到物流 / 3C / 汽车站点 | 按场景做集成 | 将世界模型能力打包进物理 AI 方案 | 可能跨场景泛化 | 公开证据仍停留在场景层面,不到站点层面 |
| 为具身模型做基准测试 | 临时评估偏重视觉 | 用 WorldArena 做功能评估 | 更贴近真实具身任务 | 基准领先不一定等于部署成功 |
公开资料未披露客户 KPI 变化,因此收益只能按公司宣称或推断来表述。
[CE006, CE007, CE012, CE013, CE033]公开来源显示,技术栈从数据采集一路用到机器人部署。
[CE006, CE012, CE022, CE028]5.2 架构与数据闭环
公开技术证据显示,Manifold 的架构围绕功能型具身表现设计,而不只是追求照片级输出。WorldArena 衡量的不只是视频质量,还包括合成数据可用性、策略评估质量和动作规划价值。RoboScape 论文同样强调物理信息训练、时间深度预测和关键点动力学学习。Reportify 又补充了更大的架构叙事:LongScape 以及混合自回归 + DiT 方法支撑多个模型家族;Pedaily 则强调 MoE 结构,把指令跟随、移动交互和操作推理等子问题拆开。 同样重要的是数据闭环故事。公开报道称,公司拥有硬件-数据-模型闭环、自建数据采集设备、硬件在环强化学习,以及数十万小时数据。每个细节是否都被独立验证,并没有它对架构的暗示那么重要:Manifold 把评估、数据捕获和部署反馈视为产品栈的一部分,而不是离线支持功能。对具身 AI 来说,这是可信的设计哲学,因为产品质量取决于能否从模型输出走到机器人行为,再回到训练。[CE004, CE005, CE006, CE007, CE009, CE010]
| 层 / 组件 | 作用 | 依赖 | 风险 |
|---|---|---|---|
| 世界模型核心 | 预测时空世界状态 | 大规模训练数据和算力 | 离开训练域后,泛化可能变差 |
| 动作 / 策略层 | 将预测状态转成控制信号 | 机器人本体、传感器、校准 | 长尾执行或安全场景可能失效 |
| MoE / LongScape 训练设计 | 拆分任务并高效扩展 | 架构调优 + 路由质量 | 如果没有稳定运营证据,复杂度可能白白增加 |
| 硬件-数据-模型闭环 | 采集数据并改进模型 | 自有或合作方数据采集栈 | 数据权利和现场支持负担会拖慢扩张 |
| 基准 / 评估层 | 验证跨功能任务的效用 | 公开数据集、代码库和挑战赛运营 | 可能优化基准,而不是买方 KPI |
| 边缘部署 / 蒸馏 | 在已部署机器人或无人机上运行模型 | 高端边缘算力和优化 | 硬件瓶颈可能限制部署广度 |
架构是论文、访谈和公开基准界面的综合整理;公司没有发布完整工程图。
[CE004, CE005, CE009, CE010, CE022, CE027]基于公开信息,从数据采集到模型、动作和评测层,分层梳理 Manifold 技术栈。
[CE003, CE008, CE009, CE010, CE022]5.3 部署成熟度与依赖
Manifold 的公开成熟度并不均衡。研究侧异常清晰:基准仓库、公开论文和挑战赛基础设施都存在。部署侧更薄。公开来源称模型正用于物流、3C 制造和机器人场景,UBTECH 合作也显示 Manifold 希望运行在第三方机器人平台之上。Reportify 还称公司已经为边缘部署量化并蒸馏模型,包括机器人移动和无人机导航。合在一起,这些都是有意义的产品化信号,但仍达不到企业客户充分承保生产系统所需的水平:集成文档、支持边界、正常运行时间指标、回滚工作流和现场可靠性证据。 最重要的依赖也摆在明面上。产品需要强边缘算力,所以 NVIDIA 的 Jetson Thor 材料作为生态背景很重要。产品需要伙伴硬件或集成商渠道,所以 UBTECH 很重要。它还可能需要来自客户环境的特权数据流,所以不能只凭榜单排名判断部署成熟度。Manifold 在具身世界模型基准上可能已经技术领先于同行,但运营成熟度仍取决于公开记录之外的基础设施、硬件伙伴和部署纪律。[CE014, CE015, CE016, CE017, CE018, CE019]
| 日期 / 阶段 | 功能 / 里程碑 | 状态 | 含义 | 来源 |
|---|---|---|---|---|
| 2025-05 | 公司发布 + RoboScape 发布窗口 | 已公开宣布 | 研究速度从一开始就拉起来 | 官方网站 / RoboScape 论文 |
| 2025-07 | AirScape 公开发布 | 已公开宣布 | 路线图延伸到地面机器人之外 | 官方网站 |
| 2026-02 to 2026-03 | WorldArena 代码、排行榜、提交通道开放 | 已公开上线 | 评估栈开始产品化,并面向社区 | WorldArena 网站 / GitHub |
| 2026-04 to 2026-06 | CVPR 2026 WorldArena 挑战赛周期 | 已公开上线 | Manifold 将自己定位成基准制定者,而不只是参赛者 | QQ 文章 / WorldArena |
| 2026-07 | 与 UBTECH 合作推进电商和物流方案 | 已公开宣布 | 释放渠道主导部署的野心 | Gasgoo 合作报道 |
路线图证据在研究和基准里程碑上更强,在正式企业版本上更弱。
[CE020, CE021, CE028]公开可见的依赖会影响 Manifold 能否把基准领先转成部署可靠性。
[CE022, CE027, CE028, CE029]5.4 信任、安全与产品技术结论
信任与安全是这套栈里公开可见度最低的一层。Manifold 所处产品类别会影响机器人移动、动作规划和工业工作流,但公开界面没有暴露通常支撑企业信心的材料:安全认证、事故报告、安全控制、与数据采集绑定的隐私政策,或部署质量指标。这并不意味着这些控制不存在,而是它们不足以支撑公开尽调。AGIBOT 等相邻机器人厂商更清楚地发布硬件保修和认证表面,尽管这些材料也不能解决模型层安全。 因此,产品技术结论是两面的。一面,Manifold 看起来确实差异化:基准原生、研究产出高,并且罕见地明确强调功能型具身评估。另一面,它仍更像一个前沿技术平台,而不是文档完备的企业产品。对投资人或客户来说,下一步尽调不是争论模型是否有趣,而是要求公司披露可支持性、安全性、失效安全行为、数据治理和部署可靠性。[CE030, CE031, CE032, CE033, CE034, CE035]
| 控制 / 认证 / 质量指标 | 状态 | 范围 | 缺口 |
|---|---|---|---|
| 模型安全文档 | 未见公开资料 | Manifold 模型栈 | 需要风险控制、护栏和故障安全流程 |
| 安全 / 隐私文档 | 未见公开资料 | 数据采集和部署栈 | 需要安全架构和数据治理材料 |
| 正式产品认证 | 未见 Manifold 产品层面的公开资料 | 公司产品栈 | 需要适用认证或审计证据 |
| 事故历史 / 状态报告 | 未见公开资料 | 生产运营 | 需要可用时间和事故透明度 |
| 硬件伙伴认证 | AGIBOT 等相邻厂商可见 | 硬件层,不是模型层 | 不能替代 Manifold 自身控制 |
没有公开证据不等于不存在;但它是尽调阻断项。
[CE030, CE031, CE032]按序位观察当前公开信息里最强和最弱的能力。
评分来自分析师基于公开证据的判断,不是公司给出的分数。
[CE019, CE021, CE029, CE031, CE034, CE035]5.5 证据摘录
06客户情况
6.1 客户分层与买方地图
公开记录对用例很清楚,对账户很模糊。Manifold 一直与电商物流、3C 制造、汽车制造,以及更试探性的混合现实探索相连。这意味着企业客户基础的采购中心大概率在运营、工业自动化和创新预算里,而不是纯 IT 或消费者渠道。产品也不太可能由一线机器人操作员自己购买。这些团队是用户,制造集团、机器人平台所有者或工业战略方才更可能是买方和付款方。 另一类重要客户是间接客户:机器人 OEM 和生态伙伴。公开证据显示,Manifold 不只是直接卖进工厂或仓库,也在尝试把世界模型能力嵌入已经触达客户的伙伴平台。这一点重要,因为客户故事会从“Manifold 有多少 logo?”变成“进入市场路径有多少会被硬件或产业伙伴中介?”公开层面,后者比前者更容易证明。[CU001, CU002, CU003, CU013, CU033]
| 分群 | 买方 / 用户 / 付款方 | 用例 | 规模 / 战略价值 | 缺口 |
|---|---|---|---|---|
| 电商物流 | 买方:运营 / 自动化负责人;用户:机器人 / 仓储团队;付款方:企业资本开支或自动化预算 | 分拣、仓储、搬运、协同 | 公开反复出现最多的部署场景 | 无具名终端客户或站点数量 |
| 3C 制造 | 买方:工厂 / 自动化负责人;用户:产线 / 机器人工程师;付款方:工厂数字化预算 | 精密操作、柔性生产、检测 | 多个公开来源反复出现 | 无具名工厂或结果指标 |
| 汽车制造 | 买方:OEM / 供应商工业负责人;用户:机器人或自动化团队;付款方:工业转型预算 | 装配、检测、柔性自动化 | 战略重要,且贴近投资人关注点 | 公开证据较弱,更偏前瞻 |
| 机器人 OEM / 渠道伙伴 | 买方:机器人公司或集成商;用户:嵌入式方案团队;付款方:伙伴商业预算 | 将世界模型嵌入机器人产品 / 方案 | UBTECH 让这条商业化路径变得具体 | 经济条款和排他性未披露 |
| 研究 / 基准生态 | 买方不明;用户:模型开发者和实验室;付款方可能是内部研发 | 基准测试、评估、挑战赛参与 | 漏斗顶端信誉很强 | 不能证明存在付费客户 |
本表将直接终端市场与渠道、生态分开,因为 Manifold 公开客户路径看起来是混合型。
[CU001, CU002, CU003, CU013, CU014]公开证据显示,企业客户旅程大致从基准发现进入试点,再走向伙伴主导的规模化。
[CU003, CU020, CU021, CU024, CU032]6.2 公开采用证明
最强公开采用证明是与 UBTECH 的关系。盖世汽车和凤凰汽车都称,合作旨在把 Manifold 的世界模型与 UBTECH 的量产能力结合,从电商和物流开始推出可盈利的综合解决方案。UBTECH 自己的解决方案页面展示了匹配的工作流——仓储、包裹处理、分拣、巡检和装配——这让合作在商业上连贯,而不是表面化。这是有意义的证明,说明 Manifold 至少拥有一个具备现实客户邻近性的具名渠道。 UBTECH 之外,证据更分散。多个 2026 年 6 月媒体重复称,Manifold 技术已经部署或正在落地于电商物流和 3C 制造,汽车制造则被提为下一步工业目标。但这些文章没有点名底层终端客户。WorldArena 又提供了另一层需求证据:超过 200 份挑战赛提交,以及大型科技和研究团队参与,说明公司获得关注和技术兴趣;但这更适合归为生态采用,而不是客户采用。结论是,Manifold 的采用证明真实但分层:伙伴 / 渠道证据最强,场景级部署主张中等,具名终端账户证明较弱。[CU006, CU007, CU008, CU009, CU010, CU011]
| 指标 | 数值 | 日期 | 来源 | 置信度 | 含义 | 缺失分母 |
|---|---|---|---|---|---|---|
| 具名商业化渠道 | 宣布与 UBTECH 合作 | 2026-07-13 | Gasgoo / Phoenix Auto | 中 | 至少确认一条严肃的商业化渠道关系 | 无合同规模或部署数量 |
| 公开垂直场景证明:电商物流 | 在 2026 年 6—7 月多篇报道中反复出现 | 2026-06 to 2026-07 | Pedaily / Tencent / TMTPost / KuCoin | 中 | 显示真实商业化重心 | 没有账户数量或生产环境拆分 |
| 公开垂直证据:3C 制造 | 在 2026 年 6 月多篇报道中反复出现 | 2026-06 | Pedaily / Tencent / 同花顺 | 中 | 显示多垂直产品叙事 | 没有结果或客户名称 |
| 公开垂直证据:汽车制造 | 部署叙事中提到 | 2026-06 | Pedaily / Tencent / Leaderobot | 中低 | 指向战略邻近性 | 没有具名项目或站点 |
| 参与基准测试生态 | 200+ 次挑战提交 | 2026-03 至 2026-06 | WorldArena / QQ 挑战赛文章 | 中 | 技术兴趣信号强 | 不是付费客户指标 |
采用轨迹主要是一条证据覆盖时间线,不是已披露的客户数量序列。
[CU006, CU007, CU010, CU011, CU012, CU014]| 客户 / 对手方 | 细分领域 | 部署 / 用例 | 生产 / 试点 | 结果 | 限制 |
|---|---|---|---|---|---|
| UBTECH | 机器人 OEM / 商业化伙伴 | 将世界模型与量产人形机器人结合;从电商和物流切入 | 伙伴发布 / 商业化意向,不是终端客户生产证明 | 公开信息中最具体的具名商业化路径 | 没有合同金额、排他性或终端账户名称 |
| 未具名电商物流运营商 | 终端企业客户 | 公开资料称技术已落地物流场景 | Unknown | 多家媒体反复提到该场景 | 没有具名运营商、站点、KPI 或周期 |
| 未具名 3C 制造商 | 终端企业客户 | 公开资料称技术已落地 3C 制造 | Unknown | 多家媒体反复提到该场景 | 没有具名工厂、项目规模或续约证据 |
| 汽车制造合作方(未具名) | 工业目标 / 战略邻近领域 | 公开资料提到汽车制造部署或探索 | 可能处于早期 / 探索阶段 | 可能受汽车战略资本助推 | 没有具名 OEM 或生产项目 |
这份枚举刻意保留为不完整,因为公开记录中具名终端客户很少。
[CU006, CU007, CU010, CU011, CU012, CU019]一个定性证明漏斗:从广泛公共兴趣,收窄到少数证据更扎实的商业化信号。
数值是证明强度指数,不是客户数量。
[CU006, CU014, CU015, CU016, CU017, CU034]证据质量因维度不同而差异明显:渠道证明远强于终端客户结果证明。
矩阵取值是分析师基于公开记录具体程度作出的判断。
[CU006, CU008, CU010, CU011, CU012, CU014]6.3 耐久性、留存与集中度
耐久性是公开记录快速变薄的地方。没有所审阅来源披露活跃账户、部署数量、利用率、客户结果、合同期限、续约率、NRR 或流失率。因此,无法判断公司是否正从技术上亮眼的试点毕业,走向可重复的生产关系。2026 年 6 月和 7 月之间行业叙事保持连续,稍有帮助——它说明物流和制造是故事中的持续部分,而不是一次性公关实验——但不能证明同一批客户在续约或扩张。 可见渠道结构也带来集中风险。UBTECH 显然是最具体的具名商业化路径,BAIC 等战略投资方也指向少数具有产业重要性的交易对手,它们可能主导早期分发。短期看,这种集中度有好处,因为能加快接触设计伙伴。中期看,它有风险,因为客户扩张可能依赖少数交易对手、伙伴准备度和特定行业采购周期。公开层面,Manifold 更像拥有少数潜在强关系的公司,而不是拥有广泛披露装机基础的公司。[CU015, CU016, CU017, CU018, CU020, CU021]
| 指标 | 数值 / null | 细分领域 | 置信度 | 尽调问题 |
|---|---|---|---|---|
| NRR / GRR | Null | 全部 | 高 | 要求按细分领域提供分群留存或续约 |
| 合同期限 | Null | 全部 | 高 | 要求提供 MSA / SOW 样本期限 |
| 具名多站点扩张 | Null | 直接终端客户 | 中 | 要求提供站点铺开历史 |
| 部署使用率 | Null | 机器人部署 | 中 | 要求提供机器人运行小时数、任务成功率、介入率 |
| 客户口径 ROI / 满意度 | Null | 全部 | 高 | 要求提供客户背调访谈或 KPI 案例研究 |
公开资料不足以支撑真正的留存分析;这些 null 是有意保留,应在尽调中补齐。
[CU016, CU017, CU018, CU030]| 扩张驱动因素 | 集中度风险 | 影响 | 尽调路径 |
|---|---|---|---|
| 基准测试领先 | 可能带来兴趣,但转不成收入 | 中 | 将基准测试线索映射到实际试点 / 赢单 |
| UBTECH 渠道 | 渠道集中依赖一个可见伙伴 | 高 | 要求按伙伴拆分销售管线和收入 |
| 战略投资方 / 产业生态 | 可能把触达过度集中在少数行业或客户品牌 | 中高 | 要求按投资方 / 渠道拆分客户来源 |
| 多垂直平台叙事 | 可能把支持能力拉扯到多个行业 | 中 | 要求按垂直行业拆分实施团队配置 |
| OEM / 集成商主导的部署 | 可能遮住终端客户归属和续约可见度 | 高 | 要求提供终端账户清单和主合同所有方 |
Manifold 早期可能受益于集中度,但投资人需要弄清,这种集中是否会变成结构性依赖。
[CU020, CU021, CU022, CU023, CU024, CU032]真实留存数据未公开,因此本队列用公开关系可见度的延续性粗略代理耐久性。100 表示该关系或场景在相应时间段内仍保持公开可见。
这不是续约队列,只展示各证据面随时间延续的公开可见性。
[CU029, CU030]6.4 客户结论
今天做客户尽调,应区分战略拉力和账户级证明。Manifold 显然有战略拉力:工业领域反复出现,渠道伙伴可见,世界模型基准叙事正在吸引关注。但公司仍缺少把战略拉力转化为高置信客户逻辑的披露:具名终端客户、结果指标、部署数量、集中度细节和留存历史。 这并不会推翻客户故事,而是说明故事正在转场。公平解读是,Manifold 很可能已经越过纯研究新奇性的门槛,进入真实工业互动;但还没有越过可透明验证客户耐久性的门槛。对投资人来说,下一步不是问客户是否以某种形式存在,而是要求数据室回答:多少客户、多活跃、多粘、多集中。[CU024, CU025, CU026, CU034, CU035]
6.5 证据摘录
07风险
7.1 监管与法律风险
具身 AI 已经不再在真空里推进。中国 2026 年人形机器人与具身智能标准体系,意味着安全、伦理、应用和全生命周期要求正在走向制度化。杭州地方监管又给出第二个路标:政府开始为具身机器人部署定义测试、可追溯和商业化框架。Manifold 即便仍处早期,也会受影响:它公开讲述的产品故事已经延伸到机器人、仓储、制造产线和汽车相关工作流——这些领域一旦部署不再只是实验,政策关注通常会升温。 全球模式也指向同一方向。英国 Automated Vehicles Act 和欧盟 AI Act 并不是直接针对 Manifold 的规则,但能看出责任归属、问责和高风险 AI 合规讨论会往哪里走。最重要的风险不是 Manifold 现在违反了某套已披露规则;而是它的公开材料还没拿出能让未来合规更顺的安全、可审计性、隐私和治理证据。眼下没有公开诉讼是好事,但远不如证明 Manifold 已做好治理准备重要。[CR002, CR003, CR004, CR005, CR006, CR007]
| 规则 / 许可 / 案件 | 司法辖区 | 状态 | 可能性 | 严重性 | 缓释 | 剩余暴露 | 尽调路径 |
|---|---|---|---|---|---|---|---|
| 具身智能标准体系 2026 | 中国 | 政策框架已启动 | 中高 | 高 | 将产品控制映射到安全 / 伦理 / 应用标准 | 高 | 要求对照 2026 标准体系做合规差距分析 |
| 具身机器人地方监管 | 杭州 | 地方框架已生效 | 中 | 中高 | 在适用地区使用沙盒 / 地方政策路径 | 中高 | 要求逐地区部署合规计划 |
| EU AI Act 高风险义务 | EU / 出口市场 | 扩张触及相关系统 / 市场后适用 | 中 | 中高 | 限制范围,或尽早搭建可审计性 | 中 | 要求提供跨境产品分类备忘录 |
| 自主系统问责框架 | 英国及其他市场 | 具风向标意义 | 短期中低 | 中 | 跟踪责任与授权义务 | 中 | 要求律师评估未来问责暴露 |
| 公开诉讼 / IP 纠纷 | 全球 | 公开未见 | 当前低 | 若隐匿则中 | 尽早留存来源与合同文件 | 中 | 要求审查 IP 权利链和训练数据权利 |
行按严重性和战略相关性排序,而不是按执法马上落地的确定性排序。
[CR002, CR004, CR006, CR007, CR025, CR026]公开证据最薄、系统复杂度最高的地方,残余风险最高。
评分为分析师截至 2026-08-03 基于公开证据综合作出的判断。
[CR001, CR008, CR015, CR017, CR020, CR038]7.2 运营与技术风险
从运营上看,Manifold 公开证据最强的一面仍是技术验证。它在公开基准中领先或排名靠前,发布与研究相连的基础设施,并把产品叙事落在具身功能效用,而不只是视觉生成。但这些都不能证明机器人在客户现场能安全、可靠、经济地运行。已审阅来源没有披露正常运行时间、人工介入率、任务成功率或事故历史。核心运营风险就在这里:公开记录充分证明模型能在受控评测里做什么,却很少证明组织能在真实工业环境中反复交付什么。 还有一个更隐蔽的技术风险:基准漂移。WorldArena 比纯审美基准更有用,但仍可能变成团队优化的靶子。如果 WorldArena、WorldScore 和相关榜单上的成绩,不能与现场层面的 KPI——吞吐、错误率、节省人工——高度相关,Manifold 可能赢下品类叙事,却仍让买家失望。再叠加集成复杂度、硬件差异和边缘推理约束,“更好的模型”与“更好的客户结果”之间的距离就成了主要运营断层。[CR001, CR010, CR011, CR012, CR013, CR014]
