Egocentric Data Collection and India's Governance Gap

Human Rights Research Center
14 minutes ago
6 min read

August 25, 2026

Workers wearing head-mounted cameras for AI and robotic training. [Image credit: R Satish Babu/AFP]

India is emerging as an important hub for egocentric data, where workers use head-mounted cameras to shoot footage of their physical movements and interactions for AI and robotics training. A variety of established and new companies have begun collecting this data, but gaps in India’s current data privacy laws, along with reported concerns about egocentric data collection practices, make this emerging industry a challenge for the country's data governance framework.

Introduction

India is emerging as an important hub for egocentric data, a type of artificial intelligence (AI) training data where head-mounted cameras capture first-person footage of workers performing routine tasks. However, the rapid growth of this industry raises important questions surrounding privacy, labour rights, informed consent, and whether existing governance frameworks are adequate to regulate this emerging form of AI training data.

The Rise of Egocentric Data

AI companies require a variety of resources to create and train their models. Key among these is training data - data that is translated into a language machines can “learn” in order to perform their function. Training data is the bedrock on which modern AI models are built. While large language models require huge swaths of textual data, often numbering in billions of words, to perform the algorithmic process of turning user prompts into coherent and useful responses, robotics and embodied AI systems require an extensive dataset of physical actions. This data is often collected through motion capture. However, the exponential growth of artificial intelligence, along with the growth of robotics, has created a huge demand for training data capturing human physical actions. Consequently, companies have sprung up to provide an alternative — egocentric data — that can be captured, curated, and sold at much greater volumes at a fraction of the cost.

Egocentric data is captured from a first-person perspective, usually with a head-mounted camera, and records hand movements, gestures, and physical actions that are used in the process of performing a particular task. This data typically consists of a variety of actions that require dexterous human input, ranging from measuring and cutting fabric, assembling an electrical component, or even cleaning a room. Thus, egocentric data captures the physical knowledge required to perform real-world tasks.

India's Emerging Egocentric Data Industry

India has emerged as a key environment for the collection of egocentric data. Industry estimates suggest that robotics labs may spend between US$1.5 billion and US$50 billion over the next several years to collect up to a billion hours of egocentric footage, and a growing number of companies have now set up shop in India to collect egocentric data for AI companies around the world.

Broadly, two types of companies are engaged in egocentric data collection: (1) companies that employ freelancers to specifically generate egocentric datasets, and (2) companies that partner with factories where data generation occurs alongside ordinary manufacturing.

While both models generate AI training data, questions have begun to emerge regarding how these datasets are collected, who benefits from their commercialization, and whether existing legal frameworks adequately protect the people involved in generating them.

For instance, Pronto, an Indian app-based homekeeping service provider, launched a pilot in which its workers wore cameras to record in-home footage to train AI systems. While the company claims that customers had the option to consent to the recording, concerns were flagged regarding the collection, storage, and downstream use of first-person recordings, and how informed consent is obtained, how long such data is retained, and whether it can be linked back to identifiable individuals.

Notably, even this ostensibly consensual model centers its protections primarily on the customer. All data privacy rules relate to the customer purchasing the service, while the domestic worker wearing the camera appears largely absent from the consent architecture. The Ministry of Electronics and Information Technology is currently examining Pronto’s egocentric data collection process.

Workplace-integrated data collection raises an additional layer of governance concerns. An investigation by Scroll.in, an Indian news outlet, examined Bengaluru-based egocentric data company Egolab and reported that factory workers interviewed during the investigation were not informed that the recordings would ultimately be used to create AI training datasets. Some workers said they were told to wear the cameras so that the company could monitor how they spent their working hours. Unlike conventional workplace monitoring, however, these recordings possess economic value — once aggregated, standardized, and anonymized, they become transferable AI assets that can be licensed or sold to robotics developers. The same act of labour therefore produces two economically valuable outputs: a physical product for the employer and a reusable AI training dataset for a separate market.

