Human Archive Wants Data to Train Physical AI—By Whatever Means Necessary

As home-cleaning services startups Pronto, Snabbit, and Urban Company scramble to either defend or deny collecting data through in-home recordings, a physical AI-linked data collection company at the centre of the storm is fighting its own data privacy demons.

In a conversation with AIM, two former employees of Human Archive, a Y Combinator-backed AI and robotics data lab, have raised ethical and privacy concerns about how the company collects data using egocentric (captured from a first-person point of view) and multimodal data through body-worn cameras, gloves, microphones, and motion sensors in restaurants, factories, and homes. They claim that the workers who wore their equipment and their customers were not always aware of how the data was being collected, stored, and eventually monetised.

The company, founded this year by UC Berkeley and Stanford researchers Raj Patel, Samay Maini, Rushil Agarwal, and Shloke Patel, taps into gig and factory workers to collect training data of humans at work for physical AI systems. It has raised $8.2 million this week, led by Wing Venture Capital, NVP Capital, and Y Combinator.

However, its data collection practices came sharply into question when an Entracker report revealed that Pronto was running a limited physical AI data pilot programme that raised data privacy concerns and led to an investigation by the Ministry of Electronics and Information Technology. 

The Anjali Sardana-led company, after bagging $20 million from physical AI-focused investor Lachy Groom, enrolled some professionals who wore cameras that recorded their tasks, such as cleaning homes, washing dishes, and folding laundry. Pronto claimed that the company deleted footage within 48 hours, but the report noted that the initiative generated real-world training data for physical AI labs.

Human Archive earlier approached Pronto, Snabbit, and Urban Company for a collaboration. While Snabbit confirmed exploring the proposal, it denied involving it in active operations. 

Co-founder Agarwal also called out Pronto on X.

Back in January I spoke to @anjalisardana from pronto, she laughed at me and basically called me stupid for doing this and for some reason has a weird obsession with shitting on early founders. Look at them now, Iwk props for catching onto the trend https://t.co/itoTjAKW7Z

— Rushil Agarwal (@Rush26agarwal) May 24, 2026

Meanwhile, Co-founder and CEO Raj Patel had also responded to Urban Company Co-founder and CEO Abhiraj Singh Bhal’s reaction to the controversy, denying engaging in any data collection activity. 

Human Archive (YC W26) was one of the first companies in the world to start collecting human data, and I said this last year, but soon you will be able to get free cleaning and cooking services anywhere in the world by just consenting to being recorded. Most people clearly still… https://t.co/38n3FzpIFx

— Raj Patel (@babugi28) May 24, 2026

The episode has now become a case study in how India’s rapidly expanding physical AI ecosystem is colliding with unresolved questions around surveillance, labour rights, governance, and data ethics.

How Human Archive Collects Data

Former Human Archive employees say the data collection devices span wearable GoPro head cameras, motion tracking gloves, microphones, and inertial measurement units that recorded workers performing physical tasks across different environments.

Suresh (name changed) alleges that the company deployed wearable hardware on workers for long hours in restaurants, houses, and industrial settings

“Some workers were underage, and many labourers did not fully understand the extent of recording taking place through microphones and wearable cameras,” Suresh tells AIM. 

AIM has reviewed the footage and visuals provided by the former employees of workers in apartments and restaurants wearing the devices. 

The former employees have also raised concerns about dubious audio recording through microphones embedded in wearable systems. Suresh alleges that some workers themselves were unaware that their voice recordings were also being collected alongside video data. 

Fenish (name changed) also claims that many workers were not fully informed about the technical depth of the data being captured. “For a long period of time, even we didn’t know voice recordings were also being collected,” he reveals.

Moreover, they say the workers had to wear the hardware devices for long stretches, and they would overheat to dangerous levels, leaving some workers with scars across the body. 

He mentions multiple accidents in different places wherein the workers were physically hurt with burns and other health issues. He claims that these workers were never given compensation for these accidents. 

He also alleges that the workers were not paid for wearing the devices. Instead, the earnings went to their employers.

He further notes that data collection was being aggressively scaled across multiple cities, including Jaipur and Gurugram. “They used to sell this data to big companies. Companies collect the data and then sell it to companies that need it. That’s how the ecosystem works.”

Fenish also notes that the company sold physical activity datasets to larger AI firms. He also acknowledges that worker-level consent was not always directly obtained.

“I agree that consent is a very big concern. In many cases, managers knew recording was happening, but workers individually may not have been informed,” he claims.