| 失效模式 | 可能性 | 严重性 | 缓释成熟度 | 剩余暴露 | 未解缺口 |
|---|---|---|---|---|---|
| 基准测试无法转化为现场 KPI | 中高 | 高 | 中低 | 高 | 需要站点级结果证明 |
| 部署可靠性弱于宣传 | 中 | 高 | 低 | 高 | 未公开正常运行时间 / 介入 / 事故数据 |
| 模型安全 / 失效安全控制不足 | 中 | 高 | 低 | 高 | 未公开安全论证或回滚流程 |
| 采集闭环中的数据治理或隐私失效 | 中 | 中高 | 低 | 中高 | 未公开数据权利 / 隐私架构 |
| 边缘推理瓶颈拖累现场表现 | 中 | 中 | 中低 | 中 | 需要延迟 / 硬件移植证据 |
这份清单反映的是研究证明与真实工业运营之间的缺口。
[CR008, CR009, CR011, CR012, CR013, CR014]技术和信任风险如何传导到客户证明、收入质量、融资和估值。
[CR001, CR011, CR013, CR020, CR033]7.3 依赖、财务与人才风险
Manifold 目前的商业化打法重度依赖外部。最清楚的商业化路径是通过 UBTECH,这马上带来渠道、节奏和合作伙伴健康度依赖。更大的品类也依赖高端算力,MERICS 和 NVIDIA 都在强化一点:先进边缘硬件仍是核心。在此之上,还有公司自己的数据闭环逻辑:护城河很大一部分取决于能否持续进入客户环境、传感器和现场数据。算力、合作伙伴硬件或数据权利任一输入受限,即便核心模型继续进步,商业化也可能卡住。 财务上,风险更多是预期压力,而不是资不抵债。现有资本缓解了近期生存风险,但一轮又一轮融资如果没有运营指标披露,下一轮举证压力会更大。人才层面,公司也显得依赖关键人物:公开叙事高度绑定创始人吴维、顶尖研究履历和稀缺世界模型人才。搭建楔子时这是优势;但如果工业交付、客户运营和安全治理落后于研究职能,它也会变成扩张风险。[CR015, CR016, CR017, CR018, CR019, CR020]
| 依赖 | 对手方 | 角色 | 集中度 | 失效场景 | 严重性 | 缓释 | 剩余暴露 |
|---|---|---|---|---|---|---|---|
| 商业化渠道 | UBTECH | 机器人平台和市场触达伙伴 | 可见集中度高 | 渠道延期、优先级调整或转化疲弱 | 高 | 分散 OEM / 集成商渠道 | 高 |
| 算力平台 | NVIDIA | 边缘计算 / 生态依赖 | 品类层面 | 硬件成本、供给或平台错配拖慢部署 | 高 | 逐步适配多条硬件路径 | 中高 |
| 客户现场数据访问 | 工业客户 / 伙伴 | 支撑数据闭环和迭代 | 中高 | 客户限制数据采集或复用 | 中高 | 把数据权利和备用数据集写进合同 | 中高 |
| 基准测试叙事框架 | WorldArena / 公开排行榜 | 发现和验证入口 | 中 | 叙事成功掩盖商业疲弱 | 中 | 把基准测试结果绑定现场 KPI | 中 |
| 战略投资方 / 产业生态 | BAIC 等产业投资方 | 资源触达和信号背书 | 中 | 过度依赖少数战略网络 | 中 | 拓宽直销和客户背书 | 中 |
这些依赖本身不是坏事;如果管理层无法逐步分散,才会变危险。
[CR015, CR017, CR018, CR019, CR022]| 角色 / 职能 | 依赖或缺口 | 可能性 | 严重性 | 缓释 | 尽调路径 |
|---|---|---|---|---|---|
| 创始人 / CEO 叙事 | 公开叙事与 Wu Wei 高度绑定 | 中 | 高 | 扩充面向客户的管理层梯队 | 会见更广泛的管理团队 |
| 前沿模型研究员 | 稀缺世界模型人才集中 | 中 | 中高 | 留任激励和文档化 | 要求组织架构图和留任计划 |
| 工业交付职能 | 可能落后于研究能力 | 中高 | 高 | 补强运营 / QA / 部署负责人 | 要求部署组织 KPI |
| 信任 / 治理负责人 | 公开不可见 | 中 | 高 | 明确安全 / 隐私负责人 | 要求确认 Responsible AI 和安全负责人 |
| 客户成功 / 现场支持 | 成熟度不明 | 中 | 中高 | 建立支持指标和复盘机制 | 要求支持流程和 SLA |
具身 AI 的执行风险常来自运营团队,而不是模型团队。
[CR027, CR028, CR036]最重要的依赖集中在伙伴渠道、算力、数据访问和公开基准。
[CR015, CR017, CR019, CR022, CR031]7.4 缓释措施、监测与终止标准
好消息是,Manifold 最重要的许多风险都能监测。投资人不必马上拿到完美披露,也能知道下一步该问什么:是否出现具名生产部署,UBTECH 及类似渠道是否真正扩张,安全与信任文档是否可见,公司能否披露哪怕基础的可靠性层,而不伤害叙事。这些都是可量化检查点。 坏消息是,投资逻辑可能在灾难性失败前就先失效。如果下一次融资到来时仍缺少更强客户证明,或合作伙伴渠道看起来有前景却无法规模化,或治理要求上升得比公司的信任基础设施更快,下行不会表现为戏剧性崩塌,而会是商业化乏力。因此,正确的风险判断是升高但可监测。Manifold 值得尽调,不是因为它看起来弱,而是因为它正在跨过前沿具身 AI 最容易失败的关口:从亮眼模型领先,走向可问责的工业系统。[CR030, CR031, CR032, CR033, CR034, CR035]
| 风险 | 可监控触发项 | 阈值 / 事件 | 行动含义 |
|---|---|---|---|
| 客户证明失败 | 下次融资前仍没有具名生产部署 | 仍只有伙伴发布和行业口径 | 转为继续研究 / 避免只为叙事付费 |
| 渠道失败 | UBTECH 关系未扩展为可见客户项目 | 2-3 个季度内没有后续部署证据 | 判定 GTM 去风险失败 |
| 治理失败 | 部署扩大后仍没有公开信任 / 安全文档 | 仍无安全、隐私或回滚文档 | 提高所需折价,或暂停尽调 |
| 运营失败 | 出现生产部署可靠性事故,或 KPI 转化表现差 | 出现具名失败案例,或无法展示可靠性仪表盘 | 将基准测试切入口视为不足 |
| 合规失败 | 新合规义务实质性拖慢部署 | 审计 / 安全要求导致商业化时间表后移 | 下调乐观 / 基准情景概率 |
这些触发因素设计成可通过公司后续披露、客户访谈和市场进展持续观察。
[CR030, CR031, CR032, CR033, CR034, CR035]7.5 附录图表
08估值
8.1 投资逻辑 vs. 反向逻辑
Manifold 的投资逻辑很直接:公司似乎在具身 AI 中占住了稀缺战略位置。它较早在中国推进世界模型商业化,基准领先比单纯视频美学更有意义,也已经融到足够资本,在许多同行还在定义楔子时继续投入建设。公开来源还显示,它与物流、制造、机器人等场景存在可信相关性,并通过 UBTECH 拥有可见商业化渠道。如果具身 AI 成为大型工业品类,站在模型与评估层的公司,相对当前规模可能获得不成比例的话语权。 反向逻辑同样直接:在品类经济性被证明之前,Manifold 可能已经按品类赢家定价。今天几乎所有高质量公开信号都在持久运营证明的上游——融资轮、基准、技术论文和伙伴公告。真正稀缺的是能告诉投资人这些信号是否滚成真实业务的数字。这个不对称让公司值得跟踪,也让追高变得危险。溢价并非不理性,只是验证不足。[CV001, CV002, CV003, CV021, CV022]
| 建议 | 置信度 | 风险评级 | 估值立场 | 决策含义 |
|---|---|---|---|---|
| 观察 / 继续研究 | 中 | 高 | 当前独角兽估值可以理解;公开证据不支持在其上方给溢价 | 保持跟踪,但出手前要么拿到更好条款,要么看到更多证据 |
这是价格敏感型建议,不是泛泛的质量评分。
[CV004, CV005, CV006, CV007, CV040]| 论点 | 什么会改变判断 |
|---|---|
| 世界模型领先地位可能变成具身 AI 的基础控制层 | 具名生产部署和类软件经济性会强化这一判断 |
| 基准测试领先可能高估商业化就绪度 | 客户 KPI 证据和可靠性仪表盘会削弱这一担忧 |
| 合作伙伴主导分发可以加速放量 | 更广的多伙伴客户基础会让这一点更稳 |
| 当前独角兽估值可能已经计入过多未来成功 | 更低进入价格或更清晰经营证据会改善入场条件 |
反向逻辑针对的是时点和价格,不是否定技术。
[CV001, CV002, CV003, CV021, CV022]公开证据先支持战略关注,估值信心排在其次。
[CV001, CV002, CV004, CV005, CV006, CV040]8.2 估值背景与入场纪律
当前估值故事有一个可信但不完整的锚。可信,是因为多家 2026 年 6 月来源称,Manifold 在 Pre-A 累计融资接近 RMB1 billion 后进入独角兽层级。不完整,是因为公开信息仍没有收入披露、股权结构细节或优先股堆叠清晰度。也就是说,投资人应把当前估值标记看作战略期权价值,而不是已完整承销的后期运营估值。它更像对一个潜在重要基础设施层的早期索取权。 检验这个期权价值,最好用可比公司施压。Physical Intelligence 这类软件优先的机器人智能公司说明,当投资人相信模型层可以横跨多种具身形态时,上行溢价可以有多大。UBTECH、AgiBot、Unitree 这类硬件较重的公司则提供反向纪律:真实部署、制造规模和最终公开市场审视,会迫使市场更尖锐地追问利润率、可靠性和复购需求。Manifold 介于这两类可比公司之间,也正因为如此,仅凭公开证据很难为激进定价辩护。[CV007, CV008, CV009, CV010, CV011, CV012]
| 可比对象 | 指标 | 倍数 / 估值 / 状态 | 相关性 | 局限 |
|---|---|---|---|---|
| Physical Intelligence | 私有市场估值 | ~US$5.6B,2025 Series B 后 | 最佳软件优先机器人基础模型可比对象 | 资本更多,全球投资人深度更强,但仍是私有公司 |
| Wayve | 融资 + 平台状态 | 累计融资 US$2.8B;软件主导自动驾驶平台 | 可用于参照经常性软件收入愿景 | 汽车自动驾驶是不同产品和买方路径 |
| UBTECH | 公开收入 + 亏损 | 2025 年收入 RMB2.0B;净亏损 RMB789.8M | 检验具身 AI 部署经济性的最佳硬参照 | 硬件重的经济性不能直接套用 |
| AgiBot | IPO 目标 / 私有市场信号 | 据报道 IPO 目标估值 ~US$5.1B-6.4B | 带部署叙事的中国具身 AI 可比对象 | 目标估值,不是已经成交的市场出清价 |
| Unitree | IPO 目标 / 进程里程碑 | 据报道围绕科创板上市路径的目标估值约 ~US$5.9B-6.0B | 近期中国公开市场参照点 | 仍需完整公开市场验证 |
| Manifold AI | 当前私有市场叙事锚 | 2026 年 6 月融资后形成独角兽 / ~US$1B+ 叙事 | 入场纪律的直接锚点 | 没有公开收入或股权结构表支撑 |
合适的可比框架是混合的;没有单一可比家族能同时覆盖上行空间和当前不透明度。
[CV011, CV012, CV013, CV014, CV015, CV016]| 触发因素 | 阈值 | 如何传导到投资逻辑 | 行动含义 |
|---|---|---|---|
| 到下一轮融资时仍无具名生产部署 | 仍只有基准测试胜出 + 伙伴公关稿 | 削弱商业转化投资逻辑 | 避免溢价进入,或完全后退 |
| 渠道未扩展到 UBTECH / 少数战略方之外 | 看不到第二条市场路径 | 提高集中度和议价风险 | 给更深折价,或继续等待 |
| 部署扩展后,治理 / 安全文档仍缺位 | 即便声称在扩张,信任界面仍没有 | 抬高合规和可靠性风险 | 暂停,直到问题解决 |
| 商业指标仍未披露 | 收入、使用率、客户数都不可见 | 无法为估值做论证 | 把当前估值只当叙事看待 |
否决触发因素设计成可在后续融资或部署披露中观察到。
[CV027, CV035, CV036, CV037]示意估值结果更取决于商业化证明,远不只是技术叙事。
条形为十亿美元口径的情景输出,不是市场报价。
[CV007, CV008, CV023, CV024, CV025, CV029]8.3 乐观 / 基准 / 悲观情景框架
Manifold 的乐观情景不能只靠“AI 继续火”。它需要具名生产部署,需要证明公司能超越一两个战略渠道扩张,还需要让人相信模型层能拿到软件式经济性,而不是永远变成服务占比很高的集成工作。若这些条件出现,公司可以向前沿具身 AI 可比公司的上沿重估,但很可能仍低于最成熟或资本最充足的私人领导者。 基准情景更克制:Manifold 仍具战略重要性,也能拿到融资,但在更好的客户证据出现前,当前独角兽叙事已经吸收了大部分已知上行。悲观情景不是技术失败,而是商业化滞后。如果下一轮融资早于具名部署、收入或可靠性证据,公司可能会发现,战略兴奋感挡不住估值压缩。换句话说,情景分歧不主要取决于原始模型质量,而取决于公司何时能证明谁付钱、多久付一次、利润率是多少。[CV019, CV020, CV023, CV024, CV025, CV026]
| 情景 | 假设 | 估值 / 回报逻辑 | 关键风险 | 概率信号 |
|---|---|---|---|---|
| 乐观 | 具名生产客户、伙伴范围更广、类软件利润率更清晰、治理界面更强 | 支持估值重估至私有具身 AI 上沿区间,但仍低于最大类别龙头 | 执行仍难,但证据追上叙事 | 需要多项公开证据升级 |
| 基准 | 基准测试领先保持,融资仍拿得到,但客户和收入证据只逐步改善 | 当前独角兽估值锚大致守住,并带有限上行期权 | 经济性显现前,叙事可能先进入平台期 | 最符合当前公开证据 |
| 悲观 | 商业化滞后,伙伴集中度持续,下一轮融资时仍没有更强证据 | 有平轮或下轮风险;公开市场式可比折价占主导 | 运营指标出来前,价格先重置 | 持续不透明叠加部署证据放慢会触发 |
情景逻辑按里程碑搭建,因为公开证据不足以支撑精确现金流建模。
[CV023, CV024, CV025, CV026, CV027, CV028]基于当前公开证据,只能支撑一个较宽的估值区间。
区间刻意拉宽,因为收入、利润率、稀释和退出时点仍未披露。
[CV023, CV024, CV025, CV026, CV027, CV028]8.4 建议与最终门槛
因此,建议不是“否”,而是“还不能不问价格就进”。有纪律的投资人可以继续贴近这条线索,因为上行真实存在,公司也可能仍处在一个大品类的早期。但同样有纪律的投资人,不应只凭基准胜出、融资势能和伙伴叙事,就按高溢价后期估值承销。门槛很具体:具名生产客户、基础运营指标、伙伴经济性、股权结构清晰度和治理准备度。 如果这些拼图改善,Manifold 可以从有趣的战略期权,升级为可投资的后期成长故事。若没有,当前叙事很可能跑在证据前面。因此,基于公开信息的正确结论是观察 / 继续研究,并严格守住价格纪律:保持关注,但让下一批运营证明——而不是错失恐惧——决定入场。[CV004, CV005, CV006, CV034, CV035, CV036]
| 主题 | 缺失证据 | 为什么重要 | 负责人或尽调路径 |
|---|---|---|---|
| 收入 / 订单 | 月收入、ARR、订单和结构 | 用来判断业务更像软件、服务还是试点 | 财务数据室 |
| 具名部署证据 | 终端客户名单、阶段和客户访谈 | 用来验证客户耐久性和集中度 | 销售 / 客户成功尽调 |
| 伙伴经济性 | UBTECH 及其他伙伴合同、收入分成、排他性 | 用来验证渠道依赖和利润流失 | 商务拓展 / 法务 |
| 股权结构表与优先权 | 逐轮股权结构表、清算优先权、按比例跟投权和附函 | 用来判断当前估值是否具备经济上的可投性 | 财务 / 法务 |
| 治理 / 安全控制 | 安全、隐私、回滚和事故流程 | 用来验证监管和运营就绪度 | 产品 / 风险 / 安全尽调 |
没有这五项,最终给出肯定的 IC 建议还为时过早。
[CV009, CV010, CV031, CV034, CV038, CV040]按投委会式打分看,Manifold AI 志向很高,验证结果参差,价格支撑还不完整。
分数为编辑性 1-10 分判断,依据截至 2026-08-03 留存的公开证据。
[CV005, CV006, CV018, CV038, CV040]8.5 附录图表
免责声明
本报告依据截至 2026-08-03 的公开来源。Manifold AI 是私营公司,财务、客户和治理披露明显有限。任何投资决定都应以一手尽调为前提, 包括客户访谈、管理账、合作伙伴合同以及完整股权结构审阅。
证据索引
| 编号 | 陈述 | 可信度 | 来源 |
|---|---|---|---|
| CO001 | Beijing 流形空间科技有限公司 (Manifold AI) was founded on 2025-05-22 and began presenting itself publicly in late May 2025. | 高 | SO014, SO021 |
| CO002 | Manifold AI uses the Chinese name 流形空间 and publicly brands itself in English as Manifold AI. | 高 | SO021, SO014 |
| CO003 | The company describes itself as building next-generation world models and applying them to AI hardware use cases such as robotics and XR equipment. | 中 | SO021 |
| CO004 | Independent June 2026 coverage describes Manifold AI as China’s first startup to use a self-developed world model as the backbone for embodied intelligence. | 中 | SO002, SO003, SO007 |
| CO005 | WorldScape is described as a real-time world model that supports both mobility and manipulation interactions in one framework. | 中 | SO003, SO007, SO013 |
| CO006 | Public descriptions of WorldScape emphasize a mixture-of-experts architecture that routes sub-tasks such as instruction following, navigation, and manipulation reasoning to different experts. | 中 | SO003, SO007 |
| CO007 | Gasgoo and Tencent coverage say WorldScape held the top WorldScore rank for roughly two months while using about one-tenth the parameters of leading rivals. | 中 | SO003, SO007, SO013, SO024 |
| CO008 | Tencent's April 2026 WorldScore article says the competing field included World Labs, MIT, Alibaba, Zhipu, MiniMax, Runway, and Tencent Hunyuan. | 低 | SO013 |
| CO009 | WorldScape Policy uses world-model state prediction plus visual input to support action execution in embodied tasks. | 中 | SO007, SO024 |
| CO010 | Reviewed sources say WorldScape Policy outperformed existing VLA models in closed-loop tests and retained robustness under lighting, background, and object-position disturbance. | 中 | SO003, SO007, SO024 |
| CO011 | Manifold AI and collaborators launched WorldArena as a unified benchmark for embodied world models. | 高 | SO003, SO009, SO025 |
| CO012 | WorldArena evaluates models across six sub-dimensions and 16 metrics, then extends evaluation into synthetic-data, policy-evaluation, and action-planning uses. | 高 | SO009, SO019, SO025 |
| CO013 | The CVPR 2026 WorldArena Challenge was co-led by AMap CV Lab, Manifold AI, and Tsinghua University with participation from other global research institutions. | 中 | SO019 |
