This startup is betting India’s gig economy can train the world’s robots
This startup is betting India’s gig economy can train the world’s robots
这家初创公司押注印度的零工经济可以训练全球的机器人
In the last few years, India’s online food delivery market has grown significantly, with both Zomato and Swiggy going public and the number of cloud kitchens increasing. Meanwhile, startups working on home services, such as on-demand household staffing platforms like Urban Company, Snabbit, and Pronto, have gained popularity. 过去几年里，印度的在线食品配送市场增长显著，Zomato 和 Swiggy 均已上市，云厨房的数量也在不断增加。与此同时，从事家庭服务的初创公司，如按需家政服务平台 Urban Company、Snabbit 和 Pronto，也获得了广泛欢迎。
Silicon Valley-based startup Human Archive is tapping into this trend, partnering with these companies to have workers wear special caps with cameras to collect egocentric (first-person point of view) video data of everyday tasks that could be used to train robots. Without naming specific partners, the startup said it is working with companies in the home services, hotel, and restaurant sectors to collect egocentric data, and it says it has more than 1,000 active headsets deployed across multiple locations. 总部位于硅谷的初创公司 Human Archive 正利用这一趋势，与这些公司合作，让工人佩戴装有摄像头的特制帽子，以收集日常任务的“自我中心”（第一人称视角）视频数据，这些数据可用于训练机器人。在未透露具体合作伙伴的情况下，该公司表示正与家庭服务、酒店和餐饮行业的公司合作收集此类数据，并称目前已有超过 1,000 台活跃的头戴设备部署在多个地点。
On the back of that traction, Human Archive said Tuesday it has raised $8.2 million in funding from Wing Venture Capital, NVP Capital, Y Combinator, and angels from OpenAI, Nvidia, Google, Mercor, AfterQuery, BAIR, SAIL, Brad Boa, and Meta. The startup was founded by three students from UC Berkeley and one from Stanford — Samay Maini, Rushil Agarwal, Shloke Patel, and Raj Patel, the latter two being cousins. (Raj Patel is CEO.) All four have research backgrounds spanning robotics, hardware, and tactile data. 凭借这一进展，Human Archive 周二宣布筹集了 820 万美元资金，投资方包括 Wing Venture Capital、NVP Capital、Y Combinator，以及来自 OpenAI、Nvidia、Google、Mercor、AfterQuery、BAIR、SAIL、Brad Boa 和 Meta 的天使投资人。该初创公司由三名加州大学伯克利分校的学生和一名斯坦福大学的学生共同创立——他们分别是 Samay Maini、Rushil Agarwal、Shloke Patel 和 Raj Patel（后两人为堂兄弟，Raj Patel 担任首席执行官）。这四人均拥有机器人、硬件和触觉数据方面的研究背景。
The company’s founding is a direct bet on where the AI industry is heading. As robotics labs and frontier AI companies race to build machines that can perform physical tasks in the real world, they face a critical bottleneck — a shortage of high-quality, real-world training data showing humans doing everyday work. Human Archive’s bet is that the workers staffing India’s booming gig economy represent an untapped and scalable source of exactly that data. 该公司的成立是对人工智能行业未来发展方向的一次直接押注。随着机器人实验室和前沿 AI 公司竞相开发能够在现实世界中执行物理任务的机器，它们面临着一个关键瓶颈——缺乏展示人类进行日常工作的高质量、真实世界训练数据。Human Archive 的赌注在于，印度蓬勃发展的零工经济中的从业者，正是这种数据未被开发且可扩展的来源。
While Human Archive is working with multiple partners, the startup said it was rejected by many Indian home services companies, including Pronto and Urban Company, for a collaboration. The company’s rejection by major players became public fodder last weekend, when Indian outlet Entrackr reported that Pronto is actively seeking partnerships to collect worker data for robotics training and that Snabbit had held early discussions with Human Archive before the project fell apart. 尽管 Human Archive 正在与多个合作伙伴开展工作，但该公司表示，它曾被许多印度家庭服务公司（包括 Pronto 和 Urban Company）拒绝合作。上周末，当印度媒体 Entrackr 报道称 Pronto 正在积极寻求合作伙伴以收集用于机器人训练的工人数据，且 Snabbit 在项目破裂前曾与 Human Archive 进行过初步讨论时，该公司被主要参与者拒绝的消息成为了公众谈资。
Urban Company CEO Abhiraj Singh Bhal responded on X, stating the company would not engage in such arrangements — prompting Patel to fire back that Urban Company would soon be forced to reconsider or risk losing relevance to customer churn. Co-founder Rushil Agarwal was blunter still, posting that Pronto founder Anjali Sardana had laughed at him and called him “stupid” when he raised the idea of a data partnership. Pronto acknowledged the conversations but said it chose not to move forward. Urban Company 首席执行官 Abhiraj Singh Bhal 在 X 上回应称，该公司不会参与此类安排——这促使 Patel 反击称，Urban Company 很快将被迫重新考虑，否则将面临因客户流失而失去市场地位的风险。联合创始人 Rushil Agarwal 则更为直率，他发帖称，当他提出数据合作的想法时，Pronto 创始人 Anjali Sardana 嘲笑了他并称他“愚蠢”。Pronto 承认了这些对话，但表示选择不继续推进。
