The Indian workers filming their work to train AI

Workers in India and elsewhere record themselves performing physical tasks. Their footage can show hair cutting, garment-making, embroidery, agriculture, pottery, candle-making, household chores and industrial work. The goal is to capture human actions that robots may eventually need to copy.

Mohammed Ashim, a Bengaluru barber, wears a camera that records his hands and fingers while he cuts hair and trims beards. Local data companies collect the recordings, filter and label them, then sell useful packages to robotics companies and frontier AI labs. Some workers also wear hardware on their hands to capture contact and force.

The buyers use this data to develop physical AI, which can control robots in the real world. Rashdan Technologies and Humyn Labs are examples of companies collecting and processing the footage. Their customers increasingly request several cameras to capture a person, the task and the surrounding environment.

Physical AI is artificial intelligence that can guide robots through real-world activities. It needs more than language or image recognition. A robot must move its limbs, handle objects, respond to surroundings and apply suitable force while completing a task.

For example, a system learning to handle containers must distinguish between crushing a plastic bottle and holding a steel glass securely. Human recordings can show finger, wrist, elbow and shoulder movements, along with the changing forces involved. Cameras and wearable hardware capture these signals during real work.

Text- or image-focused AI works mainly with information represented as words or pictures. Physical AI must connect perception to action. The article describes an emerging field whose data is still being collected from scratch. Its aim is to help robots perform industrial tasks, household chores and other activities outside computer simulations.

The article describes a global workforce of hundreds of thousands recording everyday and industrial activities for AI development. The workers are spread across many countries and tasks. Their recordings give technology companies access to physical actions that previously were not available in a large shared archive.

Humyn Labs currently has 96,000 data collectors across countries in Latin America and Southeast Asia. Rashdan Technologies pays $5 to $10 for eight to 10 hours of recorded work. Ashim, the Bengaluru barber, wears his camera for five or six hours a day, although the article does not give his individual pay.

These figures show both the scale and the uneven structure of the industry. Pay is described as meaningful for many workers, but companies pay more when collecting harder data, such as footage from four to six cameras. More competitors are also expected to enter India, increasing the industry’s reach.

Companies try to protect privacy by limiting what cameras record and removing identifying information. Rashdan Technologies says its usual setup records hands rather than faces. This lets clients study actions without directly identifying the worker or customer.

Ashim’s camera records his hands and fingers while his client’s face is blurred. When several cameras capture more of the body or surrounding environment, Rashdan says it blurs faces there too. The company also says it sends customers no names, contact details or other personally identifying information.

Privacy becomes harder as clients request four to six cameras mounted around a person’s head. Wider coverage can show the entire scene, but it can also capture more people and surroundings. Rashdan says it seeks consent, checks that workers are comfortable and pays more for this demanding data-collection tier.

Computer simulations cannot fully provide the variety and physical detail found in real human work. Robotics companies need evidence of how bodies move through tasks, how objects respond and how much force actions require. These details matter when robots operate outside controlled digital environments.

Agarwal gives the example of gripping a plastic bottle without crushing it while holding a steel glass securely. Recordings can capture fingers, wrists, elbows, shoulders and dynamic force application. Wearable hardware may capture physical contact, while cameras show the wider task and setting.

The article says there is no existing internet archive of embodied human activity. Companies must collect this material from scratch across diverse environments. Humyn Labs therefore gathers footage from different factories, tasks and geographies. Data from Indian homes, for example, may not provide the right context for robots intended for US houses.

Workers are helping create data for machines that may eventually perform some of the same jobs. This creates a possible tension: recording work can provide income now, while better robots could reduce demand for human labour later. The article does not say that widespread replacement has already happened.

Bino Paul of TISS says AI might be a useful tool for some workers, but robots could automate other jobs. A barber, craft worker or factory employee may contribute recordings that teach a robot the movements needed for similar work. The data companies pay workers for collecting it, but ownership of later outcomes is not described.

Paul says it is too early to tell what will happen. He warns that by the time workers understand what they have participated in, developments may have gone too far. He wants governments to regulate AI and treat it as a socio-technical system involving ethics, not merely as a technical system.

Machine-learning systems need organised examples before they can learn from human activity. The companies in the article collect recordings, remove unsuitable material, identify what appears in each clip and prepare the results for technology clients. This turns raw activity into structured training data.

Humyn Labs filters thousands of hours of video and selects the most useful content. It then labels the footage so an AI system knows what it is looking at, and packages the data in a format clients can use. Video can show limbs and movements, while hand and finger hardware can capture physical contact and force.

The article does not describe the exact software or training algorithm that converts these labels into robot instructions. It does explain the foundation: diverse examples of tasks, movements, environments and force. Robotics companies use that prepared data to develop physical AI capable of controlling robots in real-world settings.