| CO014 | Dr. Wu Wei is publicly identified as founder and CEO of Manifold AI. | 中 | SO001, SO003, SO007 |
| CO015 | Reviewed sources consistently describe Wu Wei as a former SenseTime executive who previously led world-model-related work. | 中 | SO001, SO003, SO013 |
| CO016 | Multiple company profiles state that Wu Wei led teams to back-to-back first-place finishes in the Waymo SimAgents Challenge. | 中 | SO001, SO003, SO007 |
| CO017 | The founding roster combines researchers from Tsinghua University's FIB Lab with industry practitioners who previously deployed world models commercially. | 中 | SO003, SO007, SO014 |
| CO018 | Public materials say the team includes a Tsinghua professor/Changjiang Scholar and former big-tech world-model leads from autonomous-driving and AI companies. | 中 | SO007, SO013, SO014 |
| CO019 | Public profiles claim the broader team has authored more than 200 top-tier papers and accumulated more than 100,000 citations. | 中 | SO003, SO007 |
| CO020 | By June 2026, Manifold AI had completed six financing rounds in roughly one year. | 中 | SO001, SO002, SO003, SO007 |
| CO021 | Public June 2026 coverage places cumulative Pre-A funding at nearly RMB 1 billion. | 中 | SO001, SO002, SO003, SO007, SO008 |
| CO022 | Named new investors in the June 2026 round included Guoxin Fund, Yifeng Capital under Temasek, BAIC Capital/industrial investment, and Xinneng Venture Capital. | 中 | SO001, SO002, SO003, SO007 |
| CO023 | Earlier investors publicly named across 2025-2026 coverage include Legend Capital and Huawei Hubble. | 中 | SO003, SO007 |
| CO024 | A near-RMB-200 million Pre-A round was publicly reported in spring 2026, led by Huakong Fund and Xichuangtou with Datai Capital joining and prior investors topping up. | 中 | SO005, SO014 |
| CO025 | Tencent's April 2026 profile says a later Pre-A+ round of several hundred million yuan was led by Shunxi Fund with Yinxinggu, Fosun RZ, Jinyu Maowu, and 同创伟业 following. | 中 | SO013, SO014 |
| CO026 | QQ, Sina, and Gasgoo coverage explicitly place Manifold AI in the billion-dollar unicorn category by June 2026. | 中 | SO003, SO007, SO008 |
| CO027 | No reviewed public source disclosed revenue, ARR, or run-rate for Manifold AI as of 2026-08-03. | 中 | SO003, SO007, SO014 |
| CO028 | No reviewed public source disclosed a precise headcount for Manifold AI as of 2026-08-03. | 中 | SO003, SO007, SO014 |
| CO029 | Reviewed public materials did not disclose board composition, independent directors, or detailed governance terms. | 中 | SO014, SO021 |
| CO030 | Management said June 2026 proceeds would fund next-generation technical frameworks, multimodal world models, and infrastructure platform construction. | 中 | SO002, SO003, SO007 |
| CO031 | Manifold AI says WorldScape and WorldScape Policy have already landed in e-commerce logistics and 3C manufacturing scenarios. | 中 | SO003, SO007, SO024 |
| CO032 | The company also frames automotive production and mixed-reality entertainment as next deployment targets rather than already-verified production-scale customer wins. | 中 | SO003, SO004, SO007 |
| CO033 | On 2026-07-13, Manifold AI and UBTECH announced a strategic partnership to commercialize world-model solutions, starting with e-commerce and logistics. | 中 | SO018 |
| CO034 | The official website records three dated public milestones: company online on 2025-05-22, RoboScape on 2025-07-03, and AirScape on 2025-11-06. | 中 | SO021 |
| CO035 | The RoboScape paper presents a physics-informed embodied world model aimed at realistic robot-video generation plus downstream policy training and evaluation. | 高 | SO022, SO023 |
| CO036 | The WorldArena paper benchmarked 14 representative models and concluded that high visual quality often fails to translate into strong embodied-task capability. | 高 | SO009, SO025 |
| CO037 | MERICS argues that China's embodied-AI sector still depends heavily on Nvidia's AI chips and software ecosystem and remains far from fully autonomous large-scale deployment. | 中 | SO015 |
| CO038 | 36Kr Research Institute estimated China's embodied-intelligence market reached RMB 915 billion in 2025 and could exceed RMB 1 trillion in 2026, helping explain continued capital inflows into startups like Manifold AI. | 中 | SO016 |
| CO039 | Public materials name robotics and XR equipment as the company's target hardware surfaces, but do not yet support an exact customer count. | 中 | SO021, SO014 |
| CO040 | The WorldScore benchmark exists as a public paper-backed leaderboard, which strengthens but does not independently verify the company's self-reported rank framing. | 中 | SO011, SO012, SO020 |
| CO041 | Tencent's June 2026 profile says the CVPR 2026 WorldArena Challenge accumulated more than 200 submissions from companies, universities, and open-source teams. | 低 | SO007 |
| CO042 | Baidu Baike says the company built a family of domain models including DriveScape, RoboScape, and AirScape that span outdoor, indoor, and aerial settings. | 低 | SO014, SO013 |
| CO043 | Baidu Baike lists Wu Wei as legal representative and gives registered capital of RMB 2.5237 million, but this registry-style fact was not corroborated by another reviewed primary source in-session. | 低 | SO014 |
| CO044 | Exact debt facilities, secondary sales, and the full current cap table were not publicly disclosed in reviewed materials. | 中 | SO003, SO007, SO014 |
| CM001 | For diligence purposes, Manifold AI sits in the embodied-AI enablement market: world models, action models, data pipelines, and deployment software that help physical systems perceive and act in the real world. | 中 | SM006, SM012, SM010 |
| CM002 | That market boundary excludes pure consumer gadgets and generic video-generation models that do not claim embodied-task utility. | 中 | SM006, SM012 |
| CM003 | 36Kr Research Institute estimates China's embodied-intelligence market grew from RMB213.3 billion in 2018 to RMB915 billion in 2025 and could exceed RMB1 trillion in 2026. | 中 | SM001 |
| CM004 | Embodied Global's H1 2026 tally puts China embodied-AI financing at roughly RMB93.5 billion across 322 deals, about five times 2025 value. | 中 | SM003 |
| CM005 | ResearchInChina estimates the narrower China EAI data market reached RMB500 million in 2025, up 203% year over year, with China holding about 40% of the global market. | 中 | SM002 |
| CM006 | BCG defines physical AI as hardware-agnostic robotic capability rather than any single form factor, emphasizing perception, manipulation, planning, and reasoning as the real strategic axes. | 中 | SM006 |
| CM007 | BCG's framework places causal world-model reasoning at a frontier level that remains largely aspirational today, while perception and targeted manipulation already create economic value. | 中 | SM006 |
| CM008 | MERICS argues China's strengths in industrial robotics, EV supply chains, and state-backed industrial policy give it structural advantages in embodied AI. | 中 | SM004 |
| CM009 | MERICS also argues that Chinese embodied-AI firms remain dependent on Nvidia chips and software and that cost, precision, and autonomy constraints still limit deployment. | 中 | SM004 |
| CM010 | China Economic Net reports that China had more than 140 humanoid manufacturers and more than 330 humanoid models disclosed in the prior year. | 中 | SM005 |
| CM011 | The same China Economic Net article says AgiBot, Unitree, and UBTECH together accounted for nearly 80% of global humanoid shipments in 2025, with AgiBot alone at 39%. | 中 | SM005 |
| CM012 | 36Kr attributes demand growth to labor shortages, aging demographics, and the limits of fixed traditional automation in flexible manufacturing and human-machine collaboration. | 中 | SM001 |
| CM013 | ResearchInChina says the bottleneck in 2026 shifts from robot bodies toward large-scale, physically realistic multimodal data acquisition and data operations. | 中 | SM002 |
| CM014 | Embodied Global describes a barbell market in which more than half of deals are early stage but nearly 60% of capital flows to the top 20 companies. | 中 | SM003 |
| CM015 | Embodied Global says Beijing, Guangdong, and Shanghai together captured about 80% of total embodied-AI capital in H1 2026. | 中 | SM003 |
| CM016 | Physical Intelligence represents the software-first “general model for any robot” segment of the broader market. | 中 | SM013 |
| CM017 | Wayve represents the autonomous-mobility branch of embodied AI, using the same world-model and end-to-end learning vocabulary for driving rather than warehouse or factory tasks. | 中 | SM014, SM022 |
| CM018 | Unitree represents low-cost legged robotics and industrial inspection, while AGIBOT represents humanoid hardware plus dataset and software infrastructure. | 中 | SM015, SM016 |
| CM019 | Fourier adds healthcare and rehabilitation as an adjacent commercialization path, while X Square and ROBOTERA emphasize foundation-model plus full-stack embodied-AI platforms. | 中 | SM017, SM018, SM019, SM023 |
| CM020 | UBTECH's 2025 annual report shows embodied-intelligent humanoid products becoming its largest revenue source, illustrating that scaled industrial deployment is starting to exist but remains concentrated among a few leaders. | 中 | SM020 |
| CM021 | UBTECH disclosed RMB2.0 billion of 2025 revenue, RMB820.6 million from full-size embodied humanoid products and services, and annualized production capacity above 6,000 units by year end. | 中 | SM020 |
