As global robotics companies race to build physical AI, India's workers are supplying the human movements that teach machines how to work
In the 1990s, American hospitals began routing physician dictations to transcriptionists in Bengaluru and Chennai. It was unglamorous work, but India was good at it.
Thirty years on, cameras are being slapped on the foreheads of India’s blue-collar workforce, and as they sew a seam, pack an iPhone or assemble a component, the gadget records — every wrist angle, hand movement and intricacy with utmost precision and detail. The data then travels by cloud to robotics labs in San Francisco, where it trains bots to move like people.
Indeed, India is becoming a training ground for robots capable of doing human-like work.
In April 2026, video clips of textile workers at a factory in Delhi NCR wearing small head-mounted cameras went viral on social media. The assumption was that someone was recording them to train robots.
The headgear was identified as originating from Egolab.AI. Founded by 19-year-old Raghav Samani and 18-year-old Varun Pareek, the startup is a first-person point-of-view (POV) data aggregator. It collects ‘egocentric data’, which is first-person footage from body-worn cameras that captures exactly what a robot’s eyes will eventually need to see: the angle of a wrist, the grip on a tool, the small mid-task corrections that no simulation can replicate.
Collecting such data is cheaper and more useful than synthetic simulation. And, just like the transcription work of the 1990s, India has become a gold mine for egocentric data generation — an opportunity that many startups are trying to cash in on.
So, is India writing a crucial chapter in the future of robotics, or once again just doing the cheapest work (outsourcing) while others continue to make the breakthroughs? Let’s find out in this edition of The AI Shift…
The reason egocentric data collection matters today is a gap that has held robotics back for years. Large language models (LLMs) had the internet to pre-train on before being fine-tuned for specific applications.
“Robots, on the other hand, have had no equivalent. A robotic arm can be post-trained for a specific task using simulation or teleoperation (where a human physically puppeteers the robot through a task), but that breaks the moment the environment changes,” said Rushil Agarwal, COO of Human Archive, a data infrastructure startup, which recently garnered attention with the Pronto controversy.
What the robotics industry is now chasing is first-person human-action data across environments to develop a foundation model for physical AI. Egocentric video data is the closest thing to that. It requires no specialised hardware, naturally transfers to a robot’s camera perspective, and costs a fraction of teleoperation data.
As of now, nobody knows whether this data leads to robotics achieving general-purpose intelligence, but it is the largest bet being placed right now, and India is where a large chunk of money is flowing.
The ecosystem includes players of all kinds. India’s on-demand work fulfilment startup Awign, which was acquired by Japanese HR conglomerate Mynavi in 2024, piggybacks on 1.5 Mn gig workers to capture as much as 1,000 hours of 4K footage daily from India alone.
Data labelling startup Objectways is doing the same volume but says the demand for egocentric data runs to 2-3 Lakh hours. It counts Amazon SageMaker AI among its clients. Humyn Labs, Neo Cambrian, and Human Archive are all building similar pipelines.
On the demand side, Scale AI, partially owned by Meta, has already logged 100,000 production hours through its dedicated Physical AI data engine. It is the kind of buyer these Indian pipelines are ultimately feeding.
The money in this business is flowing upward and unevenly. At the collection end, a factory worker in a garment unit earns roughly ₹400 daily for wearing a camera during their shift. According to Thaslim Pattan, who runs a bootstrapped startup called RoBoEra, the startup deploying the hardware pays the factory ₹450-500 per hour.
Meanwhile, more mature startups like Human Archive are pricing data at around $1-10 per hour (about ₹95-₹946), depending on task complexity.
For instance, capturing a garment worker stitching a sleeve needs just a smartphone strapped to their head. However, capturing a warehouse worker sorting randomly placed objects, where a robot needs to identify, orient, and place each item, requires depth cameras, motion sensors, and equipment that tracks every micro-movement. The more complex the task a robot needs to learn, the more expensive the data needed to teach it.
Once annotated and packaged, the data is sold to global robotics labs for $15- $ 50 per hour. Raw unlabelled footage is priced at the lower end, while fully annotated multimodal data is at the top.
