Pronto Turns Home Cleaning Visits Into Robot Training Data

Pronto, an Indian quick-commerce-style home services startup that says it has served more than half a million households across 11 cities, is now converting the daily work of its cleaning professionals into training data for robots. In October 2026, founder Anjali Sardana set up Provision Labs, a physical AI venture based in San Francisco that will collect data to train and test AI systems designed to perform real-world tasks with robots and other machines.

A few days after Provision Labs was announced, the company acquired Mumbai-based Basilisk Industries, a startup founded in 2023 that builds perception technologies, which help machines interpret their surroundings. Basilisk founder Shourya Watwani will lead Provision's hardware efforts and work on its in-house post-processing models. Watwani and his team are reported to have developed a 3D hand-tracking model and built a first version of a lightweight tactile glove for capturing touch data.

The move marks a sharp expansion for a company that entered India's domestic services market in 2025 to solve the problem of finding reliable house help on demand. Pronto competes with Snabbit and Urban Company.

How does the data collection work?

The first reports of Pronto's data ambitions appeared in May 2026, when the company was reported to be exploring physical AI training data from footage captured during household chores. In July, it formalised the experiment through Verified, an opt-in premium service in which housekeepers and cleaners wear head-mounted or body-mounted cameras during visits.

The service triggered a backlash over surveillance and privacy. Pronto says customers are informed about recording terms several times before it begins, and personally identifiable information is automatically blurred or removed on the device.

"The raw footage is deleted immediately after the work is done, and no one ever accesses it," the company told Inc42. "For 48 hours, an anonymised, end-to-end encrypted recording is available only to the customer, who can watch it in the app to check the work. After that, we use the anonymised footage for AI training, as our terms set out."

Pronto says more than half of the professionals participating in Verified already earn over ₹40,000 a month. It claims workers feel safer knowing a record of their work exists, but no independent studies have validated these claims.

The relationship between Pronto's household data collection and Provision Labs' operations remains unclear. Inc42 asked Pronto how the two businesses work together, and the company declined to comment.

Why is this footage so valuable?

Unlike chatbots, which learn from vast repositories of online text and images, physical AI systems must understand the real world. A robot has to learn how to pick up objects, fold clothes, navigate a room, and respond when something goes wrong. That requires large volumes of real-world, multimodal data capturing movement, spatial information and object interactions.

ANI quoted a Goldman Sachs report estimating that general-purpose robots could require around 10 Mn hours of high-quality physical interaction data to reach production-level capabilities. Such data is far harder and more expensive to collect than internet-sourced text and images.

The physical AI stack has four layers. The data layer captures real-world interactions, while the model layer trains vision-language-action models that translate what a robot sees into actions. The simulation layer, through platforms such as NVIDIA's Isaac Sim, tests these models virtually before deployment. The edge layer runs them on physical robots, with real-world feedback improving training.

NVIDIA, Google DeepMind, Amazon, OpenAI and Anthropic are building capabilities in the model and simulation layers. India has robotics players such as Addverb, CynLr and Perceptyne at the deployment end. But the data layer is where a growing crop of Indian startups is concentrated, alongside global data companies running collection operations that rely heavily on India's workforce and physical environments.

Humyn Labs, EgoData, Egolab.ai, Neocambrian AI, Human Archive and Objectways have deployed people on the ground to collect footage across household chores, manufacturing units and other industrial settings. Head-mounted cameras, sensor-equipped gloves and full body suits have become commonplace.

For India, the opportunity echoes its earlier role in the AI supply chain. Over the past decade, the country became a hub for data annotation. Physical AI takes that outsourcing model a step further: Indian workers' movements, workplaces and object interactions are becoming the training material itself.

Abhinav Kukreja, founder of Neocambrian AI, sees India's large and diverse physical environments as an advantage over the US, where access to comparable real-world data is more limited. "I strongly felt that this layer would not be compute, architecture or algorithms, but the data layer," he said.

Y Combinator-backed Human Archive pursues a similar model, with cooks, cleaners and domestic help wearing cameras inside homes and private properties. Cofounder Rushil Agarwal told Inc42 earlier that India's appeal lies in the diversity of industries within one geographical boundary and its large labour pool, which is difficult to replicate in the US. Human Archive raised $8.2 Mn in a seed round led by Wing Venture Capital and NVP Capital.

Snabbit also evaluated Human Archive as it considered launching a Pronto-like data collection pilot, but eventually dropped the proposal. Pronto's $45 Mn Series B saw participation from Lachy Groom, cofounder of robotics model startup Physical Intelligence, who has also backed Provision Labs. Bengaluru-based Humyn Labs, led by former Nazara CEO Manish Agarwal, has committed $20 Mn towards expanding its data collection network.

What about pricing and the workers?

The supply side is becoming more competitive. Data collection rates fell 30% between January and June this year, an industry executive told Inc42. More suppliers have entered the market and synthetic data generation has improved rapidly, though demand for human-captured footage could shrink as physical AI models mature.

Compensation structures remain murky. Pronto says over half of Verified professionals earn more than ₹40,000 a month, while Human Archive has cited rates of $1-10 per hour for data collection work. Companies are spending heavily to attract workers who may collect the data that replaces them in a few years.

Social media videos have raised questions about whether workers understand why they wear cameras. Some report discomfort from devices worn 5-10 hours a workday; others believe the devices monitor their productivity. Every company maintains that workers are informed, but when collection runs through partner businesses, the responsibility for communication shifts to those partners.

An openDemocracy investigation found that workers at two Delhi factories were asked to wear cameras supplied by Egolab.ai without being told their purpose. India's Digital Personal Data Protection Act, scheduled for implementation in May 2027, adds regulatory uncertainty, and it remains unclear how far the government can restrict individual participation.

The larger risk is value drain: Indian workers and factories may supply the footage while the resulting models and intellectual property are developed and owned elsewhere. For data suppliers, the challenge is moving beyond one-off collection contracts to differentiated datasets that hold lasting value as first-person footage becomes commoditised. For Sardana, the data from Pronto's network may soon be worth more than the service itself — and whether Pronto and Provision Labs formally combine remains the unanswered question.