| CM022 | AGIBOT publicly claims a mass-production-line deployment in consumer-electronics manufacturing with Longcheer, supporting 3C manufacturing as a leading commercialization beachhead. | 中 | SM024 |
| CM023 | AGIBOT's “Deployment Year One” language indicates the sector is shifting from demo-led messaging toward measured rollout and productivity claims. | 中 | SM025 |
| CM024 | Manifold AI and UBTECH frame e-commerce logistics as a first commercialization wedge for world-model deployment, making logistics a near-term demand surface rather than a distant adjacency. | 中 | SM009, SM010 |
| CM025 | The buyer in manufacturing and logistics is usually an enterprise automation or operations leader, while the user is line staff, warehouse staff, or robotics engineers and the payer is a capex or operations budget holder. | 中 | SM001, SM006, SM020, SM024 |
| CM026 | The sector's adoption path typically moves from benchmark or lab proof, to pilot, to line-cell or site rollout, and only then to broader multi-site scale. | 中 | SM006, SM020, SM025 |
| CM027 | Waymo One demonstrates that autonomous mobility is commercially real, but it monetizes through an operated service model that is economically and regulatorily distinct from Manifold's model-enablement strategy. | 中 | SM021 |
| CM028 | WorldArena matters commercially because it tries to measure whether embodied world models can serve as data engines, policy evaluators, and action planners rather than just as visually impressive demos. | 中 | SM012 |
| CM029 | The broad trillion-yuan China embodied-intelligence market estimate overstates Manifold AI's near-term serviceable market because Manifold sells enabling models and deployment software rather than the entire robot hardware economy. | 中 | SM001, SM002, SM010 |
| CM030 | ResearchInChina explicitly describes a market transition from customized one-off collection projects toward standardized data services, data stores, and cloud data malls. | 中 | SM002 |
| CM031 | BCG argues that current value capture is strongest at Levels 2 and 3—perception and dexterous manipulation—while Level 5 reasoning remains the gating constraint for general-purpose robotics. | 中 | SM006 |
| CM032 | NVIDIA's Jetson Thor announcements show that edge-compute platforms are becoming a key enabling layer for embodied AI, especially for multimodal, VLA, and world-foundation-model inference. | 中 | SM007, SM008 |
| CM033 | The Jetson Thor partner roster shows that the embodied-AI ecosystem is coalescing around shared compute tooling rather than every vendor building a fully independent stack. | 中 | SM007 |
| CM034 | EU AI Act Annex III shows that some embodied-AI deployments touching critical infrastructure, employment, or public services can fall into high-risk regulatory categories. | 中 | SM026 |
| CM035 | Market estimates conflict because some analysts count the full embodied-intelligence economy, others count robot-data infrastructure, and others still count current shipment or revenue flows. | 中 | SM001, SM002, SM003, SM005 |
| CM036 | BCG's estimate that about 75% of TCO in traditional robotics comes from setup and reengineering helps explain why buyers care about software-defined flexibility, not just robot bodies. | 中 | SM006 |
| CM037 | Public sources do not support a defensible precise SOM for Manifold AI today because customer count, pricing, conversion rates, and realized deployment economics remain undisclosed. | 中 | SM009, SM010, SM011 |
| CM038 | Mixed-reality entertainment appears in Manifold's outward narrative as an adjacency, but current public commercialization proof is much stronger in logistics and 3C manufacturing. | 中 | SM010, SM011 |
| CM039 | X Square's July 2026 PR says the company is already deploying robots across household, industrial, and logistics scenarios, showing that leading peers are racing to secure scenario breadth rather than one vertical only. | 中 | SM023 |
| CM040 | The market still lacks public, segment-level pricing and ROI transparency, even from the most visible Chinese embodied-AI vendors, which keeps buyer willingness-to-pay harder to model from public evidence alone. | 中 | SM020, SM024, SM025 |
| CP001 | The competitive landscape splits into software-first foundation-model labs, full-stack robot-platform vendors, autonomous-mobility model vendors, and traditional automation or internal-build substitutes. | 中 | SP003, SP005, SP007, SP009, SP017 |
| CP002 | Manifold's closest direct peers on the model layer are companies that explicitly combine world models, action models, and proprietary data loops rather than only robot hardware. | 中 | SP002, SP014, SP017 |
| CP003 | Status-quo substitutes include fixed industrial automation and narrower VLA or perception stacks that solve point tasks without a generalized world-model layer. | 中 | SP024, SP025 |
| CP004 | Internal-build competition is real because leading robot vendors and autonomy labs increasingly present themselves as integrated model-plus-hardware developers. | 中 | SP009, SP012, SP014, SP017 |
| CP005 | Physical Intelligence says it is building learning algorithms and models that can control any robot to do any task. | 中 | SP003 |
| CP006 | openpi gives Physical Intelligence an open-source distribution wedge by releasing π0, π0-FAST, π0.5, and fine-tuning packages based on 10k-plus hours of robot data. | 中 | SP004 |
| CP007 | Wayve positions itself as an embodied-AI company for autonomous mobility and publicly reports US$2.8 billion of total funding across four rounds. | 中 | SP005, SP026 |
| CP008 | GAIA-1 shows that Wayve also uses world-model language and generative simulation for action-conditioned prediction in driving, making it technically adjacent even if it targets a different end market. | 中 | SP006 |
| CP009 | Unitree positions itself around low-cost, high-volume quadruped and humanoid robotics with early commercial retail and inspection use cases. | 中 | SP007, SP008 |
| CP010 | AGIBOT positions itself as a one-stop embodied-AI development platform with multiple humanoid product lines and an enterprise-quality task dataset ecosystem. | 中 | SP009 |
| CP011 | AGIBOT's Longcheer announcement provides one of the clearest public proofs of scaled manufacturing deployment among Chinese peers. | 中 | SP010 |
| CP012 | AGIBOT's “Deployment Year One” framing signals a push from technical demos toward utilization and rollout metrics. | 中 | SP011 |
| CP013 | UBTECH is a stronger commercialization benchmark than most startup peers because it reported RMB2.0 billion of 2025 revenue and RMB820.6 million from embodied humanoid products and services. | 中 | SP012 |
| CP014 | UBTECH also disclosed annualized production capacity above 6,000 full-size humanoid robots and extensive industrial application focus, which gives it a trust and scale advantage over younger peers. | 中 | SP012 |
| CP015 | Waymo One demonstrates that operated-service autonomy can achieve public deployment at scale, even though it is not a direct licensor peer to Manifold. | 中 | SP013 |
| CP016 | X Square positions itself as a full-stack embodied-AI company combining foundation models, robotics hardware, a model-driven data pipeline, and real-world deployment. | 中 | SP014, SP015, SP027 |
| CP017 | X Square's July 2026 PR says valuation exceeded US$2.8 billion after four consecutive financing rounds and highlights deployments across household, industrial, and logistics scenarios. | 中 | SP015 |
| CP018 | X Square's GitHub presence adds an open-source distribution channel that increases ecosystem reach and developer familiarity. | 中 | SP016 |
| CP019 | ROBOTERA positions itself as the industry's only full-stack self-developed embodied-intelligence company in its public messaging. | 中 | SP017, SP028 |
| CP020 | ROBOTERA's July 2026 financing release says it raised over US$200 million, had deployments across more than ten logistics centers, and started thousand-unit deliveries in Q2 2026. | 中 | SP018 |
| CP021 | Fourier adds a differentiated healthcare-and-rehabilitation angle to embodied AI while still participating in the broader humanoid robotics race. | 中 | SP019, SP020 |
| CP022 | NVIDIA's Jetson Thor materials show that many embodied-AI companies will share a common edge-compute platform rather than build every compute layer in-house. | 中 | SP021, SP022 |
| CP023 | The same shared-compute trend can reduce raw infrastructure differentiation and intensify competition at the model, data, and deployment layer. | 中 | SP021, SP022 |
| CP024 | Public pricing and packaging transparency is weak across the peer set; most companies disclose narratives, milestones, and funding rather than contract terms or list prices. | 中 | SP003, SP005, SP009, SP012, SP014, SP017 |
| CP025 | Capability comparisons that matter most to buyers are model depth, hardware ownership, deployment proof, data-loop strength, and trust or safety posture. | 中 | SP012, SP024 |
| CP026 | Distribution power is strongest where a vendor combines capital, industrial partners, and scenario ownership, as seen in UBTECH's filings, AGIBOT's manufacturing proof, and ROBOTERA's logistics-center partnerships. | 中 | SP010, SP012, SP018 |
| CP027 | Trust and regulatory posture are strongest for players with public filings or operated services, such as UBTECH and Waymo, and weakest for opaque startups whose claims rely mainly on PR and benchmark narratives. | 中 | SP012, SP013, SP023 |
| CP028 | Switching cost in embodied AI comes from data pipelines, embodiment-specific training, safety validation, workflow integration, and partner-channel dependence rather than from one algorithm alone. | 中 | SP004, SP012, SP024 |
