The Robots Already Working in American Factories | Tutor Intelligence

Josh Gruenstein says he co-founded Tutor about five and a half years ago around a simple thesis: deploy robots into the physical economy first, then use that deployment to build the data, models, and supply chain needed to spread robotics. The Boston-based company has about 160 employees, trains frontier robot-AI models, manufactures robots, and deploys them directly into major supply-chain businesses. Robots now operate across the United States, from small businesses to Fortune 50 and Fortune 100 companies, and have helped manufacture and ship more than a quarter-billion products to American consumers.

Josh argues that robotics cannot be solved by focusing only on intelligence, hardware, or customers. His thought experiment starts with eight billion humans who possess baseline physical intelligence: if software could connect any person to any robot, the intelligence problem would look solved, but robots would still need to be manufactured, supplied, sold, and used in real jobs. Robotics therefore requires trade-offs across models, hardware, software, manufacturing, supply chain, product design, and go-to-market. He says earlier roboticists focused on custom machines, while starting a company taught him that the entire business stack must be integrated.

The hosts observe that robot companies often begin by presenting a general model, then end up buying manufacturing equipment and specializing. Josh attributes this mainly to market immaturity, comparing robotics with early computing, when specialized systems preceded general platforms. He believes humanoid or similar general-purpose forms can eventually provide the physical equivalent of a CPU: not perfectly optimized for every task, but general enough to support many businesses. As markets mature, specialization may return. Tutor’s goal is to build a platform that lets ordinary companies use robotics without becoming robotics companies themselves.

The name Tutor reflects Josh’s central data strategy. Human physical work is not automatically recorded the way online writing, email, and social media create training data for language models. Tutor therefore connects robots to human teleoperators—“tutors”—who teach them new tasks. The robots generate economic value while operating in customer facilities, and the resulting real-world work becomes training data. Josh says this is preferable to collecting arbitrary demonstrations, because a robot performing a task for a paying customer proves that the data is economically relevant. Customers often do not notice when a tutor is involved, which Josh treats as evidence that the system is working.

The hosts ask whether Tutor’s long-term goal is to eliminate human tutors. Josh frames the answer as a unit-economics decision rather than an ideological commitment. The company compares the cost of human intelligence with the cost of inference on hardware such as an Nvidia H-200, considering duty cycle, utilization, and the specific capabilities models still lack. He accepts that humans may handle the final difficult portion of a high-dimensional task for some time. Tutor also uses simulation and internet-scale data to broaden its training distribution, rather than relying exclusively on physical robot traces.

Josh describes a recent change in robotics models: frontier labs are post-training on 3D data such as CAD and 3D modeling because there is a commercial market for those capabilities. That work supplies the spatial reasoning needed to connect general model knowledge to robot actions. Tutor began in 2021 using convolutional neural networks and has lived through several architecture shifts, but Josh says the durable advantage is the flow of data. Architectural improvements matter when they help deploy more robots, collect more useful experience, manufacture at greater scale, or expand the task range.

Tutor builds two robot types. Cassie is a heavy-duty, wheeled robot with one arm and suction cups, designed for pallet and box handling and powered by foundation models in roughly the 200-million-to-400-million-parameter range. Sunny is a wheeled semi-humanoid designed to run for three to ten years at human rates and reliability, with higher payload capacity, greater speed, bilateral manipulation, and collaborative industrial safety. Josh says these robots address the largely unseen work below highly automated “How It’s Made” factories: people standing at conveyors or folding tables, handling boxes, assembling products, and doing backbreaking work.

Tutor operates 100 Sunny robots in Boston in what it calls Data Factory One, which Josh describes as the largest semi-humanoid robot data-collection operation in the United States to their knowledge. The fleet provides a test bed for real-world reinforcement learning, a family of methods that generally needs large batch sizes and therefore cannot be explored effectively with only a few robots. Tutor uses the factory for internal data collection and research while also deploying Sunnies into outside facilities. Josh says the company is hiring across model architecture, software, hardware, materials, electronics, manufacturing, and robotics technicians.

Because most U.S. factories and warehouses have never purchased a robot, Tutor avoids leading with models, tutors, or artificial intelligence. Its sales proposition is effectively, “Have you ever considered a robot for 16 bucks an hour?” The average time from first call to signed contract is 40–60 days, and from contract to installation is about 40 days. Robots can be stood up in production within a day, sometimes within hours, because Tutor ships a general product rather than a custom automation project and does not require customer CapEx. Customers range from local businesses to large enterprises.

Factory managers generally treat a working robot as equipment—“a robot is a robot unless it works, in which case it’s a machine”—but line workers interact with it daily, so Tutor uses human-robot-interaction practices. Every robot has a name; technicians name new units, while repeat customers may name their own, sometimes after children, athletes, or pop stars. Josh says robots make jobs less backbreaking and can elevate workers into higher roles. He expects robotics to lower the cost of goods and create discontinuities such as factory-built housing becoming cheaper than on-site construction, while general-purpose labor could let more people start physical businesses.

Josh says the United States deployed only about 38,000 new industrial robot workcells in 2025, excluding older installed machines, versus roughly a million in China. Existing suppliers such as Fanuc and ABB mainly serve systems integrators with fast, repetitive machines, not robots co-designed for frontier models and flexible work. Tutor therefore manufactures its robot bodies in Watertown, Massachusetts, while sourcing some components such as cameras and actuators. As Tutor becomes a major buyer, some suppliers increasingly resemble contract manufacturers, pushing the company toward further verticalization. Josh calls building U.S. capacity both an obligation and an opportunity in an interconnected 2026 supply chain.