Crowdsourced Data Collection

Distributing robot data-collection tasks to large numbers of non-professional workers, who are paid for each validated, usable clip.

Crowdsourced data collection doesn't rely on a small team of trained operators. Instead, a website or app distributes tasks to large numbers of ordinary people, who either teleoperate a robot remotely or record themselves performing tasks at home or at work; the platform reviews submissions and pays based on validated, usable duration. An early example is Stanford's RoboTurk (2018), which let users control a robot arm through a browser using their phone as a 6-DoF controller. Between 2025 and 2026, as first-person video and embodiment-free capture devices became more common, crowdsourcing shifted toward real-world human data: Figure reportedly opened a paid project called “Index” in August 2026 to collect users' task videos, and the Chinese startup Mifeng Technology (觅蜂科技) launched a crowdsourced app called “Mifeng Pai” (觅蜂派) in September, where users take on data-collection jobs while wearing its MEgo capture device. The advantages are low cost and broad scene coverage; the challenges are inconsistent quality, privacy protection, and the cost of cleaning and labeling the results.

ExampleStanford's 2019 real-robot RoboTurk dataset was collected by 54 non-expert users teleoperating robots remotely over the course of a week, yielding 2,144 demonstrations totaling 111 hours.

Also called

Crowdsourcing

Related

Data Collector (Teleoperator) · Robot-free (Embodiment-free) Data Collection · Egocentric Video · Valid (Usable) Data · Data Anonymization · Data Quality Control

Sources

RoboTurk - Crowdsourcing Robotics（斯坦福） (Chinese)
机器人开始向更多人类买数据（新京报） (Chinese)
日薪120元全民数采：谁在训练下一个机器人保姆？（36氪） (Chinese)

As of

2026-09