| CP029 | Multi-homing is likelier during the pilot stage, when customers test competing stacks before data collection and workflow tuning create higher lock-in. | 中 | SP024, SP025 |
| CP030 | Strategic investors and ecosystem partners matter because they provide not just capital but deployment channels, supply access, and trust transfer. | 中 | SP002, SP015, SP018 |
| CP031 | WorldArena-style benchmark leadership can give Manifold a discovery advantage with researchers and early adopters, but it is not the same as a durable distribution moat. | 中 | SP001, SP002, SP024 |
| CP032 | The peer set increasingly converges on integrated loops among models, data, and hardware, as shown by Physical Intelligence, Wayve, AGIBOT, X Square, ROBOTERA, and UBTECH. | 中 | SP003, SP005, SP009, SP012, SP014, SP017 |
| CP033 | BCG argues that high-profile robotics demos often exaggerate practical readiness because dexterity and causal reasoning mature more slowly than perception. | 中 | SP024 |
| CP034 | MERICS similarly argues that China's embodied-AI push remains compute-dependent and far from fully autonomous mass deployment, which weakens generalized “winner-take-all” claims across the sector. | 中 | SP025 |
| CP035 | Competitor claims are hardest to verify where public operating metrics are absent—especially pricing, renewal, customer concentration, and deployment economics. | 中 | SP012, SP018, SP020 |
| CP036 | UBTECH's role in standards-setting and national working groups adds a trust moat that early-stage private startups cannot easily replicate. | 中 | SP012 |
| CP037 | Wayve's investor roster and capital base suggest that autonomy competitors can use funding scale itself as a moat by compressing time-to-deployment and attracting ecosystem partners. | 中 | SP005 |
| CP038 | The strongest competitive takeaway for Manifold is that benchmark leadership helps at the top of funnel, but durable moat will depend on pairing model reputation with repeatable deployment channels before better-capitalized full-stack rivals do. | 中 | SP001, SP002, SP012, SP018, SP024 |
| CI001 | No reviewed public source disclosed Manifold AI revenue, ARR, or run-rate as of 2026-08-03. | 中 | SI001, SI002, SI003 |
| CI002 | No reviewed public source disclosed customer count or revenue concentration, despite public deployment claims in logistics and 3C manufacturing. | 中 | SI001, SI002, SI003 |
| CI003 | The official site positions the company around world models for robotics and XR, implying revenue would come from enterprise software, deployment work, and partner-enabled applications rather than consumer subscriptions. | 中 | SI005 |
| CI004 | Public deployment narratives in logistics, manufacturing, automotive, and MR imply a monetization mix of model licensing, integration services, and scenario-specific solution work. | 中 | SI002, SI003 |
| CI005 | WorldArena and benchmark leadership create a plausible evaluation, benchmarking, or data-service wedge, but no pricing or bookings are publicly disclosed for such services. | 中 | SI001, SI005 |
| CI006 | The most plausible GTM motion is founder-led or investor/channel-led enterprise sales into a small number of design partners rather than self-serve product-led adoption. | 中 | SI002, SI003, SI020 |
| CI007 | Public sources do not disclose CAC, payback, pipeline conversion, or sales-cycle duration for Manifold AI. | 中 | SI001, SI002, SI003 |
| CI008 | Major cost buckets likely include model training compute, multimodal data collection and labeling, benchmarking and evaluation infrastructure, field deployment support, and partner integration. | 中 | SI002, SI005, SI018, SI021, SI022 |
| CI009 | Compared with hardware-first peers, Manifold likely carries lower inventory and manufacturing burden but still faces meaningful capex-like spend through compute, data, and field support. | 中 | SI015, SI016, SI017, SI021 |
| CI010 | BCG argues that roughly 75% of traditional robotics TCO sits in setup and reengineering, implying that software-defined automation can monetize by reducing engineering friction rather than only by selling hardware. | 中 | SI016 |
| CI011 | ResearchInChina says the Chinese EAI data industry is shifting from one-off custom collection toward standardized data stores, cloud data malls, and DaaS-style delivery. | 中 | SI018 |
| CI012 | 36Kr ties current demand to labor shortages and flexible manufacturing, supporting an enterprise ROI narrative based on throughput and labor substitution. | 中 | SI019 |
| CI013 | By June 2026, Manifold AI had completed six rounds of financing within roughly one year. | 中 | SI001, SI002 |
| CI014 | June 2026 coverage places cumulative Pre-A financing at nearly RMB1 billion. | 中 | SI001, SI002, SI003 |
| CI015 | Spring 2026 reporting also included a near-RMB200 million Pre-A round before the larger June unicorn-step financing. | 中 | SI004 |
| CI016 | Management said new financing would fund next-generation technology frameworks, multimodal world models, infrastructure, and scenario deployment. | 中 | SI002, SI003 |
| CI017 | No reviewed public source disclosed debt facilities, credit lines, or project-finance obligations for Manifold AI. | 中 | SI001, SI002, SI003 |
| CI018 | The absence of disclosed revenue combined with aggressive funding cadence implies continued dependence on external financing. | 中 | SI001, SI002, SI014 |
| CI019 | A plausible next-round trigger is evidence that benchmark leadership converts into repeatable industrial deployments rather than just additional leaderboard wins. | 中 | SI001, SI002, SI020 |
| CI020 | UBTECH reported RMB2.0 billion of 2025 revenue, RMB820.6 million from embodied humanoid products and services, 37.7% gross margin, and a net loss of RMB789.8 million. | 中 | SI015 |
| CI021 | UBTECH's numbers show that even a scaled embodied-humanoid company can still be capital intensive and loss-making despite meaningful revenue. | 中 | SI015 |
| CI022 | AGIBOT's official store lists the X2 at US$24,240 and the A2 Lite at US$44,560, while the A2 Ultra page emphasizes enterprise deployment, certification, and support terms without a public list price. | 中 | SI006, SI007, SI008 |
| CI023 | AGIBOT store pages also show one-year warranty terms, shipping/import responsibilities, and optional add-on software or package charges, illustrating how adjacent hardware companies bundle support rather than publish clear realized ASPs. | 中 | SI006, SI007, SI008 |
| CI024 | Wayve's investors page says the company has raised US$2.8 billion across four rounds and frames its business as a vehicle-agnostic software platform with recurring software economics. | 中 | SI009 |
| CI025 | Wayve AI Driver product pages emphasize licensing a vehicle-agnostic software stack across levels of autonomy, reinforcing how software-led embodied-AI vendors can target recurring economics without owning the final vehicle. | 中 | SI010 |
| CI026 | Physical Intelligence's pi0 blog illustrates another software-first archetype: powerful model release, strong technical narrative, but no public pricing or revenue disclosure. | 中 | SI011 |
| CI027 | ROBOTERA's products and solutions pages show a full-stack commercial story built around bundled hardware-plus-solution delivery, which makes price discovery less transparent than a pure software product. | 中 | SI012, SI013 |
| CI028 | ChoZan's Fourier analysis suggests another adjacent monetization path: using a pre-existing rehab and care footprint to support future humanoid revenue. | 中 | SI014 |
| CI029 | NVIDIA Jetson Thor materials highlight the high-end edge-compute requirements of embodied AI, reinforcing compute as a nontrivial ongoing cost center or partner dependency. | 中 | SI021, SI022 |
| CI030 | MERICS warns that China's embodied-AI sector remains dependent on Nvidia and still needs significant cost reduction before truly widespread deployment. | 中 | SI017 |
| CI031 | Embodied Global's H1 2026 funding tally suggests capital remains abundant for category leaders, which can help Manifold refinance, but also raises the bar for eventual proof of revenue quality. | 中 | SI020 |
| CI032 | Public traction for Manifold today is technical and capital-market traction—benchmark ranking, investor list, and named scenarios—rather than auditable revenue traction. | 中 | SI001, SI002, SI003 |
| CI033 | That means revenue quality cannot yet be assessed on customer concentration, contract duration, renewal behavior, or gross-margin durability. | 中 | SI001, SI002, SI003 |
| CI034 | A software-first embodied-AI company like Manifold should be capable of structurally higher gross margins than hardware peers if it can keep services and customization from dominating revenue mix. | 中 | SI015, SI016, SI025 |
| CI035 | The main capital-intensity risks are prolonged R&D before monetization, compute and data spending, and the possibility that field deployment requires more service-heavy labor than expected. | 中 | SI016, SI017, SI021 |
| CI036 | Because public price, revenue, and burn data are absent, runway cannot be measured directly from disclosed capital raised. | 中 | SI014, SI017 |
| CI037 | The financial verdict today is that Manifold looks well financed for a private frontier-AI startup, but still financially under-evidenced as an operating business. | 中 | SI014, SI016, SI020 |
| CE001 | Manifold publicly positions itself as a world-model company applying its models to robotics and XR hardware applications. | 中 | SE001 |
