For most of the last three years, the AI boom has lived on screens: chatbots drafting emails, copilots writing code, models generating images on demand. The next frontier is messier and far harder — getting that same intelligence to reliably pick up a bolt, seat a connector, or weld a seam on a real factory line. That is the promise of “physical AI,” and this week an Indian startup put a small but pointed marker down on it.
Bengaluru- and Detroit-based Mowito has raised a $3 million pre-seed round to scale foundation models that teach standard industrial robot arms to learn tasks by watching, rather than by being painstakingly programmed. The round, reported by Inc42 and confirmed across multiple outlets, was led by Version One Ventures — and, notably, drew a personal cheque from PyTorch co-creator Soumith Chintala. For a two-year-old hard-tech company, that is a meaningful signal of technical validation.
The raise
According to reporting on the deal, Mowito’s $3 million pre-seed was led by Version One Ventures, with participation from All In Capital, Unisol and iSeed. The angel roster is where it gets interesting for anyone tracking the AI stack: it includes Soumith Chintala, the co-creator of PyTorch — the deep-learning framework that underpins a huge share of the world’s AI research and production — who is now CTO of Mira Murati’s Thinking Machines Lab. Also on the cap table are Adarsh Kulkarni of Foundry Robotics, Ashish Kulkarni of Coformer.ai, and Vaibhav Domkundwar of Better Capital.
Founded in 2024 by Puru Rastogi, Adityanag Nagesh and Safar V, Mowito says the fresh capital will fund an expansion of its US presence, strengthen its engineering and go-to-market teams, and scale deployments across automotive and electronics manufacturers.
Why does one angel cheque matter so much in a $3 million round? Because in deep tech, capital is not the scarce resource — credible technical judgment is. When the person who helped build the tooling that most modern AI runs on decides to put money into an industrial-robotics startup out of India, it is less about the dollar amount and more about the diligence it implies. It tells other investors, prospective hires and enterprise customers that someone who understands the hard parts looked closely and liked what he saw.
Why physical AI now
“Physical AI” is the shorthand for applying modern machine learning to perception and control in the real world — robots and machines that sense their surroundings and act on them, rather than models that only produce text or pixels. It sits alongside terms like embodied AI and world models, and it has become one of the hottest themes in the field precisely because the software side is maturing while the physical side remains stubbornly unsolved.
Mowito’s specific bet is a good illustration of what the category actually means in practice. Rather than building robots, the company builds the foundation models that run standard industrial robot arms. The pitch is that a robot can learn a manufacturing task by observing an operator perform it — in some cases from as few as a single demonstration — and then repeat it with the precision an assembly line demands, without the weeks of conventional programming that traditional industrial automation requires.
That framing matters. Classical factory robots are extraordinarily precise but brittle: they do exactly what they are coded to do, and re-tooling a line for a new part or a new task is slow and expensive. The physical-AI thesis is that learned models can collapse that setup time and let the same hardware flex across many tasks — the difference between programming a machine and teaching it. Mowito claims its models are already live, with robots running on manufacturing lines at a Fortune 500 automotive company and at one of the world’s largest electronics contract manufacturers.
The hard parts
None of this is easy, and it is worth being clear-eyed about why physical AI has lagged its screen-bound cousins.
- Sim-to-real transfer. Much robot learning happens in simulation because real-world data is slow and costly to collect. But the gap between a tidy simulator and a noisy factory — friction, lighting, worn parts, vibration — is where models tend to break. Closing that gap reliably is one of the central unsolved problems of the field.
- Safety and reliability. A language model that hallucinates produces a bad sentence. A robot arm that misjudges force or position damages a part, a machine, or a person. On a real assembly line the tolerance for error is close to zero, which raises the bar for validation far above what consumer AI demands.
- Hardware, data and time. Embodied AI couples software to physical machines, sensors and long deployment cycles. Demonstration data has to be gathered on real hardware, iteration loops are slower than in pure software, and enterprise manufacturing customers move deliberately. These are long-horizon businesses, not overnight scale-ups.
- Competing globally. Well-funded robotics and embodied-AI efforts are underway in the US, China and Europe, several with far larger war chests. An Indian pre-seed company is playing in a genuinely global arena.
The flip side of every one of those difficulties is a moat. Because the hard parts are hard, a team that actually gets learned control working reliably on real factory hardware — and can prove it with live deployments — has something that is difficult to copy and expensive to catch up to.
The India read
Mowito is a small round, and it would be a mistake to over-read a single pre-seed. But the shape of the bet is worth noting for what it says about where Indian deep tech might compete.
India’s software and IT-services strength is well established. Physical AI points at a different, arguably more defensible opportunity: embodied intelligence for manufacturing, logistics and services, sectors where the country has both scale and a growing appetite to move up the value chain. A model layer that makes existing industrial hardware smarter is a natural fit for an economy trying to expand its manufacturing base — and for founders who would rather sell software-style margins into factories than build robots from scratch.
It also plays to a specific kind of talent. Building foundation models for robot control is squarely hard-tech work — the sort that rewards deep research capability over growth-hacking. That a company doing it from Bengaluru could attract a lead institutional investor and an angel of Chintala’s technical stature suggests the talent-and-credibility flywheel for Indian deep tech is starting to turn, even if the checks are still early-stage.
Where can India actually win in physical AI? Likely not by out-spending the largest global labs on general-purpose humanoids, but by going deep on specific, high-value industrial problems — precise manipulation on real production lines, learned automation that slots into existing factories — and by pairing world-class model talent with proximity to manufacturing demand. Mowito is one data point, not a trend. But it is the kind of data point worth watching, because it is exactly where the embodied-AI wave and India’s industrial ambitions overlap.
This report is based on funding details disclosed by the company and reported by Inc42, The AI Insider and other outlets on July 7, 2026. Figures are attributed to those sources.