AI firms get the data, and factories get the money. From a distance, it’s a win-win for everyone. However, the problem runs deeper.
The Gurugram factory at the centre of the April controversy reportedly engaged in productivity benchmarking, in which worker performance scores are compared across factories. The camera became a management tool.
This is also where concerns related to consent begin to surface.
“There is a part of our contract where we ask the factories to give us rights to data with consent. Whatever the factory people are doing with that, that’s on their part,” said Human Archive’s Agarwal.
India’s DPDP Act 2023 does allow employers to process worker data without explicit consent under ‘employment purposes’, but legal experts argue that using body-worn cameras to benchmark worker productivity against colleagues in other factories stretches that exemption well beyond its intent.
The problem also extends beyond the factory floor. When a house help arrives at a home wearing a head-mounted camera, or a vegetable vendor interacts with a customer being recorded, a third party enters the frame, one that has consented to nothing and been told nothing.
“When you go to a mall or any public place, you will always see a notice that you are under surveillance, something to signify that you are being recorded. If that is not there, then there might be an issue with regard to a third person being recorded without consent or knowledge,” said Harshit Prakash, an advocate at the Delhi High Court.
The worker wearing the camera may have implicitly accepted the recording as a condition of their work, but not the customer whose kitchen is being filmed, the customer whose face appears in the footage, or the child who accidentally walks into the frame. These are not fictitious examples.
Human Archive’s model explicitly involves home service workers, cooks, cleaners and domestic help wearing cameras inside homes and private property. The customer is offered a discount in exchange.
In a country where domestic workers routinely interact with children, elderly family members, and even guests, the question of who else ends up in the training data has no clean answer under the current law.
Almost none of the Indian players in this chain own what the data eventually becomes. They collect, annotate, and ship footage to US robotics companies. The models trained on that footage, the robots those models power, and the intellectual property running through the entire chain belong to companies sitting overseas. India, as always, provides labour, and that too for a song.
To be true, the Indian data collection opportunity is not a technology play yet. It is more like a services contract with high volume, thin margins and no moat whatsoever.
According to an industry executive, data collection rates fell by 30% between January and June 2026 as more suppliers entered the market. Another risk is that synthetic data generation is improving rapidly. As foundation models for physical AI mature, the recurring need for large volumes of human-captured egocentric footage is destined to shrink.
Ashish Taneja from GrowX Ventures, which tracks the physical AI space and has invested in startups like CynLr and Bellatrix Aerospace, frames it as a short-term opportunity. According to him, most businesses in this space may be flourishing right now, but are not scalable ventures.
The more durable bet, as per Taneja, is for a startup that treats data collection as an input, not a product, one that uses proprietary footage to build its own physical AI models, creating assets that compound rather than footage that gets sold once.
As of now, most startups in this space risk getting obsolete unless they evolve, meeting the same fate as the transcription companies of the 1990s.
JustAI Raises $17 Mn: AI-native martech startup JustAI has raised $17 Mn in a Series A round to strengthen its engineering and GTM teams, expand its agentic AI infrastructure, and build products for ecommerce and B2B marketing use cases.
Kapture CX Bags $10 Mn: Enterprise AI startup Kapture CX has secured $10 Mn in a Pre-Series B funding round to expand into new international markets, strengthen R&D, and accelerate product development.
India, US Hold Talks Over Access To Anthropic’s Advanced AI Models: India and the US are engaged in high-level discussions to restore India’s access to Anthropic’s advanced AI models after US export controls restricted availability to foreign nationals
Amazon Commits Additional $13 Bn To India: Amazon has announced an additional $13 Bn investment in India through 2030, taking its total commitment to the country to $48 Bn. The funds will primarily expand AWS data centre capacity in Mumbai and Hyderabad.
Vishal Sikka’s AI Startup Hang Ten Raises $32 Mn: Former Infosys CEO Vishal Sikka’s new enterprise AI venture Hang Ten Systems has raised $32 Mn in seed funding led by Mayfield Ventures. The startup develops agentic AI tools that help enterprises build and operate software using AI-driven code generation and reusable skills libraries.