| CE002 | The official site publicly names at least three model families or assets: RoboScape for robotics, AirScape for drones, and the broader company platform around WorldScape. | 中 | SE001 |
| CE003 | Pedaily and Tencent coverage add WorldScape Policy as the action model built on top of WorldScape. | 中 | SE021, SE024 |
| CE004 | RoboScape is described in its paper as a unified physics-informed world model that jointly learns RGB video generation and physics knowledge for robotic scenarios. | 中 | SE009 |
| CE005 | The RoboScape paper says the model improves physical plausibility via temporal depth prediction and keypoint dynamics learning. | 中 | SE009 |
| CE006 | WorldArena evaluates embodied world models along both perceptual and functional dimensions, including data engine, policy evaluator, and action planner roles. | 中 | SE003, SE005 |
| CE007 | The benchmark site lists sixteen metrics across six sub-dimensions plus human evaluation, showing that Manifold's preferred evaluation framing extends beyond visual quality alone. | 中 | SE003, SE005 |
| CE008 | Reportify says Manifold's technical roadmap spans DriveScape, RoboScape, and AirScape under a broader “全域世界模型” narrative. | 中 | SE002 |
| CE009 | Reportify also attributes LongScape and a hybrid Auto-regressive + DiT training approach to the company's stack. | 中 | SE002 |
| CE010 | Pedaily describes the company's MoE architecture as separating instruction following, mobility interaction, and manipulation reasoning across expert subspaces. | 中 | SE021 |
| CE011 | WorldScape is publicly described as a real-time world model supporting both mobility and manipulation interactions. | 中 | SE022, SE024, SE027 |
| CE012 | WorldScape Policy is publicly described as the world-action layer that uses predicted spatiotemporal state plus visual input for spatial reasoning and control. | 中 | SE021, SE022 |
| CE013 | Public sources repeatedly tie the stack to logistics, 3C manufacturing, automotive manufacturing, and MR exploration, which means the product is sold as scenario-specific physical-AI capability rather than as a generic API. | 中 | SE021, SE023, SE024 |
| CE014 | The official surface does not publish a detailed product catalog, SDK, API reference, or support manual for external customers. | 中 | SE001 |
| CE015 | The WorldArena GitHub repository is a real developer surface with public code, commit history, and community-facing submission mechanics. | 中 | SE004 |
| CE016 | The WorldScore GitHub repository publishes installation, dataset, and model-registration instructions, which makes benchmark participation reproducible and gives Manifold-aligned evaluation a practitioner foothold. | 中 | SE007 |
| CE017 | The WorldScore Hugging Face leaderboard adds another developer/community surface for public benchmarking visibility. | 中 | SE008 |
| CE018 | The RoboScape GitHub repository provides public code visibility around at least one Manifold-linked model family, although the repo is still small in public social proof. | 中 | SE010, SE032 |
| CE019 | The combination of papers, benchmark repos, and public leaderboards makes Manifold unusually legible for a China robotics startup on research infrastructure, even though the commercial product surface is still sparse. | 中 | SE004, SE007, SE008, SE009, SE010, SE029 |
| CE020 | The public roadmap chronology visible on the official site runs from company launch in May 2025 to RoboScape in May 2025 and AirScape in July 2025, indicating rapid research iteration immediately after formation. | 中 | SE001 |
| CE021 | WorldArena opened submissions in March 2026 and anchored a CVPR 2026 challenge, which shows Manifold's stack is being packaged into a public evaluation workflow rather than only internal demos. | 中 | SE003, SE011, SE028, SE030 |
| CE022 | Tencent and Pedaily report that the company has built a hardware-data-model closed loop, including egocentric, UMI, and hardware-in-the-loop data pipelines with hundreds of thousands of hours of data. | 中 | SE021, SE024 |
| CE023 | That closed-loop data claim is central to the company's moat narrative because it links proprietary capture infrastructure to model improvement. | 中 | SE021, SE024 |
| CE024 | Compared with Physical Intelligence's π0 and openpi narrative, Manifold exposes less generic developer tooling but more public emphasis on benchmark leadership and industrial scenario specificity. | 中 | SE016, SE017, SE029, SE031 |
| CE025 | Compared with Wayve's GAIA-1 and AI Driver, Manifold appears earlier in productization but more explicit about embodied manipulation and closed-loop physical interaction. | 中 | SE014, SE015, SE021 |
| CE026 | Public sources claim that WorldScape can be quantized and distilled for edge inference to drive robot mobility and drone navigation. | 中 | SE002 |
| CE027 | NVIDIA's Jetson Thor materials underscore that advanced embodied-AI products depend on increasingly capable edge compute, making compute availability and porting efficiency real product dependencies. | 中 | SE012, SE013 |
| CE028 | The UBTECH partnership shows the product is being positioned to sit on top of third-party robot platforms rather than only on Manifold-owned hardware. | 中 | SE022, SE026 |
| CE029 | That partner-led deployment model can accelerate commercialization but also means reliability and support are partly contingent on external robot manufacturers and integrators. | 中 | SE022, SE018, SE019 |
| CE030 | AGIBOT's certification-heavy A2 Ultra page highlights how adjacent robotics products publicize hardware certifications and warranty terms, while Manifold does not yet publish equivalent trust artifacts on its own surface. | 中 | SE020, SE001 |
| CE031 | No reviewed Manifold public source disclosed formal safety certifications, security attestations, incident history, privacy controls, or compliance documentation for the product stack. | 中 | SE001, SE021 |
| CE032 | Because the product influences physical action in robotics contexts, the lack of public trust and safety documentation is more material than it would be for a pure video-model company. | 中 | SE005, SE009, SE022 |
| CE033 | Most visible public proof is benchmark and paper proof, not customer-operated uptime, MTBF, or deployment-support evidence. | 中 | SE003, SE005, SE021, SE027 |
| CE034 | That means the most mature public assets today are the research stack and benchmarking apparatus, while the least proven assets are repeatable deployment operations, support tooling, and trust controls. | 中 | SE003, SE021, SE022 |
| CE035 | The product-tech verdict is that Manifold looks technically differentiated and unusually benchmark-native, but still under-documented as an enterprise product platform. | 中 | SE003, SE021, SE022, SE026 |
| CU001 | Public Manifold sources position the company around enterprise physical-AI scenarios rather than consumer distribution. | 中 | SU001, SU002 |
| CU002 | The clearest public end-use segments are e-commerce logistics, 3C manufacturing, automotive manufacturing, and MR exploration. | 中 | SU002, SU003, SU004 |
| CU003 | In workflow terms, likely buyers are operations, automation, or manufacturing leaders; likely users are robotics / automation teams; likely payers are enterprise innovation or capex-backed industrial budgets. | 中 | SU002, SU004, SU009 |
| CU004 | No reviewed public source disclosed total customer count, active accounts, deployed sites, or utilization. | 中 | SU002, SU003, SU004 |
| CU005 | No reviewed public source disclosed geography-by-customer mix or revenue concentration by account. | 中 | SU002, SU003 |
| CU006 | The strongest named customer/channel proof in the public record is UBTECH, not a named end logistics operator or 3C factory. | 中 | SU005, SU006 |
| CU007 | Gasgoo and Phoenix Auto say the UBTECH partnership is intended to combine Manifold's world models with UBTECH's mass-production capability to deliver profitable comprehensive solutions, starting with e-commerce and logistics. | 中 | SU005, SU006 |
| CU008 | UBTECH's own logistics and industrial-solution pages show that warehousing, parcel handling, sorting, inspection, and assembly are already targetable customer workflows on the partner side. | 中 | SU012, SU013, SU014 |
| CU009 | That makes UBTECH a credible distribution or embodiment channel for Manifold, but it does not prove how many end customers Manifold has closed directly. | 中 | SU005, SU012 |
| CU010 | Pedaily, Tencent, TMTPost, and KuCoin all repeat that Manifold technology is already applied in e-commerce logistics. | 中 | SU002, SU003, SU004, SU010 |
| CU011 | The same mid-2026 source cluster also repeats 3C manufacturing as a live or actively deployed scenario. | 中 | SU002, SU003, SU008 |
| CU012 | Automotive manufacturing appears in public source lists, but the evidence is weaker and more forward-looking than for logistics or 3C manufacturing. | 中 | SU002, SU003, SU007 |
| CU013 | Reportify and Leaderobot frame the company as platform-like infrastructure spanning hardware products, data tools, models, and custom solutions, implying buyer diversity across industrial accounts and robotics partners. | 中 | SU007, SU009 |
| CU014 | WorldArena has more than 200 challenge submissions and visible community participation from large tech companies and research teams, but that is ecosystem adoption rather than direct customer revenue proof. | 中 | SU018, SU020, SU021 |
