Mint Explainer | Why Pronto's physical AI training pilot is drawing scrutiny

What is physical AI?

Physical AI refers to systems designed to function in the real world— from factories and warehouses to homes. These systems are typically robots or autonomous machines built to navigate spaces and perform physical tasks.

Unlike large language models, which are trained on text and images, physical AI systems require audio-visual and sensor-based data to learn how humans perform actions. These include learning how a person grips a plate, fits a car part, and performs other actions in a cluttered, unpredictable environment.

Simulations can help train such systems, but real-world data is considered far more valuable because of the unpredictability and variation involved in physical environments.

Given the scale at which the home services business operates in India, for companies offering physical AI training data, the opportunity is vast, and valuable.

According to Entrackr, which first reported the news, citing an internal memo from investor Gladebrook Capital, Pronto is "developing a data business leveraging its workforce to capture real-world household data for robotics labs.”

The company has said the pilot is “very small” and fully opt-in, with customers required to give consent before each booking rather than through an approval during sign-up.

Pronto founder Anjali Sardana told Mint it's also an effort to protect their partners when bookings are made by consumers away from home. “There are people who told us they hesitate to book services when they are away from home, or wish they had more reassurance around what happened during a booking.”

Other home services companies, including Urban Company and Lightspeed Venture Partners-backed Snabbit, however, said they wouldn't explore such a model.

India's Digital Personal Data Protection Act 2023 mandates that companies obtain specific, informed consent for every distinct purpose for which personal data is collected.

“If the data is genuinely anonymized and individuals are no longer identifiable, then it may fall outside the DPDP Act, since the Act regulates personal data. India still lacks a non-personal data framework that squarely governs anonymised datasets used for AI training,” said Nikhil Narendran, partner—technology, media and telecommunications practice at law firm Trilegal.

Sumeysh Srivastava, partner at the Quantum Hub, a policy consulting firm, said that the anonymization defense has limited standing under current law—since the DPDP offers no definition or benchmark for what constitutes anonymized data, there is no enforceable standard against which a company's practices can be measured.

While the DPDP's Data Protection Board provisions took effect in November 2025, the substantive obligations on consent and purpose limitation don't become binding until May 2027.

“The obligations set a clear test: consent has to be specific to a stated purpose, so a customer agreeing to a cleaning service wouldn't by itself be agreeing to have footage used for AI model training,” said Srivastava.

This isn't just a home services phenomenon, but is seen across industries ranging from warehousing and logistics to production and assembly lines that require human intervention to fit specific parts.

It extends even to agriculture. Agricultural machinery manufacturer John Deere is seeking to integrate AI systems into its product lines to enable more efficient large-scale farming.

Other companies, like Cyn:LR, have created their own training data, but without using humans. In February, the company created its own intelligence model, which, when deployed on its own industrial robots, can be used in areas which have typically been harder to automate.

Prosus-backed Deccan AI, which works on reinforcement learning, pays experts across different subjects to determine which LLM responses are better.