A significant shift in how frontier AI gets released played out this week after OpenAI launched its GPT-5.6 model series, Sol, Terra, and Luna, but restricted access to a small group of US-based partners approved by the Trump administration.
The announcement stood out because of the political pressure behind it. Earlier this month, the White House signed an executive order, establishing a federal review framework for advanced AI models and giving the government up to 30 days to assess national security risks before public release.
OpenAI briefed the government ahead of the launch and, at its request, began with a limited preview. The company said it does not believe “this kind of government access process should become the long-term default”.
The move follows a harder intervention against Anthropic two weeks earlier, when the Trump administration ordered it to block all foreign nationals from accessing its Fable 5 and Mythos models. Anthropic pulled both models entirely, saying it could not reliably enforce the restriction on a per-user basis.
The common thread across both cases is cybersecurity. The latest frontier models from both companies have drawn government concern over their ability to identify software vulnerabilities at a scale and speed that could be weaponised by malicious actors.
Notably, frontier AI is quietly becoming a nationally controlled resource. If this review framework becomes routine, it would effectively create a two-tier global AI ecosystem, where the most capable models reach US-based developers and partners first, and the rest of the world waits.
For countries like India, which are simultaneously trying to attract AI talent back home and build on top of frontier models, that access gap is worth watching closely.
India has over a million students preparing for JEE and NEET every year, but personalised doubt-solving, the kind that a good coaching teacher provides in real time, remains out of reach for most students outside urban centres and expensive institutes.
Noida-based YoLearn.ai is betting that voice-first AI tutoring can democratise that access. Founded in 2025 by Kirti Prakash Mishra and Vishal Kashyap, YoLearn.ai is building an AI tutoring platform focused on India’s competitive exam ecosystem, covering preparation for JEE, NEET, CBSE, ICSE, and CUET.
The core idea is simple but technically non-trivial: students ask doubts through voice in any of 22 Indian languages, and the AI tutor responds in real time, generating diagrams through a live sketchpad while adapting explanations based on each student’s learning history.
What distinguishes YoLearn from generic AI tutoring tools is its India-first architecture. The platform builds a persistent “Learner Memory Graph” for each student, tracking gaps, progress and patterns over time, and aligns responses to Indian exam curricula rather than generic academic content.
It also offers AI tutor personas modelled on popular educators, which reduces adoption friction for students already accustomed to specific teaching styles.
The startup operates across web, Android and iOS under a hybrid model that combines direct B2C subscriptions with school partnerships. It claims to have crossed 1 Lakh app downloads and 10,000 monthly active users without paid marketing, and has been selected for NVIDIA’s Inception programme.
YoLearn is entering a market that research firm Grand View Research estimates will reach $1.15 Bn in India by 2033, growing at a 37.1% CAGR.
What prompts and hacks are CTOs, CEOs and cofounders using these days to streamline their work?
Here’s the prompt Krupesh Bhat, founder and CEO of Melento, used to generate boardroom-level competitive intelligence for entering the BFSI market.
“You are an elite BFSI advisory council of banking CIOs, compliance leaders, fintech investors, AI governance experts, and enterprise SaaS strategists, advising Melento, an AI-powered CLM and contract intelligence platform expanding into BFSI.
Create a boardroom-level intelligence brief on how contract risk, compliance pressure, vendor governance, operational inefficiencies, AI adoption, internal buying dynamics, and regulatory expectations shape software decisions across banks, NBFCs, insurers, and fintechs.
Analyse workflow pain points, why legacy CLMs fail, GenAI’s impact, whitespace opportunities, buying triggers, competitive landscape, strategic recommendations, and key risks for [company], and present it as a report.
The report should be built based only on credible sources of data such as the RBI.
Also build an executive brief and a six-month GTM plan for it.”
Editor’s Note: Some prompts may need to be adjusted by users for best results or may not work as intended for certain users.
Edited by Shishir Parasher
Source: Inc42 - Startups
