He took down the elevators in his childhood apartment building while welding in the basement. Two decades later he runs Tutor Intelligence, an MIT spinout teaching robots to earn their keep on the factory floor.
Most robotics companies sell you a machine, hand you a manual, and wish you luck. Josh Gruenstein decided that was the wrong business entirely. Tutor Intelligence, the company he co-founded out of MIT's Computer Science and Artificial Intelligence Laboratory in 2021, does not sell robots. It rents them by the hour. No capital expense, no programming, no integration consultants camped out for six months. A Tutor robot shows up, plugs into a packing or palletizing line, and starts working alongside the people already there.
In December 2025 that bet got a $34 million vote of confidence. Union Square Ventures led the Series A, Fundomo co-led, and seed backer Neo came along for more. Total funding to date sits at roughly $42 million. The money is pointed at one stubborn idea Gruenstein has been chasing since graduate school: that the thing holding robots back was never the cleverness of the algorithm. It was the absence of data.
Here is the observation that became a company. Computer vision had ImageNet. Language models had the entire internet. Robots, meanwhile, had almost nothing collected from real physical environments, no equivalent corpus of what it actually feels like to grip a dented box or recover from a missed pick. Gruenstein and his co-founder Alon Kosowsky-Sachs looked at that gap and saw a flywheel rather than a wall. Put robots into real factories, capture what they do, feed it back into the models, and the robots get better, which lets you deploy more robots, which produces more data. As Gruenstein puts it: "more learning unlocks more robots, unlocks more data, so on and so forth."
Running a robot company, it turns out, is not mostly a software job. Tutor's team is a strange coalition under one roof: field technicians, machine learning researchers, salespeople, operations staff, and data labelers who turn raw factory footage into training signal. "One thing that is unique about our business is the breadth of functional skill sets necessary to deliver and operate robots," Gruenstein has said. The robots-as-a-service model is what holds that coalition together, because Tutor stays on the hook for uptime instead of disappearing after the sale.
The robots themselves are deliberately unglamorous. They palletize. They pick cases. They handle the high-mix, low-volume work that has resisted automation precisely because every order looks a little different and traditional robots need every motion spelled out in advance. Tutor's pitch is that its machines bring "out-of-the-box performance" and learn the rest on the job, with a cloud platform watching the whole fleet and feeding improvements back down.
"More learning unlocks more robots, unlocks more data, so on and so forth."
The founding myth of Tutor Intelligence is refreshingly literal. Gruenstein describes Kosowsky-Sachs as "the guy with the DIY Segway robot." His own contribution to the lore is a confession: he knocked out the elevators in his childhood apartment building while welding in the basement. These are not the polished origin anecdotes of a pitch deck. They are the tells of people who have been taking things apart, and occasionally breaking the building, since well before anyone was paying them to.
Before he was old enough to enroll at MIT, Gruenstein wrote Fido, a lightweight C++ machine learning library for embedded electronics and robotics. It is still on GitHub, still collecting stars, north of 460 of them. It is a small artifact, but a revealing one: a teenager deciding that the right way to make a robot smarter was to write the learning code himself.
At MIT he collected the credentials almost as a side effect. An SB in electrical engineering and computer science, then an MEng in artificial intelligence. Along the way he did quantitative finance at Bridgewater Associates, deep learning acceleration at Intel's programmable solutions group, and recommendation systems at Cornell Tech. He ran Battlecode, MIT's famous student programming competition, and later taught the graduate course on robot learning, 6.884, while doing research in the Improbable AI lab.
In 2020 he joined the MIT lab where Kosowsky-Sachs was already working on robot AI through deep learning. The two of them kept circling the same conviction. A year later, Tutor Intelligence existed, with a mission stated plainly enough to fit on a sticker: put a robot in every factory.
Writes Fido, a C++ machine learning library for embedded robotics, which goes on to collect 462+ GitHub stars.
Earns an SB in EECS and an MEng in AI. Stints in quant finance at Bridgewater, deep learning at Intel, and recommendation systems at Cornell Tech.
Runs Battlecode, MIT's student programming competition, and teaches the graduate robot learning course (6.884).
Joins an MIT lab developing robot AI through deep learning, working alongside future co-founder Alon Kosowsky-Sachs. Publishes robot learning research.
Co-founds Tutor Intelligence out of MIT CSAIL. Mission: a robot in every factory.
Tutor raises a $34M Series A led by Union Square Ventures, co-led by Fundomo, with seed investor Neo. Total funding reaches roughly $42M.
Vision and language had oceans of training data. Robots had a puddle. Tutor exists to fill it, one real factory shift at a time.
Robots-as-a-Service means no capital expense and no programming. Tutor keeps the risk and the responsibility for uptime.
Every deployed robot feeds the cloud platform. More robots make better robots, which justify more robots. The flywheel turns.
Warehouse and factory automation is a crowded, expensive, often disappointing field. Buyers have been burned by systems that promised flexibility and delivered a rigid arm that needed a specialist every time a product line changed. So the more interesting question about Tutor's Series A is not the size of the number. It is what Union Square Ventures, Fundomo, and Neo think they are buying.
They are buying a data engine disguised as a robot company. Every hour a Tutor robot spends palletizing boxes or picking cases in a real facility is an hour of labeled experience that a competitor selling hardware outright never gets to keep. The robots-as-a-service model is the mechanism that makes that loop possible: because Tutor owns the fleet and charges for usage, it also owns the footage, the failures, and the recoveries. The business model and the research strategy are the same decision.
That alignment is rare, and it is the part Gruenstein keeps returning to. Selling a robot ends the relationship at the loading dock. Renting one keeps Tutor present, accountable, and learning. The flywheel he describes only spins if the company stays in the room.
The work itself is the unsexy, high-mix, low-volume corner of manufacturing and logistics, the kind of operation where every order is a little different and traditional automation gives up. Tutor leans into exactly that, pitching machines that arrive with usable out-of-the-box performance and then sharpen on the job, watched over by a cloud platform that pushes improvements across the whole fleet.
It is a deeply American bet, too, aimed at factories across North America at a moment when reshoring and labor shortages have made the economics of a flexible robot worker suddenly legible to operators who never thought they could afford automation. Gruenstein's framing is not that robots replace the people on the line. It is that they work alongside them, human-machine collaboration as a feature rather than a slogan.
Whether the thesis holds is, of course, unproven at scale. But the shape of the wager is clear, and unusually honest: build robots that are good enough to deploy today, keep them on a leash so you learn from every shift, and let the fleet teach itself toward something better than any single robot could become alone.
"The guy with the DIY Segway robot." The two met at MIT and were both working on robot AI through deep learning when they decided to build Tutor together.
Led the $34M Series A in December 2025. Fundomo co-led; seed backer Neo returned for the round.
Two degrees, a teaching post in robot learning, research in the Improbable AI lab, and the lab where Tutor was born.
Quant finance, deep learning acceleration, and recommendation systems. A résumé built to span the whole stack a robot company needs.
The co-founders' calling cards: one built a DIY Segway robot, the other disabled an apartment building's elevators mid-weld. Hardware people, through and through.
A machine learning library for robots, written in C++ before he started college, still pulling hundreds of GitHub stars years later.
He taught MIT's graduate robot learning course, then walked across the river to find out whether the syllabus survives contact with a real packing line.
Technicians, researchers, salespeople and data labelers on the same team. The org chart is the strategy.
The goal is not a smarter robot in a lab. It is a robot that earns its keep in every factory, learning at fleet scale until it operates with superhuman intuition and flexibility.
Where to follow the work, the code, and the company.