Maxinsights 向人们付费，让他们在烹饪、打扫和工作时佩戴摄像头。这家位于圣克拉拉的初创公司现已完成 3000 万美元种子轮融资，旨在将这些影像转化为机器人训练数据。Peak XV 领投该轮融资，公司于周四宣布。Tola Capital、South Park Commons、INCE Capital 和 Alumni Ventures 也参与了投资。
Maxinsights pays people to wear cameras while they cook, clean and work. The Santa Clara startup has now raised $30M in seed funding to turn that footage into training data for robots. Peak XV led the round, the company announced on Thursday. Tola Capital, South Park Commons, INCE Capital and Alumni Ventures also took part.
公司在融资的同时推出了两款产品。MaxLoop 是一款在影像传入时即进行检查的软件。MaxGlove 是一款传感器手套，可记录手部抓握的力度。创始人兼首席执行官袁宇光向《福布斯》的 John Koetsier 解释了这一押注。
It launched two products alongside the round. MaxLoop is software that checks footage as it arrives. MaxGlove is a sensor glove that records how hard a hand grips. Founder and chief executive Yuguang Yuan explained the bet to John Koetsier at Forbes.
“语言模型通过阅读互联网学会了写作，而人们花了数十年时间用文本填充互联网。机器人没有类似的东西。没有人从第一人称视角、双手入镜地装载洗碗机的影像档案，”袁宇光对《福布斯》表示。
“Language models learned to write by reading the internet, which people spent decades filling with text. Robots have nothing like that. There’s no archive of people loading dishwashers from a first-person view with their hands in frame,” Yuan told Forbes.
5000 名承包商佩戴摄像头 来自六个国家的 5000 多名承包商佩戴公司的头戴式、腕戴式和胸戴式摄像头。他们拍摄自己装载洗碗机、制作冰沙或洗车的过程。当地人力资源公司雇佣他们，Maxinsights 直接管理他们。
Cameras on 5,000 contractors More than 5,000 contractors in six countries wear the company’s head, wrist and chest cameras. They film themselves loading dishwashers, making smoothies or washing cars. Local staffing firms hire them, and Maxinsights manages them directly.
公司表示，它在家庭、餐厅和酒店记录真实工作，速度自然。公司称，竞争对手在模拟环境中安排任务，并要求操作员放慢手部动作。
The company says it records real work at natural speed in homes, restaurants and hotels. Rivals stage tasks in mock settings and ask operators to slow their hands down, it says.
Maxinsights 表示，已采集超过 200 万小时影像。其中，150 万小时已可供模型使用。公司拥有 30 多家客户，包括前沿 AI 实验室，其中 8 家签有经常性合同。
Maxinsights says it has captured more than 2 million hours of footage. Of those, 1.5 million hours are ready for models. It has more than 30 customers, including frontier AI labs, and eight are on recurring contracts.
Genesis AI 和 Dyna 背后的数据 这家初创公司称，它提供了 Genesis AI 最新模型 GENE-26.5 背后所有 20 万小时的第一人称数据。Genesis 于 7 月曾洽谈融资约 5 亿美元，用于其机器人软件。Maxinsights 还自称是 Dyna Robotics 模型 Dyna-2 最大的数据提供方。
The data behind Genesis AI and Dyna The startup says it supplied all 200,000 hours of first-person data behind GENE-26.5, the latest model from Genesis AI. Genesis was in talks to raise about $500m for its robot software in July. Maxinsights also calls itself the largest data provider behind Dyna-2, a model from Dyna Robotics.
“机器人需要从日常生活的复杂性和多样性中学习，仅靠精心控制的数据集无法捕捉这一点，”Genesis AI联合创始人王森在一份声明中表示。
“Robots need to learn from the complexity and variation of everyday life, and carefully controlled datasets alone cannot capture that,” Genesis AI co-founder Johnson Wang said in a statement.