| CU015 | Public customer evidence is therefore strongest on discovery and strategic interest, weaker on account-level conversion, and weakest on renewal or retention. | 中 | SU004, SU014, SU018 |
| CU016 | No reviewed public source disclosed production-vs-pilot counts, multi-site rollout counts, or robot-hours delivered under customer contracts. | 中 | SU002, SU003, SU005 |
| CU017 | No reviewed public source disclosed NRR, GRR, churn, contract length, or renewal rates. | 中 | SU002, SU003, SU005 |
| CU018 | No reviewed public source disclosed customer-quoted KPI outcomes such as error reduction, throughput improvement, or labor savings from Manifold deployments. | 中 | SU002, SU003, SU005 |
| CU019 | The absence of named end-customer references means production deployment and pilot language cannot be cleanly separated from public evidence alone. | 中 | SU002, SU003, SU005 |
| CU020 | A plausible expansion motion is benchmark visibility -> technical evaluation -> pilot deployment -> partner-led scaling across robotics platforms or adjacent factory workflows. | 中 | SU018, SU020, SU005, SU012 |
| CU021 | Strategic investors such as BAIC Capital and industrial partners like UBTECH likely do more than fund the company; they likely open domain access, reference opportunities, or deployment pathways. | 中 | SU002, SU005, SU007 |
| CU022 | The public record points to high concentration risk because only a tiny set of counterparties—especially UBTECH and investor-linked industrial ecosystems—are visible as channel or strategic proof. | 中 | SU005, SU006, SU007 |
| CU023 | Customer and channel concentration may be amplified by the fact that public proof is sector-specific and hardware-partner-mediated rather than broad-based self-serve usage. | 中 | SU005, SU012, SU025 |
| CU024 | Procurement friction is likely high because embodied-AI buyers must align robot hardware, world-model software, site integration, and safety acceptance at once. | 中 | SU012, SU014, SU024 |
| CU025 | Compared with Waymo and Wayve, Manifold has much thinner public customer proof: fewer named production customers, fewer deployment metrics, and less geography-level visibility. | 中 | SU022, SU023, SU024 |
| CU026 | Compared with competitor solution pages like ROBOTERA, Manifold also discloses less about concrete customer scenarios and packaged solution boundaries on its own official surface. | 中 | SU001, SU025 |
| CU027 | UBTECH order disclosures—13,361 channel orders on one Tencent report and 11,000+ preorders on Sohu—show channel demand for humanoid products, but they do not reveal how much of that flow can be attributed to Manifold-enabled solutions. | 中 | SU015, SU017 |
| CU028 | A second Tencent article frames UBTECH’s order and commercialization push as meaningful but still financially stressed, which matters because Manifold’s most visible customer/channel proof sits on top of that partner. | 中 | SU016, SU005 |
| CU029 | Public continuity across June and July 2026 sources supports the idea that logistics and 3C messaging is persistent rather than a one-day claim burst. | 中 | SU002, SU003, SU007, SU008 |
| CU030 | However, continuity of sector messaging is not the same as continuity of paying customer relationships. | 中 | SU002, SU003, SU005 |
| CU031 | WorldArena community participation is better interpreted as top-of-funnel technical credibility that can attract customers, not as a substitute for named accounts. | 中 | SU018, SU019, SU021 |
| CU032 | Because the company is still early, its best near-term customer motion likely runs through lighthouse projects, OEM partners, and strategic industrial backers rather than a diversified direct-sales base. | 中 | SU005, SU007, SU021 |
| CU033 | MR exploration is part of the public application set, but it should not be treated as customer proof because no public deployment detail accompanies it. | 中 | SU002, SU003 |
| CU034 | Public evidence quality is highest for partner/channel existence, medium for vertical deployment claims, and low for end-customer outcomes or retention. | 中 | SU005, SU006, SU018 |
| CU035 | The customer-proof verdict is that Manifold clearly has industrial demand signals and at least one meaningful named commercialization channel, but still lacks the account-level transparency needed to judge durability. | 中 | SU005, SU006, SU015, SU018 |
| CR001 | The most severe risk is not lack of technical ambition, but the gap between benchmark leadership and publicly evidenced production reliability. | 中 | SR016, SR017, SR001 |
| CR002 | China released a national standard system for humanoid robotics and embodied AI in 2026, covering applications, safety, ethics, and lifecycle standards. | 中 | SR009, SR010 |
| CR003 | That means embodied-AI vendors like Manifold are entering a policy environment that is becoming more formalized, not less. | 中 | SR009, SR010 |
| CR004 | Hangzhou’s 2026 embodied-robotics regulation shows local governments are beginning to define testing and commercialization frameworks for the category. | 中 | SR011 |
| CR005 | China AI ethics and safety guidance adds a second layer of obligation around fairness, human control, and risk management for embodied AI applications. | 中 | SR012 |
| CR006 | The EU AI Act creates possible future compliance burden if Manifold or its partners want to commercialize relevant autonomy or safety-adjacent systems in Europe. | 中 | SR015, SR031, SR025 |
| CR007 | The UK Automated Vehicles Act highlights a broader regulatory trend toward assigning responsibility, authorization, and safety accountability in autonomy systems. | 中 | SR013, SR014 |
| CR008 | No reviewed public Manifold source disclosed formal safety certifications, incident-response procedures, or model-governance documentation specific to the company. | 中 | SR001, SR002 |
| CR009 | No reviewed public Manifold source disclosed privacy architecture, data-governance controls, or customer data-rights terms for its collection loop. | 中 | SR001, SR002 |
| CR010 | Because Manifold’s stack can influence robot action in industrial settings, the lack of public trust artifacts is a more material risk than for a pure content model. | 中 | SR017, SR029, SR023 |
| CR011 | No reviewed public source disclosed uptime, MTBF, intervention rate, task success rate, or incident history for Manifold deployments. | 中 | SR002, SR003, SR004 |
| CR012 | That makes operational reliability a core unknown, not a secondary diligence item. | 中 | SR011, SR021, SR023 |
| CR013 | WorldArena and WorldScore provide strong evidence of benchmark performance, but they also create a classic risk of optimizing for public evaluation instead of customer KPI reality. | 中 | SR016, SR017, SR028 |
| CR014 | The product appears to depend on partner robot hardware and site integration, which means field quality can fail even if model quality is strong. | 中 | SR004, SR022, SR023 |
| CR015 | NVIDIA’s Jetson Thor materials and MERICS’s analysis both point to compute availability and platform dependence as category-level risks for embodied AI. | 中 | SR006, SR019, SR020 |
| CR016 | MERICS explicitly warns that China’s embodied-AI sector remains dependent on Nvidia and needs significant cost reduction before widespread deployment. | 中 | SR006 |
| CR017 | UBTECH is currently Manifold’s clearest public commercialization channel, creating visible dependency on partner health, priorities, and execution. | 中 | SR004, SR024 |
| CR018 | UBTECH’s public order momentum is encouraging, but its stressed profitability narrative means Manifold inherits some channel fragility through that relationship. | 中 | SR005, SR024 |
| CR019 | Manifold’s data-collection moat also implies data-rights and site-access dependence; if customers limit data capture, model improvement could slow. | 中 | SR002, SR008, SR027 |
| CR020 | The company has completed six rounds in roughly one year and still does not disclose revenue or runway, implying meaningful financing dependence. | 中 | SR002, SR003 |
| CR021 | Current funding mitigates near-term survival risk, but it can amplify pressure to prove commercial scale quickly at a still-immature stage. | 中 | SR003, SR007, SR024 |
| CR022 | Public customer evidence is concentrated in a small set of sectors and counterparties, which creates concentration risk even before revenue concentration is disclosed. | 中 | SR004, SR022, SR023 |
| CR023 | Hardware-led peers like UBTECH show that scaling embodied AI can remain margin-compressive and loss-making even with substantial revenue. | 中 | SR005 |
| CR024 | BCG’s TCO analysis implies deployments can fail economically if setup, reengineering, and integration stay too labor intensive. | 中 | SR007 |
| CR025 | No public litigation, enforcement action, or IP dispute involving Manifold was identified in the reviewed source set. | 中 | SR001, SR002, SR003 |
| CR026 | That absence lowers current visible legal risk, but it does not reduce diligence need around IP provenance and training-data rights. | 中 | SR008, SR009 |
| CR027 | Founder and research-talent concentration is a real people risk because the public narrative leans heavily on Wu Wei, Tsinghua FIB lineage, and specialist world-model expertise. | 中 | SR003, SR027 |
| CR028 | A second people risk is organizational: translating frontier research into industrial support, QA, and customer operations often requires different leadership muscle than benchmark leadership. | 中 | SR007, SR024, SR027 |
| CR029 | The company’s strongest mitigant is that it has real capital, a coherent technical wedge, and at least one visible commercialization partner. | 中 | SR003, SR004 |
