Tutor Intelligence CEO Josh Gruenstein on Building Robots for Real Work

Tutor Intelligence co-founder and CEO Josh Gruenstein argues that putting useful robots in factories requires a system spanning models, hardware, manufacturing, customer deployment and the work itself. He discusses the company with a16z general partners Martin Casado and Raghu Raghuram.

Gruenstein says he founded Tutor around the idea that deploying robots in the physical economy could generate the data, models and supply chain needed to make them more useful. The company has about 160 people, primarily in Boston, and builds models, manufactures robots and deploys them.

Tutor’s robots operate at businesses ranging from small firms to large corporations, he says. He cites an AI-generated estimate that they have helped manufacture and ship more than a quarter of a billion products to American consumers; he presents it as an estimate, not an audited count.

The founders saw robot intelligence and software as bottlenecks for machines that had historically required extensive custom engineering. Deploying robots could both deliver useful work and produce labeled data for improving models.

The hosts observe that many robotics pitches start with a model rather than a job. Gruenstein describes the reverse thought experiment: even if any person could provide a robot with unlimited intelligence remotely, the problem would not be solved.

A company would still need to manufacture machines at scale, build software and find customers whose jobs the machines can perform. Gruenstein sees robotics as a systems problem that requires working across these layers together.

His technical background led him to think about manufacturing, hardware and electronics as well as models. Starting a company expanded the stack further to include supply chains, customers and product design; avoiding a layer also removes opportunities to make useful trade-offs.

He attributes previous robotics companies’ difficulties partly to insufficient intelligence and partly to a fragmented market. The hosts ask whether vertical integration is inherent to physical work or simply a feature of an immature industry.

Gruenstein favors the maturity explanation. He compares the goal to computing platforms that let software businesses operate without designing chips: robotics needs general platforms, but those platforms must be proven by successful businesses built on them.

A host challenges the analogy, asking whether physical environments have the same clean abstractions as computing. Gruenstein suggests a humanoid can be a general-purpose form factor, while different machines may still be appropriate for different jobs, just as computing uses CPUs, GPUs and specialized chips.

The name Tutor reflects a data problem. Human activity on the internet naturally creates text and other material used for training, but everyday physical actions do not automatically record the forces, poses and other information a robot would need.

Tutor’s approach is to deploy robots and connect them with human remote operators, or tutors, who help teach new tasks. The business aims to collect training data while the machines perform economically useful work rather than treating data collection as a standalone expense.

One host says customers they spoke to often did not notice when a human tutor was involved. Gruenstein sees that as a sign that buyers care about whether work gets done and whether their costs improve, rather than the internal division between model and operator.

Asked whether the goal is to eliminate tutors, he frames it as an operational and economic trade-off: how much engineering is needed to improve the model further, and what does human assistance cost at the utilization level required?

A host questions whether robot training data will cover economically useful tasks as neatly as internet text covers many language applications. The last portion of useful work may remain outside the training distribution and require humans for a long time.

Gruenstein agrees the source of data matters. He argues that useful commercial work can align revenue with collecting valuable examples, whereas collecting physical demonstrations without a connection to demand does not guarantee the data will be useful.

He is not committed to a single data source. If the long tail remains difficult, Tutor can use simulation and internet-scale data to widen its training distribution as well as collect real-world robot experience.

Gruenstein says frontier models are becoming more useful for robotics as labs post-train on three-dimensional material such as CAD data. In his view, better spatial reasoning helps connect general knowledge in large models to robot actions.

He sees stronger models as an accelerator rather than a replacement for the company’s work. Tutor has already lived through several model changes; he regards the continuing loop of deployment, data collection, manufacturing scale and new tasks as the durable asset.

Asked what building a robot means in practice, he says the existing supply chain is too small and often optimized for fast repetitive industrial systems rather than machines co-designed with frontier models. Tutor therefore designs and manufactures its own robot bodies.

Tutor still buys components such as actuators and cameras from larger suppliers. It builds the mobile machines and arms that perform customer tasks in Watertown, Massachusetts, according to Gruenstein.

He describes Cassie as a heavy-duty single-arm machine with suction cups, using models in roughly the 200 million to 400 million parameter range. Its design targets widespread pallet and case handling in manufacturing and logistics.