Across the country, other startups are collecting egocentric data from different work environments, including factory floors. To differentiate itself, Human Archive is using and developing additional devices, such as tactile gloves, a full-body motion capture suit, and wrist cameras to capture data, including motion and tactile force, synchronously aligned with RGB-D (color imagery paired in real time with depth information), to sell to AI labs. The startup believes that video data alone is not sufficient but that pairing it with other sensor data makes it much more valuable. 在印度全国范围内，其他初创公司也在从包括工厂车间在内的不同工作环境中收集自我中心数据。为了实现差异化，Human Archive 正在使用和开发额外的设备，例如触觉手套、全身动作捕捉服和腕部摄像头，以捕捉包括运动和触觉力在内的数据，并将其与 RGB-D（实时配对深度信息的彩色图像）同步对齐，从而出售给 AI 实验室。该初创公司认为，仅有视频数据是不够的，将其与其他传感器数据配对会使其价值大增。
Initially, Human Archive used makeshift setups or off-off-the-shelf rigs to capture the data. Now it is working on custom hardware that works together and captures different kinds of data. It already has more than 50 different devices deployed to collect different data points. “To capture data, we started with iPhones; then we built our own custom rigs and caps. Now we have more than seven different hardware products that we use interchangeably across different modalities. After data collection from different devices, we worked on synchronizing data from all these different sources,” Patel said in a call. 最初，Human Archive 使用临时搭建的装置或现成的设备来捕捉数据。现在，它正在开发能够协同工作并捕捉不同类型数据的定制硬件。它已经部署了超过 50 种不同的设备来收集不同的数据点。“为了捕捉数据，我们最初使用 iPhone；后来我们构建了自己的定制装置和帽子。现在我们有超过七种不同的硬件产品，可以在不同的模态之间互换使用。在从不同设备收集数据后，我们致力于同步所有这些不同来源的数据，”Patel 在一次通话中说道。
The company said it is developing ways to fine-tune AI models with its own data and test them on robots to evaluate task effectiveness. By doing this, the startup can demonstrate the quality of its data to potential customers and post-train internal models. Zach DeWitt, a partner at Wing VC, said the startup has a unique advantage in collecting data from multiple sensors. “No one else in the world has been able to synchronize and collect headset RGB-D, force feedback, full-body motion capture, and synchronized chest and wrist camera data at scale. They’ve been doing internal model training on this data, and every major lab and university is interested in running experiments on it due to the novelty of the sensors and the scale of the new dataset they are releasing soon,” he told TechCrunch. 该公司表示，正在开发利用自有数据微调 AI 模型的方法，并在机器人上进行测试以评估任务有效性。通过这样做，该初创公司可以向潜在客户展示其数据的质量，并对内部模型进行后训练。Wing VC 的合伙人 Zach DeWitt 表示，该初创公司在从多个传感器收集数据方面具有独特优势。“世界上还没有其他人能够大规模地同步和收集头戴式 RGB-D、力反馈、全身动作捕捉以及同步的胸部和腕部摄像头数据。他们一直在利用这些数据进行内部模型训练，由于传感器的创新性和他们即将发布的新数据集的规模，每个主要的实验室和大学都有兴趣在其上进行实验，”他告诉 TechCrunch。
Collecting data in India and expansion plans 在印度收集数据及扩张计划
Despite rejection from notable players in the home services industry, Human Archive teamed up with smaller startups to offer discounted services to customers. When a worker arrives at a home, consumers are offered a choice through the app: pay a discounted price in exchange for consenting to data collection, or pay the full price for an unrecorded visit. Patel mentioned that customers have been happy to opt for the former, as disputes about service quality are common, and video recordings can help resolve them. 尽管遭到了家庭服务行业知名企业的拒绝，Human Archive 还是与规模较小的初创公司合作，向客户提供折扣服务。当工人到达家中时，消费者可以通过应用程序获得选择：支付折扣价格以换取同意数据收集，或者支付全价以获得不被记录的服务。Patel 提到，客户很乐意选择前者，因为关于服务质量的纠纷很常见，而视频记录有助于解决这些纠纷。
The company pays workers a base rate of $1 per hour for participating in egocentric data collection. A report from the Economic Times suggests that other companies pay ₹250 to ₹400 per hour (roughly $2.63 to $4.20). Patel said competitors pay more than Human Archive, but its on-the-ground presence in India allows it to keep compensation lower. “Human Archive’s network provides immediate, flexible earning opportunities globally, lowering the barrier to participating in the AI economy. We see this as a critical bridge that funds immediate livelihoods while building the infrastructure for a safer, more productive future,” DeWitt said. 该公司向参与自我中心数据收集的工人支付每小时 1 美元的基本工资。《经济时报》的一份报告显示，其他公司支付的时薪为 250 至 400 卢比（约合 2.63 至 4.20 美元）。Patel 表示，竞争对手的薪酬确实高于 Human Archive，但其在印度的实地业务使其能够保持较低的薪酬成本。“Human Archive 的网络在全球范围内提供了即时、灵活的收入机会，降低了参与 AI 经济的门槛。我们将其视为一座关键的桥梁，在为更安全、更高效的未来构建基础设施的同时，也为当下的生计提供了资金，”DeWitt 说道。