| CR030 | The most monitorable regulatory failure would be any requirement for safety, auditability, or data-handling standards that Manifold cannot quickly document. | 中 | SR009, SR011, SR015 |
| CR031 | The most monitorable channel failure would be loss, delay, or non-expansion of the UBTECH route to market. | 中 | SR004, SR024 |
| CR032 | The most monitorable customer-proof failure would be continued absence of named production deployments despite further funding rounds. | 中 | SR002, SR003, SR004 |
| CR033 | A thesis-break trigger would be evidence that benchmark wins fail to convert into repeatable paid industrial deployments by the next financing event. | 中 | SR016, SR017, SR020 |
| CR034 | Another thesis-break trigger would be any safety or reliability incident in a live industrial deployment that reveals weak controls or brittle action planning. | 中 | SR021, SR023, SR029 |
| CR035 | A third thesis-break trigger would be material regulatory burden on data capture or model accountability that slows deployment economics. | 中 | SR009, SR011, SR031 |
| CR036 | Near-term mitigations management can actually control include publishing trust documentation, tightening partner qualification, and disclosing deployment quality metrics. | 中 | SR001, SR004, SR021 |
| CR037 | Near-term mitigations management cannot fully control include embodied-AI macro hype cycles, global compute supply, and counterparties’ financial health. | 中 | SR006, SR019, SR024 |
| CR038 | The public record suggests risk is skewed toward execution, trust, and concentration—not toward a lack of technical relevance. | 中 | SR001, SR016, SR029 |
| CR039 | In investment terms, Manifold is a high-upside but high-residual-risk company whose best public proof still sits earlier in the commercialization curve than its valuation narrative implies. | 中 | SR003, SR005, SR024 |
| CR040 | Overall, the risk verdict is “elevated but monitorable”: the company is not broken, but it remains exposed to exactly the category failures that often emerge between frontier-model demos and industrial scale. | 中 | SR007, SR016, SR024 |
| CV001 | The pro-thesis is that Manifold sits at a strategically attractive intersection of world models, embodied AI, and China industrial deployment. | 中 | SV001, SV022, SV024 |
| CV002 | The strongest evidence for that thesis is benchmark leadership plus fast capital formation, not disclosed operating metrics. | 中 | SV001, SV003, SV024, SV025 |
| CV003 | The anti-thesis is that Manifold may already be valued like a category winner before public evidence proves repeatable commercial deployment. | 中 | SV001, SV004, SV020 |
| CV004 | Current public evidence supports a recommendation of track / research-more rather than an affirmative “buy at any price.” | 中 | SV001, SV004, SV021 |
| CV005 | Confidence should be medium because product and market evidence are strong, but financial and customer proof remain thin. | 中 | SV005, SV021, SV025 |
| CV006 | Risk rating should remain high because concentration, trust, and financing dependence are all still material. | 中 | SV006, SV021, SV004 |
| CV007 | Valuation stance should be price-sensitive: the current unicorn narrative is understandable, but public evidence does not support paying a clear premium above that floor. | 中 | SV001, SV003, SV006 |
| CV008 | The best public valuation anchor for Manifold itself is still narrative rather than arithmetic: near-RMB1B Pre-A financing and confirmed unicorn framing by June 2026. | 中 | SV001, SV002, SV003 |
| CV009 | No reviewed public source disclosed cap-table detail, liquidation preferences, or exact dilution from the six-round funding cadence. | 中 | SV001, SV002 |
| CV010 | That missing preference-stack information is one reason a new investor should demand entry discipline even if the company quality is real. | 中 | SV001, SV009 |
| CV011 | Physical Intelligence is a relevant software-first comparable because it also sells the promise of a generalist robot intelligence layer rather than vertically integrated robot manufacturing. | 中 | SV011, SV031 |
| CV012 | Physical Intelligence’s late-2025 valuation of roughly $5.6B shows how large a premium public and private markets may pay for robot-model platforms with strong investor syndicates. | 中 | SV011, SV012, SV031 |
| CV013 | UBTECH is a relevant hardware-heavy comp because it supplies real public revenue and margin data for scaled humanoid commercialization, albeit with large losses. | 中 | SV004, SV020 |
| CV014 | AgiBot is relevant as a high-velocity Chinese embodied-AI private comp because it pairs deployment claims with IPO valuation targets in the $5B-$6B+ range. | 中 | SV013, SV015, SV019 |
| CV015 | Unitree is relevant because its IPO process pushes an embodied-robotics comp toward public price discovery at roughly a $6B target valuation. | 中 | SV017, SV018 |
| CV016 | Wayve is relevant as a software-led autonomy platform comp with recurring-software aspirations and total funding of $2.8B, though its automotive path differs from factory robotics. | 中 | SV009, SV010 |
| CV017 | The right comp set is therefore mixed: software-model platforms for margin ambition, and hardware-heavy robotics companies for deployment realism. | 中 | SV011, SV013, SV015, SV016 |
| CV018 | Manifold’s public customer and revenue opacity means any comparable should be discounted for earlier commercial proof and lower disclosure quality. | 中 | SV004, SV021, SV025 |
| CV019 | The main downside from missing revenue disclosure is not just modeling difficulty; it is the inability to tell whether the company behaves like software, services, or expensive pilot work. | 中 | SV004, SV005, SV021 |
| CV020 | The main downside from channel concentration is that the clearest public commercialization path runs through a small number of partners rather than a disclosed broad customer base. | 中 | SV021, SV027, SV030 |
| CV021 | The upside case rests on benchmark leadership converting into the default evaluation and control layer for embodied deployment in China. | 中 | SV024, SV025 |
| CV022 | A second upside driver is partner-led commercialization through OEMs or industrial platforms such as UBTECH. | 中 | SV021, SV027, SV030 |
| CV023 | The bull case requires three things: named production deployments, evidence of repeatability beyond one channel, and a convincing software-like margin path. | 中 | SV004, SV011, SV021 |
| CV024 | The base case assumes the current unicorn mark broadly holds because strategic value remains real, but the company still lacks enough proof to rerate sharply upward. | 中 | SV001, SV003, SV005 |
| CV025 | The bear case assumes that commercialization lags the valuation narrative, forcing either a flat round or a down-round once markets demand customer evidence. | 中 | SV004, SV006, SV017 |
| CV026 | An upward rerating would require public or diligenced proof of named deployments, better revenue quality, or repeat channel expansion. | 中 | SV021, SV024 |
| CV027 | Down-round risk would rise if the next financing arrives before the company can show named production deployments or basic operating metrics. | 中 | SV001, SV021 |
| CV028 | The broader embodied-AI funding cycle remains supportive, as Embodied Global shows in China, but supportive capital markets should not be confused with proof of fair price. | 中 | SV007, SV006 |
| CV029 | A reasonable public-evidence discount is to haircut late-stage comp valuations for Manifold’s earlier operating disclosure and customer proof. | 中 | SV012, SV015, SV017, SV020 |
| CV030 | That haircut still leaves room for a unicorn-like strategic option value, but not for treating Manifold like a de-risked leader on par with larger comps. | 中 | SV001, SV012, SV017 |
| CV031 | Entry discipline should target either a valuation discount to the current narrative, unusually strong rights, or milestone-based evidence before committing. | 中 | SV009, SV018, SV029 |
| CV032 | Plausible exit pathways include a larger strategic round, a later IPO if China embodied-AI public markets deepen, or acquisition by an industrial or platform player seeking world-model capability. | 中 | SV015, SV017, SV021 |
| CV033 | Those exit pathways are not yet ready enough to justify a late-stage-style underwriting approach. | 中 | SV004, SV017 |
| CV034 | The most important final diligence asks are revenue, deployment count, concentration, preference stack, and governance / safety controls. | 中 | SV001, SV004, SV021 |
| CV035 | The top thesis-break trigger is continued absence of named production deployments by the next financing event. | 中 | SV001, SV021 |
| CV036 | A second thesis-break trigger is evidence that partner-led commercialization does not scale beyond announcements and pilot narratives. | 中 | SV021, SV027 |
| CV037 | A third thesis-break trigger is regulatory or trust friction that slows deployment just as more demanding investors seek proof. | 中 | SV006, SV025 |
| CV038 | The evidence-quality discount should remain meaningful because Manifold has more public benchmark proof than revenue, retention, or margin proof. | 中 | SV024, SV025, SV021 |
| CV039 | The current mark most resembles a software-platform option with frontier-AI premium attached, not a transparently underwritten industrial robotics business. | 中 | SV009, SV011, SV022 |
| CV040 | The overall valuation verdict is that Manifold is strategically interesting enough to track closely, but not transparently proven enough to chase at an undisciplined unicorn entry. | 中 | SV001, SV018, SV021 |