Physical AI: Inner Sky Labs launches a safety-first stack for machines that move
Inner Sky Labs, the Mumbai company behind the children’s robot Miko, has launched a full stack of foundation models for Physical AI — the branch of artificial intelligence that gives machines the ability to sense their surroundings, decide what to do and act safely around people. The company announced the platform on 10 October 2026 and, at the same time, said it is closing a ₹300 crore (about $30.9 million) funding round.
The launch marks a strategic pivot. For a decade, Inner Sky Labs sold a consumer product: Miko, a companion robot for children, now sold in more than 140 countries. Its new platform sells the underlying intelligence instead, to other companies that build robots, drones and industrial machines. The company positions Physical AI as the next productivity frontier, noting that most of the world’s economic output happens in the physical world, where language models alone fall short.
What Inner Sky Labs launched
The stack organises into three layers. The perception layer reads the environment: Drishti handles vision, Shruti handles audio and Sparsh gives a machine a sense of touch, so a robotic hand can tell the difference between gripping an egg and gripping a wrench. The action layer — Kriya, Prana and Karma — plans and executes movement such as reaching, gripping, walking and flying, and predicts what happens next. The third layer is the safety system, Kavach.
The company describes the platform as hardware-agnostic. It says the models run on industrial arms, humanoids, mobile robots, quadrupeds, cars and camera systems, and deploy through cloud APIs, on-premise installations or on-device processing, including offline, air-gapped environments. Inner Sky Labs frames this as sovereign Physical AI: a customer or government can run the models inside its own infrastructure and define its own safety rules, rather than depend on an outside provider’s cloud.
Kavach, and what it claims to solve
Kavach — Sanskrit for “shield” — is the most distinctive part of the offer. It acts as an independent safety check on every action a machine is about to take. Before a robotic arm swings across a factory floor, Kavach asks whether the space is clear of people, whether the planned speed suits the area and whether the action stays inside the zone the machine may work in. If any answer is no, the company says, the action does not proceed at all — not slowed down, not flagged afterward, but blocked before it starts. Inner Sky Labs says the system runs more than 20 guardrails.
That idea — a separate governor with veto power over motion — addresses a real problem in Physical AI. A model can understand a room and still make the wrong call, and a wrong call from a machine near a person can injure them. Giving safety its own layer, rather than leaving it to the model’s judgement, is a sensible architecture, and one that safety engineers would recognise.
The safety proof and the gap in it
Inner Sky Labs points to a strong proof point: an independent evaluation by BDO’s cyber security practice in which Miko’s safety system caught every unsafe prompt across seven public AI-safety benchmarks, ranking ahead of four widely used commercial models. The company says its Kavach safety layer reports benchmark scores above GPT-4o mini, Claude Haiku 3.5 and Gemini 1.5 Flash, and it has published methodology on its website.
This is a real result, and a company that has spent a decade filtering what a robot says to a child has genuine experience to draw on. But the claim deserves a careful reading, because it proves something narrower than the launch implies. Those benchmarks test language safety — whether a system refuses or flags an unsafe prompt. The new claim is physical safety: whether Kavach stops an unsafe robotic action before it happens. A system can be excellent at refusing a harmful sentence and still fail to stop a robotic arm.
The release bridges the two — “the same safety thinking that has spent a decade deciding what Miko may say to a child” — but the evaluation does not directly validate the physical claim, and no independent test yet shows Kavach stopping a dangerous motion on a third-party machine. The company’s own framing invites the inference; the evidence does not yet supply it.
The market it is entering
Physical AI is one of the most heavily funded categories in technology, and Inner Sky Labs enters it as a small player. NVIDIA supplies the open Isaac GR00T foundation model along with simulation tools and robot chips. Google DeepMind ships Gemini Robotics. Among independent platform companies — the path Inner Sky Labs follows, selling a brain rather than a body — Physical Intelligence has raised about $600 million at a $5.6 billion valuation and Skild AI about $1.4 billion at more than $14 billion. Figure, which builds its own humanoids, has raised over $1 billion at a $39 billion valuation.
Against those numbers, a ₹300 crore round — roughly $31 million — is modest, though it is large by Indian deep-tech standards. Inner Sky Labs argues that training foundation models has become too expensive for individual machine makers, particularly at low production volumes, so a shared provider makes sense. That is a coherent thesis, and the same one Skild AI and others are pitching. Those numbers set the scale of the challenge: to win customers, Inner Sky Labs must convince buyers that a smaller, newer model can do the job as well as one trained on far more data and compute. The competitive question is whether a company with a decade of consumer-robot data can match rivals training on far larger fleets and compute budgets.
The India angle
Inner Sky Labs leans on its Indian roots as a selling point. It was founded in 2015 by IIT Bombay alumni — Sneh Vaswani, who is chief executive, along with Prashant Iyengar and Chintan Raikar — and it built Miko in India for a global market. Its pitch to governments is that the platform lets any country define, in its own legal language, what its machines may do, rather than accept rules written elsewhere. That sovereign-AI framing lands at a time when many countries, India included, want critical technology to run on domestic infrastructure.
The company is not alone in India’s Physical AI ecosystem. Addverb develops industrial humanoids and Mowito builds models that help standard robotic arms learn tasks from human operators. But Inner Sky Labs is among the first to offer a full stack — perception, action and safety — as a licensable product, which makes it a test case for whether India can build platform software for robotics, not just robots.
The money and the business model
The ₹300 crore round involves existing backers YourNest Venture Capital and Keshav R Murugesh, a former NASSCOM chairperson, plus new global investors. It would take the company’s total funding to about $100 million, following earlier backing from Chiratae Ventures, IvyCap Ventures, 360 ONE and others. Inner Sky Labs says it will spend the new capital on real-world data collection, computing infrastructure and model training.
The business model is licensing: machine makers adopt individual models or the full platform, rather than build their own. Inner Sky Labs says commercial deployments began two to three quarters before the public launch and that it has completed four to five large deployments across enterprises and governments, including outside India. It has not named a single customer, sector or performance figure. For a platform company, those details are the evidence that matters most, and their absence is the biggest gap between the launch and a proven business.
What to watch
For robot makers and enterprise buyers, the milestone to watch is a named deployment with published performance on a machine Inner Sky Labs did not build. For safety teams, the question is whether Kavach’s guardrails can be independently tested on physical motions, not just language prompts — the evaluation the company has not yet shown. And, for investors, the test is whether a ₹300 crore round can keep pace in a category where peers raise ten to forty times as much. And for India’s deep-tech sector, Inner Sky Labs is a case study in whether platform software for Physical AI can be built and sold from Mumbai.
What the launch does and does not establish
It establishes a real, coherent product: a full stack for Physical AI, with a safety layer that reflects a decade of operating a robot near children, and a funding round that makes the company a serious Indian entrant in the category. It does not establish verified physical safety, a named customer, or performance on hardware the company did not build.
The idea at the centre — that machines need an independent safety layer with the power to say no — is a good one, and it is the part of the launch most worth taking seriously. The proof the company offers, though strong, speaks to a different problem than the one it now claims to solve. Closing that gap, with third-party results on other people’s machines, is the work that will decide whether Inner Sky Labs becomes a platform or stays a promising pitch.
Editor’s Note: This article draws on the company’s announcement and publicly available materials, plus independent reporting and published information on competitors. Inner Sky Labs‘ product, safety and deployment claims have not been independently tested or verified. No interview or demonstration was conducted for this article. See the source list in the publication package.