Human Archive, Snabbit Respond

AIM sent a detailed list of questions regarding all the allegations. However, the company did not directly address specific allegations regarding worker consent, voice recordings, or claims involving underage labour.

On data privacy, Co-founder and CEO Raj Patel responded that the company was committed to complying with India’s upcoming data protection framework and to reviewing the concerns that had surfaced. 

“We take compliance with data privacy and consumer protection laws extremely seriously. We are fully committed to meeting all applicable legal and regulatory obligations, including the requirements of India’s forthcoming DPDP framework, ahead of its implementation in 2027,” he replied.

The startup further stated that it was “conducting a thorough review of the additional matters that have been raised to ensure full compliance with all applicable laws and regulatory standards,” and added that “where appropriate, we will take any necessary corrective or reinforcing actions promptly and decisively.”

Pronto, Snabbit Deny Rollouts

Following the controversy, both Pronto and Snabbit issued detailed statements distancing themselves from allegations that they collected customer home data through Human Archive.

Pronto told AIM that it had recently started a “strictly opt-in” pilot involving outward-facing wearable cameras used by professionals during specific jobs. The company said the feature applied to less than 0.01% of customers and required explicit consent during booking.

“Pronto does not, by default, record in customers’ homes, nor does it use customer data,” the company said in its statement. The company added that recordings are anonymised, automatically blurred for personally identifiable information and deleted within 48 hours.

“We are not sharing data with any third-party AI labs,” the statement noted.

Snabbit, meanwhile, confirmed that it never entered into any operational partnership with Human Archive.

“No partnership, collaboration, contract, commercial arrangement, joint development effort, or operational deployment was ever entered into with Human Archive,” a Snabbit spokesperson told AIM.

The company confirmed that exploratory discussions had taken place under a standard NDA but stated that no pilot, customer-home rollout, or commercial deployment ever occurred.

Snabbit also said that a one-time assessment was conducted within a controlled training-centre environment with consent from participating workers.

“Customer trust, expert dignity, and privacy come first, and we would not pursue any technology or process that compromises those principles,” the company affirmed.

Why Household Data is Becoming Valuable

A 2025 academic paper published in the Journal of Artificial Intelligence and Engineering Applications noted that physical AI systems rely on technologies such as multimodal processing, deep learning, and real-world sensor-driven environments to train robots capable of functioning autonomously. However, physical AI faces significant ethical and operational concerns, including “privacy, bias, social acceptance and adversarial attacks,” especially as these systems begin operating in unregulated spaces and human environments.

Experts say the controversy reflects a broader global shift in how AI companies are training robots and autonomous systems.

Kalyani Khona, an independent AI researcher, explains that AI companies are no longer interested in isolated tasks such as folding clothes or ironing independently. Instead, they want systems capable of understanding complete household environments in real time.

She explains that companies are now pursuing “4D world models” that combine time, movement, space, and environmental understanding to build general-purpose robotic intelligence.

“This is not new. What is new is the level people are going to for data now,” she exclaims.

She argues that countries like India, Vietnam, and the Philippines are becoming attractive markets for physical AI data collection because of cheap labour, widespread internet access and fewer regulatory barriers than in Western countries.

Many companies may not even intend to build robots themselves, but instead monetise datasets by selling them to larger AI firms globally. “The bigger question is where this data eventually goes,” she wonders.

The allegations have reignited concerns over surveillance inside private domestic spaces. One of the biggest unresolved concerns involves children and other individuals who may appear inside recorded environments without directly consenting to it.

“A child becomes ten years older, but their household data may still exist online. The parents gave consent, the worker gave consent, but the child never did,” Khona says.

She adds that even AI-based anonymisation systems are not foolproof. “What if something gets missed? What if household dynamics become traceable through leaked datasets?” she questions.

A Regulatory Vacuum

The data privacy concerns surrounding Pronto’s use of data collection devices and Human Archive’s collection methods have ultimately exposed how quickly physical AI development is outpacing governance frameworks in India. Startups are increasingly positioning homes, factories, and workplaces as training grounds for future AI systems.

Although India’s Digital Personal Data Protection (DPDP) Act and responsible AI guidelines offer some protections, experts argue that physical AI introduces entirely new categories of ethical and legal questions.

Who owns human behavioural data? Can workplace consent be treated as meaningful consent in informal labour environments? What happens when household recordings include people who never agreed to be part of AI training datasets? And who remains accountable if such datasets are leaked, monetised, or misused?