智元系的数据生意:时薪20块,喂不饱机器人

林越2026-10-081310具身智能

2016年,自动驾驶公司Comma.ai推出了一款行车记录应用Chffr。用户下载安装后,手机就变成一个行车记录仪,应用会自动录制驾驶过程中的数据,随后上传至公司的云端服务器,用于自动驾驶模型训练。十年过去,这家公司没有成为自动驾驶的主角。如今,具身智能行业把这套打法又搬了出来。9月末,觅蜂科技正式发布具身数据众包平台觅蜂派,包括自研的MEgo数采设备、觅蜂派App和数据治理引擎MEgo Engine三个组成部分。

 

In 2016, autonomous driving company Comma.ai launched a dashcam app called Chffr. After users downloaded and installed it, their phone became a dashcam. The app automatically recorded data during driving and then uploaded it to the company's cloud servers for training autonomous driving models. Ten years later, the company did not become a leading player in autonomous driving. Now, the embodied intelligence industry has brought this playbook out again. At the end of September, Mifeng Technology officially released an embodied data crowdsourcing platform, Mifeng Pai, which includes three components: the self-developed MEgo data collection device, the Mifeng Pai App, and the data governance engine MEgo Engine.

一小时赚20块


物流中心的快递小哥戴着头显设备分拣一个小时的快递,宝妈戴着设备做一顿饭,旅游景区一位工作人员记录下自己日常打价与贴标的过程。这些数据上传审核后,可以按每小时10到20元结算报酬。下载打开觅蜂派App后,会先看到一个任务大厅,其中有22个场景大类,包含生产类、生活服务类、餐饮类、物流类、医疗类、居住类等。如果想要赚取这一份外快,需要先去申领一套数据采集设备。设备的租金为每天39元,推广期19元。如果按每小时10到20元算,推广期也要先做满一到两个小时有效数据才能挣回当天租金。官方给出的数据是,觅蜂派内测一个月内,注册用户约2万人,累计提交任务1.3万份。不过简单换算下来,人均提交不足一份任务。

 

Earning 20 yuan an hour

 

A courier at a logistics center wears a head-mounted device while sorting packages for an hour. A stay-at-home mom wears the device while cooking a meal. A staff member at a tourist attraction records their daily process of pricing and labeling. After these data are uploaded and reviewed, compensation can be settled at 10 to 20 yuan per hour. After downloading and opening the Mifeng Pai App, users first see a task hall, which includes 22 major scenario categories, covering production, life services, catering, logistics, healthcare, residential, and others. If they want to earn this extra income, they first need to apply for a set of data collection equipment. The rental fee for the equipment is 39 yuan per day, or 19 yuan during the promotion period. Calculated at 10 to 20 yuan per hour, even during the promotion period one would first have to produce one to two hours of valid data to earn back that day's rent. Official figures show that within one month of Mifeng Pai's internal testing, there were about 20,000 registered users and a cumulative 13,000 tasks submitted. But a simple calculation shows that this is less than one task submitted per person on average.

 

三种数据路线


目前行业内解决数据不足问题的方式主要有三种。真机遥操作路线的数据质量最高,但产出量最有限。无本体数据采集由人来佩戴设备进行数采,产出量更多,但数据需要通过清洗、标注、治理等多个环节才可用。最后是仿真数据和互联网视频数据,量最大,但质量最差。大多数数据采集厂商不会押注单一的数据路线。从行业竞争的角度来看,目前的数据厂商可以大致分为三类。第一类是机器人本体公司自有的数据和模型业务,代表是宇树、星海图、银河通用等。第二类是专业数据服务商,包括光轮智能、帕西尼、灵初智能、它石智航、手亿科技和觅蜂科技等。第三类是互联网大厂的数采平台,目前比较有代表性的是京东。

 

Three data routes

 

At present, there are mainly three ways the industry addresses the problem of insufficient data. The real-machine teleoperation route produces the highest-quality data, but its output is the most limited. Bodyless data collection, in which people wear devices to collect data, produces more, but the data must go through multiple steps such as cleaning, labeling, and governance before it can be used. Finally, there are simulation data and internet video data, which are the largest in volume but the lowest in quality. Most data collection vendors do not bet on a single data route. From the perspective of industry competition, current data vendors can be roughly divided into three categories. The first is the data and model business owned by robot body companies themselves, represented by Unitree, Galaxea, and Galaxy General. The second is professional data service providers, including LightWheel Intelligence, PaXini, Lingchu Intelligence, Tashi Zhihang, Shouyi Technology, and Mifeng Technology. The third is data collection platforms from major internet companies, with JD.com currently being a relatively representative example.

 

几笔账还没算清


觅蜂科技在这个行业里的站位比较特殊。虽然定位是独立的数采平台,但是与智元的联系紧密,董事长姚卯青同时也是智元合伙人。这层关系给觅蜂带来了订单、场景和背书,也带来一个绕不开的疑问:其他本体厂商是否愿意把数据需求交给一家智元系公司。在真正形成数据规模壁垒前,觅蜂科技还要算清几笔账。第一笔是经济账,目前行业内的行情是高质量的无本体数据每小时可以卖到200到300元,而觅蜂科技数据众包做的纯头显EGO数据,价格差不多是每小时10到15元,跟觅蜂派里面的报酬水平基本持平,算下来是亏本收数据。第二笔是数据有效性的账,目前单一头显的EGO数据有效率差不多在10%到20%,也就是说1小时的数据里面,可能只有10分钟左右是有用的。第三笔账最容易被忽视,就是数据的存储成本,大概是每50小时数据需要1.4T左右的存储空间。觅蜂科技通过免费发设备和数据众包的方式,也许可以在短时间内积累起很漂亮的数据,但数据的量上去了,有效性、经济性、存储成本问题如何解决,可能才是更大的考验。

 

Several accounts have not yet been settled clearly

 

Mifeng Technology's position in this industry is rather special. Although it positions itself as an independent data collection platform, it is closely connected to AgiBot, and chairman Yao Maoqing is also a partner at AgiBot. This relationship brings Mifeng orders, scenarios, and endorsement, but it also brings an unavoidable question: are other robot body manufacturers willing to entrust their data needs to a company in the AgiBot ecosystem? Before truly forming a data scale barrier, Mifeng Technology still has several accounts to settle clearly. The first is the economic account. The current market rate in the industry is that high-quality bodyless data can sell for 200 to 300 yuan per hour, while the pure headset EGO data that Mifeng Technology crowdsources is priced at about 10 to 15 yuan per hour, roughly in line with the compensation level inside Mifeng Pai, which works out to collecting data at a loss. The second is the account of data validity. At present, the validity rate of EGO data from a single headset is roughly 10% to 20%, meaning that in one hour of data, perhaps only about 10 minutes are useful. The third account is the easiest to overlook, namely the cost of data storage, which is roughly about 1.4T of storage space for every 50 hours of data. Through free device distribution and data crowdsourcing, Mifeng Technology may be able to accumulate impressive-looking data in a short time, but once the volume of data goes up, how to solve the problems of validity, economic efficiency, and storage cost may be the bigger test.