两家公司正在共同开发MaxGlove。它能记录握力、物体重量以及任务过程中接触点的变化，并与视频同步。Maxinsights表示，市面上已有的手套需要长时间校准，每天仅能提供10到20分钟可用的采集时长。该公司计划在2027年收集100万小时的力数据。
The two companies are building MaxGlove together. It records grip force, object weight and how contact shifts during a task, in sync with the video. Gloves already on sale need long calibration, Maxinsights says. They give only 10 to 20 minutes of usable capture a day. The company wants 1 million hours of force data in 2027.
MaxLoop会对每段视频片段在其入库时进行评分。它会检查操作人员是否在工作、摄像头位置是否正确、任务是否完成。操作人员现在当天就能收到反馈，而不再需要等待一到两周。该公司称审核成本下降了约80%。其声明表示，人工审核团队缩减了90%。该系统每月处理约45万小时的视频素材。
MaxLoop scores each clip as it lands. It checks that the operator is working, that the cameras sit properly and that the task gets finished. Operators now hear back the same day, not one to two weeks later. The company puts the drop in review costs at roughly 80%. Its release says the human review team shrank by 90%. It processes about 450,000 hours of footage a month.
机器人需要多少素材？据Crunchbase报道，投资者在2026年上半年向实体AI投入了474亿美元，几乎是2025年下半年总额的近四倍。训练这些机器所需的数据更为稀缺。
How much footage robots need Investors put $47.4bn into physical AI in the first half of 2026, according to Crunchbase. That is nearly four times the total for the second half of 2025. The data to train those machines is scarcer.
其他公司也在追逐同一空白领域。NEURA Robotics正在建设机器人训练馆。Encord于2月为其数据平台筹集了5000万欧元，Enigma则筹集了7000万美元的种子轮融资，让网上的人员远程操控其机器人。Figure通过名为Index的项目付费让人们拍摄家务视频。袁某告诉《福布斯》，Index服务于Figure自身的模型，而Maxinsights则向众多公司销售。
Others are chasing the same gap. NEURA Robotics is building gyms to train robots. Encord raised €50M in February for its own data platform, and Enigma raised a $70M seed round to let people online control its robots. Figure pays people to film their chores through a programme called Index. Yuan told Forbes that Index serves Figure’s own models, while Maxinsights sells to many companies.
他还对《福布斯》表示，能够像人类一样适应环境的机器人可能需要一千万小时或更多的数据。并非所有人都认同这一规模。至少有一家初创公司押注机器人不需要那么多真实世界数据，并据此筹集了资金，《福布斯》指出。袁某表示，客户告诉他，需求正在急剧上升。
He also told Forbes that robots able to adapt as people do may need ten million hours of data or more. Not everyone agrees on that scale. At least one startup has raised money on the bet that robots will not need that much real-world data, Forbes noted. Yuan said customers are telling him demand is rising sharply.
“几乎该类别中的每一家公司都在出售工时。Maxinsights是唯一一家将质量控制视为工程问题而非人员成本的公司，”Peak XV管理合伙人罗希特·阿加瓦尔说。
“Almost every company in this category is selling hours. Maxinsights is the only one that treated quality control as an engineering problem rather than a staffing cost,” said Rohit Agarwal, managing director of Peak XV.
袁此前曾联合创立Mega Gear Robotics，该公司已被另一家公司收购。首席技术官王昆是Google X和Waymo的早期工程师，随后联合创立了自动驾驶公司QCraft。Maxinsights在中国和印度尼西亚设有办事处，并计划在土耳其和墨西哥开设办事处。袁在接受《福布斯》采访时表示，大约两年内，他预计客户的模型将反馈到公司收集和测试的内容中。
Yuan previously co-founded Mega Gear Robotics, which another company bought. Chief technology officer Kun Wang was an early engineer at Google X and Waymo. He then co-founded QCraft, a self-driving company. Maxinsights has offices in China and Indonesia and plans to open in Turkey and Mexico. Within about two years, Yuan told Forbes, he expects customers’ models to feed back into what the company collects and tests.