The second machine, Sunny, is a wheeled industrial semi-humanoid built for factory reliability, higher payloads and collaborative safety. Gruenstein says 100 operate in a Boston data factory and a growing number work at customer sites on more dexterous two-handed tasks.

Gruenstein estimates that Cassie and Sunny could eventually address a large share of manufacturing and logistics work. He contrasts popular images of fully automated factories with facilities where people still work by hand at conveyors and folding tables.

He says difficult-to-fill manual jobs include moving heavy cases and assembling or packaging products, including high-volume consumer goods. The hosts ask what the current machines cannot do.

The current robots are designed for industrial interiors and cannot navigate terrain such as gravel, Gruenstein says. Construction and landfill recovery may require different, potentially legged machines; the industrial designs are also not suited to homes today.

Most prospective customers have never bought a robot, he says. Rather than leading with models or AI terminology, Tutor introduces a starting price of about $16 per hour and explains the operating proposition.

Gruenstein reports an average of 40 to 60 days from first call to signed contract, about 40 more days to delivery, and production setup within a day, often hours. He attributes the speed to standardized robots and a model that avoids requiring customers to take an upfront equipment risk.

Once a machine proves it can do a job, he says, operations teams often identify other sites where it could be useful. A factory manager may see equipment, but the buyer and the worker who interacts with the machine are different people.

Tutor draws on human–robot interaction research and gives each machine a name. Manufacturing technicians usually name the initial robots; customers placing repeat orders can name new units, which Gruenstein says helps workers develop a sense of ownership.

He reports customers naming robots after children, athletes and musicians. He says workers value reduced physical strain and managers value operational help, presenting these as the company’s observations about its deployments.

The hosts ask why many customers are small and midsize businesses when robotics startups often target very large companies. Gruenstein says these firms represent much of manufacturing and logistics and that a diverse customer base forces the product to be broadly useful.

He criticizes the pattern of spending years building a custom system for a single large customer: the product may fail to work, or the customer’s needs may change. Tutor instead wants a general platform that can be deployed quickly across different businesses.

On the potential economic effects, he argues that cheaper production could expand demand or cross thresholds that make new business models viable. He cites a housing-factory example in which a small cost advantage over on-site construction might change developer demand.

A host asks about home use. Gruenstein says the company is open to it, but homes pose different technical, safety and economic constraints. Starting in factories lets the company make useful compromises while building a platform that may become more capable.

Tutor tries to make hardware iteration resemble software iteration. Gruenstein says owning manufacturing allows changes roughly every two weeks, while software ships daily or in larger two-week releases.

Product staff sit close to customer success and implementation, while researchers remain connected to customer problems. A modeling advance can therefore inform a product redesign and reach the field more quickly, he says.

Potential customers can contact the company and use pricing calculators to estimate savings. Gruenstein says the offer is intended to produce near-term operational savings without requiring the customer to make a capital purchase upfront.

He defines a robot’s value as the work it can perform divided by its cost. Both capital and operating expenses matter, but he says ongoing costs such as inference and remote operation currently make up the larger share for Tutor.

Customers often view automation as a long-term strategic need, he says: they face an aging workforce, rising demand and tight margins. They want to concentrate on making their own products while relying on a robotics partner.

Asked how general-purpose robots might affect factory locations and supply chains, Gruenstein predicts broad changes and argues that new flexible automation could affect the relative positions of the United States and China. These are his expectations, not established outcomes.

He emphasizes that the 2026 robotics supply chain remains globally interconnected and interdependent. At higher volumes, relationships with component suppliers change, raising decisions about contract manufacturing versus bringing production in-house.

Gruenstein says Tutor wants to help develop a more mature US robotics supply chain and expects to make more components domestically over time. He points to the company’s robot manufacturing in Massachusetts.

On research, he argues that real-world reinforcement-learning traces are valuable because they capture the task the robot is meant to perform. Tutor has built a 100-robot research fleet in Boston, called Data Factory 1, to investigate reinforcement learning at scale.

He says larger fleets allow experiments that are difficult with small batches, then describes hiring across research, software, hardware, materials, electronics and manufacturing. The team can move between model design and observing robots on factory lines.

The hosts close by asking where the team works. Gruenstein says it is spread across the United States, including customer-facing staff and technicians located near deployments.