What iPhone Packaging Reveals About the Robotics Opportunity
Tutor Intelligence CEO Josh Gruenstein explains why robotics is a systems problem spanning AI, hardware, manufacturing, and customers, and how deploying robots into real factories and warehouses, with human tutors as backup, generates the data that makes them more capable. The discussion covers the company's own robot designs, its focus on small and mid-sized businesses, the economics of flexible robotic labor, and the implications for American manufacturing and supply chains.
Robotics is treated here as a systems problem, and the company’s strategy follows from that premise: you cannot make real progress on intelligence, hardware, supply chain, customers, and product as separate disciplines. Josh Gruenstein describes the bet behind Tutor Intelligence: deploy robots into real businesses now, let human operators (“tutors”) teleoperate them when needed, and treat the resulting work as the data that trains more capable models. The argument is that the flow of real-world data from paying customers is the durable asset, not any single architecture.
Why the Market Has Struggled
Gruenstein’s diagnosis is that most robot startups begin with a model or a data insight and then try to find a customer, while very few start from the question of what the robot is supposed to do. He argues that no robot startup has yet been massively economically successful, for two reasons: robotics lacks a core primitive that brings unit economics to a workable level, and the market is fragmented across many suppliers who must coordinate with each other. The result is that most companies start horizontal and then verticalize, buying equipment and building their own hardware, because the layers cannot be separated in practice. Whether this is a temporary feature of an immature market, like early computing, or something inherent to physical systems is an open question, but Gruenstein’s view is that the path forward is a full-stack platform proven out through working business units.
Tutors and the Data Flywheel
The name reflects the core thesis. The internet produced vast text and image datasets as a byproduct of human activity, but nobody has recorded the fingertip pressures and poses involved in picking up a coffee cup. Rather than wait for an algorithm that learns from nothing, Tutor connects deployed robots to human teleoperators who teach new tasks, and the work itself generates training data. Gruenstein stresses that this is positive-sum: customers get cheaper goods as the robots run, and the company accumulates data that is economically valuable by construction, since customers pay for the task being done. He contrasts this with strapping cameras onto workers in another country to collect data, where nothing guarantees the data matters.
Whether tutors eventually disappear is treated as a unit-economics question rather than a philosophical one. The comparison is to the cost of serving a model on an H200 accelerator, with duty cycle and utilization. Human intelligence is worth paying for where it supplies a heterogeneous capability that models do not yet have, and the open question is how much engineering it takes to get the last few nines of reliability versus using a tutor. Martin Casado raises the concern that physical tasks may have a high-dimensional solution space, where the final fraction of cases keeps requiring humans. Gruenstein’s answer is pragmatic: the company is not purist about data sources and uses simulation and internet-scale data to widen the distribution alongside real-world robot traces.
Frontier Models Accelerating Robotics
Gruenstein says the arrival of frontier models has changed the picture. Large language models have absorbed a great deal of general knowledge about the world, but until recently they lacked the spatial reasoning needed to connect that knowledge to action heads. Labs now post-train on 3D spatial data such as CAD and modeling data, driven by real commercial demand, and that is precisely the spatial reasoning Tutor’s models need. The company began in 2021 with convolutional networks and has gone through each generation since. Its view is that architectural tricks are worth using when they accelerate the flywheel, but the lasting value sits in the data flow and the loops themselves.
Hardware: Cassie and Sunny
Because the United States deploys only about 38,000 industrial robots a year and most factories and warehouses have none, Gruenstein argues the existing supply chain was not designed for the trade-offs that frontier models require. Tutor therefore designs and manufactures its own robots in Watertown, Massachusetts, while sourcing components such as actuators and cameras from OEMs operating at larger scale.
Cassie is built for heavy palletizing and case handling: a single arm with suction cups, powered by foundation models, designed for work that remains stubbornly manual in logistics, where turnover is high and the average worker is near retirement age. Sunny is a wheeled semi-humanoid meant to run for years in a factory at human rates, with higher payload, faster motion, and collaborative safety standards. Wheels rather than legs suit polished indoor floors, though the team acknowledges that gravel sites in construction or landfill material recovery will eventually require legged embodiment. The home is not the near-term target, since the economics and the safety requirements are far harder there. About 100 Sunny units run in Boston as the data factory, forming the company’s main research test bed for real-world reinforcement learning.
Selling to First-Time Buyers
Almost every customer buys a robot for the first time, and most of the market is greenfield. Gruenstein says the pitch does not lead with AI or models; it leads with a price of roughly $16 an hour. The company reports an average of 40 to 60 days from first call to signed contract and about 40 days from contract to a robot doing productive work on site, with some units running within hours. The business model shifts risk away from the customer, which is what lets distribution move much faster than traditional automation.
The company deliberately serves small and mid-sized businesses, which make up most of the US manufacturing and logistics base, rather than chasing a few large Fortune companies with bespoke pilots. Gruenstein’s critique of robotics startups is that they often behave like custom integrators rather than product companies. A diverse customer base forces the product into the shape of a general robot body running general models, and a successful deployment in two weeks with a mid-market firm is evidence the platform works.
Adoption inside the plant depends on a distinction between buyer and user. The COO or general manager buys the robot, while line workers operate alongside it. Tutor relies on human-robot interaction research to build that relationship, and one practical trick is naming: each robot gets a name, and repeat customers are allowed to name new units. Names have ranged from children’s names to favorite athletes and musicians. Gruenstein’s broader point is that workers who are no longer doing backbreaking work can move into higher-value positions, and that a robot which works quickly becomes equipment in the eyes of management.
Economics and the Shape of the Market
Gruenstein frames value simply as the amount of work a robot can do divided by its cost, and he argues that this applies to the whole physical labor market. Surprisingly for the industry, operating cost is now a larger share of unit economics than upfront capital expense, with inference and teleoperation dominating. Customers are not making a one-time purchase so much as securing a long-term partner, because a deodorant manufacturer should be an expert in deodorant and not in robots.
He sees the broader effect as a cheapening of goods and the creation of many more factories. Infinite flexible labor would let anyone start or scale a physical business, much as cloud software lowered the barrier to digital ones. He gives the example of prefabricated housing, which costs roughly 5 percent more to build in a factory than on site; a 1 percent advantage over traditional construction would redirect developer demand toward factory methods.
Supply Chain and Geography
Gruenstein argues that the global supply chain is highly interdependent, and that US and Chinese companies are both building industrial capacity under long-term strategies that include cross-border investment. He sees an opportunity for the United States: China invested heavily in the previous generation of fixed automation, which may prove costly to unwind, while the United States is building greenfield capacity on generally capable robots. As Tutor becomes a major buyer of components, it expects to verticalize further, investing in US robot componentry because a mature domestic supply chain is necessary for the company to succeed.
Research and Team
Tutor’s research program treats real-world reinforcement learning as the highest-quality signal, analogous to what matters most in large language models. Reinforcement learning needs large batch sizes that small fleets cannot provide, so the 100-robot data factory is designed to make those methods practical in the physical world. The team is interdisciplinary across software, hardware, materials, and electronics, because there is no layer where a company can abdicate responsibility, and it spans roles from model architecture to factory-floor debugging. The company is hiring researchers, engineers, and technicians across the United States, with go-to-market staff and technicians spread widely because the customers are spread widely.