According to the investigation, no written agreements were signed, explicit verbal consent was not obtained, and none of the workers interviewed received additional compensation related to the downstream commercial use of the recordings. This combination of employment-based power asymmetries, uncertain consent, and commercial data generation raises governance questions that are largely absent from the freelance model.

The Governance Gap

Egocentric data occupies an unclear space within India’s legal framework. While experts such as Astha Kapoor of the Aapti Institute have argued that parts of the collection process may fall within the scope of the Digital Personal Data Protection (DPDP) Act, the Act provides limited guidance on how workplace-generated AI training datasets should be governed.

Consider the process of creating egocentric data, which involves two basic steps: (1) recording the physical task performed by the worker, and (2) aggregating and converting the raw footage into usable training data. Once the data is collected, it is processed, standardized, and sold to AI companies as training data.

In the case of Egolab, the recordings are reportedly assigned IDs linked to the worker wearing the camera. Because the footage can therefore be connected to an identifiable individual at the point of collection, this stage appears more likely to fall within the scope of the DPDP Act, obliging the company to take steps to protect the data and inform the workers about the usage and purpose of the data extracted. Once the recordings are aggregated and anonymized, however, the data is processed and contains no such traces. The DPDP Act provides no explicit test for determining when this transition occurs, leaving uncertainty over whether processed egocentric datasets continue to fall within its scope.

Provisions within the Act complicate this further. Section 7(i) permits processing personal data without consent "for the purposes of employment" or to safeguard the employer from loss or liability. Egolab's model of recording workers in exchange for productivity reports would plausibly fall under safeguarding the employer from loss. But aggregating that footage and selling it to third-party AI firms would not - monitoring workers is one thing; selling their data to third parties is another.

These cases reveal a broader governance gap in dealing with egocentric data. The DPDP is similar to existing privacy frameworks, including the European Union's General Data Protection Regulation (GDPR) and California's Consumer Privacy Act (CCPA), in that it primarily provides individuals with digital rights such as notice, access, correction, deletion, and transparency regarding the collection and sharing of personal data. While these rights are important, they do not directly address what happens once workers' expertise has been transformed into a commercially valuable AI training asset, nor do they clarify whether workers retain any continuing rights in the commercialization of that data. Once aggregated, standardized, and anonymized, they become transferable AI assets that can be licensed or sold to robotics developers.

Conclusion

As India emerges as a global hub for egocentric data collection, policymakers may need to look beyond conventional privacy law. Clarifying the applicability of the DPDP Act is an important first step, but future frameworks must consider the obligations of companies that transform recordings into AI training data, privacy concerns for workers, not only consumers, and the rights of workers whose labour simultaneously produces both physical goods and valuable AI training datasets.

Glossary

AI Training Data: Data that is used to teach a machine learning model how to make predictions, recognize patterns or generate content.

California Consumer Privacy Act (CCPA): California’s privacy law that grants consumers rights over how businesses collect, use, and disclose their personal information.

Dexterous Human Input: Input generated through human dexterity—the ability to perform a desired motor task precisely and deftly with ease and skillfulness.

Digital Personal Data Protection (DPDP) Act: India's central legislation governing the processing of digital personal data.

Egocentric Data: Data captured from a first-person perspective, typically using wearable cameras, that records an individual's actions and interactions with their environment.

Embodied AI: Artificial intelligence that learns to perceive, interact with, and perform tasks in the physical world through agents such as robots.

General Data Protection Regulation (GDPR): The European Union's central legal framework governing the processing of personal data.

Informed Consent: Agreement or permission to do something from someone who has been given full information about the possible effects or results.

Motion Capture: Technology that records human movement digitally using cameras or sensors, often for animation, robotics, or AI applications.

Pilot: A small-scale experiment or set of observations undertaken to decide how and whether to launch a full-scale project.

References

HRRC accepts NO funding from governments for our research or campaigns. At every step of the way, our impact is made possible because of people like YOU. Make your contribution now!