MUMBAI, India — Oct. 10, 2026 — Inner Sky Labs today unveiled a complete stack of foundation models that lets machines sense, decide and act safely around people. Every great leap in human productivity has come from a new kind of machine working alongside people: the loom, the steam engine, the industrial robot. Most of the world’s economy, by some estimates as much as 80% of global GDP, is created in the physical world, where language models alone fall short. For a construction project in Delhi, a warehouse in Bhiwandi, a farm in Punjab or a car manufacturer in Stuttgart, physical AI means a smarter machine that follows their lead, intelligently takes the strain out of the job and lets their skill go further. A machine that works this close to people must be one they can trust, because a machine’s mistake could hurt someone.
The stack gives a machine three things a human has by instinct: the ability to sense what’s happening around it, the ability to decide what to do about it, and — critically — an independent safety check to stop any unsafe move.
Teaching machines to see, hear and feel
Before a machine can act safely, it has to understand what’s actually happening around it — not just see it, but make sense of it the way a person does. Inner Sky Labs’ perception models give a machine that understanding: the Drishti series of vision models reads a scene the way a pair of eyes would, the Shruti series of audio models makes sense of sound — a spoken instruction, the crash of something falling — and the Sparsh models give a machine a working sense of touch, so a robotic hand can tell the difference between gripping an egg and gripping a wrench. Put together, a machine stops being a sensor bolted to a motor and starts being something that actually perceives the room it’s in.
Teaching machines to decide, and to think ahead
Understanding a room is only half the job — a machine also has to decide what to do next, and good judgment means thinking a step ahead. Inner Sky Labs’ action models — Kriya, Prana and Karma — supply generative action intelligence to plan and carry out physical movement — reaching, gripping, walking, flying — and also to predict what happens next: if this arm moves this way, does the box slip, does the person standing nearby step forward. It’s the same kind of anticipation a person uses to catch a glass before it tips, built into the machine’s decision-making instead of left to reflex.
Kavach: the part of the brain whose only job is to say no
Sensing and deciding well is still not the same as being safe — a machine can understand a room perfectly and still make the wrong call. That’s what Kavach — Sanskrit for “shield” — is for: a system that checks every single action a machine is about to take, and has the power to stop it, no matter what the rest of the machine has decided.
Picture a robotic arm about to swing across a factory floor. Before Kavach lets that motion proceed, it checks: is that space clear of people right now, is the planned speed safe for that area, does this action stay inside the zone it’s allowed to work in. If the answer to any of those is no, the arm does not move. Not “slows down.” Not “flags it for review afterward.” The action is blocked before it starts.
A decade of trust, now scaling up
Inner Sky Labs already knows what it means to be trusted with something fragile. It is the company behind Miko, the companion robot that has spent the past ten years engaging and educating children to improve their social-emotional skills and is now in more than 140 countries. An early version of the same safety thinking that has spent a decade deciding what Miko may say to a child is now the foundation deciding what any machine — an industrial arm, a delivery robot, a drone — is allowed to do. In an independent evaluation by BDO’s cyber security practice, Miko’s safety system caught every unsafe prompt across seven public AI-safety benchmarks, ahead of four widely used commercial AI models, and ranked first at flagging unsafe AI responses. Full methodology at www.innerskylabs.ai. A decade spent earning a parent’s trust with a robot for their child is now the foundation for earning a factory’s, a hospital’s, and a government’s trust with machines that move.
Why this matters beyond one company
As machines move into hospitals, factories, streets and homes, whoever writes the rules a machine lives by decides what that machine is allowed to do near a citizen — and today, most of those rules are written somewhere else, in a language, a legal system and a set of values that aren’t their own. Inner Sky Labs’ system lets any end customer or authority write its own rules — its own definition of what is safe, in its own legal language — and have a machine obey them, without depending on another company’s judgment call. Built and proven in Mumbai out of more than ten years of research by IIT Bombay alumni, this is a technology that lets any country decide for itself what its own machines are allowed to do.
The full stack already runs today inside factories, classrooms and consumer products in millions of homes.
“Every robot company will eventually have to answer one question: why should anyone trust this machine near a person,” said Sneh Vaswani, co-founder and chief executive of Inner Sky Labs. “For ten years, our answer was Miko — a robot trusted enough to be left with a child. Today we’re answering that same question for every kind of machine, with technology built, tested and proven out of India.”
“I’ve advised Sneh and the team since 2016, when they first shared this multi-decade vision with me—long before ‘Physical AI’ was even a category,” said Keshav R. Murugesh, early investor, advisor to Inner Sky Labs, and former Chairperson of NASSCOM. “Today, Physical AI is being called a $50 trillion opportunity. I am proud to see this built in India for the world. Miko was the proof point. This is the scale-up.”
Availability
The platform is available today to qualified partners by invitation, with cloud, on-premise and on-device deployment, including fully offline, air-gapped environments for operators who cannot depend on an outside network. Each model — perception, action and Kavach’s safety layer — is available on its own. The company has raised over $80M to drive frontier research from investors including Chiratae Ventures, IvyCap Ventures, YourNest Venture Capital and 360 ONE.