Josh Gruenstein
Related people
Episodes
Notes
a16z · The Robots Already Working in American Factories | Tutor…
-
What Work Do Tutor Intelligence Robots Do in Factories?
Josh Gruenstein says Tutor Intelligence builds two factory robots. Cassie is the heavy-duty machine: one arm, many suction cups, and foundation models in the 200 to 400 million parameter range, aimed at the handwork of moving pallets and cases. Sunny is a wheeled industrial semi-humanoid with a higher payload, more speed, and collaborative safety so it can work near people. A fleet of 100 Sunnies runs in Boston collecting data, and a growing fleet does two-handed work at customer sites.
Inside manufacturing and logistics, those two can probably do 80 to 90 percent of the work. The picture from How It's Made is the rare fully automated plant. Most million-square-foot facilities still have pe
-
What Is Data Factory One?
Josh Gruenstein built a 100 robot fleet called Data Factory One in Boston. To his knowledge, Data Factory One is the largest sort of semi-humanoid data collection farm in the United States. The research program is set up to get the most out of robot experience in the real world.
In language models, many different sources of data matter, but the top of the data is real world reinforcement learning traces of doing the thing a system should be able to do. Data Factory One is the research testbed for real world reinforcement learning at scale with robots.
Reinforcement learning is a technique family that does not really work at small batch sizes. A very large fleet out in the real world is wha
-
What Happens If a Factory-built House Gets 1% Cheaper?
Josh Gruenstein says everything gets better: goods that are valuable get bought in larger amounts when they are cheaper. A housing-factory worker in Wisconsin described houses that cost about 5 percent more to build in a factory than with traditional construction on site. If those houses were even 1 percent cheaper than a site-built house of the same quality, developers would move all of that demand into the factory.
Reclaiming materials from landfills falls under the same point. If Tutor is successful, the robots become a labor force other businesses can use to make the physical economy a cheaper place to live.
-
How Do Factory Workers Bond With Tutor Robots?
Josh Gruenstein says Tutor Intelligence gives every robot a name so the factory worker who spends the day with it can build a relationship with it. A factory manager treats a working robot as a piece of equipment. The buyer is the COO or the general manager, and the user is the factory worker. Tutor Intelligence does not automate 100% of a factory. It automates some jobs in a factory or a warehouse, so the worker and the robot still share the floor.
Human-robot interaction research supplies the practical tricks, and naming is one of them. On a repeat order, the customer names the new robots and gets a higher sense of ownership. People have named robots after their kids, a favorite football
-
How Much Does a Factory Robot Cost Per Hour?
Josh Gruenstein says Tutor Intelligence calls factories that have never bought a robot and offers one for $16 an hour. Almost every buyer is new to robots. The United States deploys only 38,000 robots a year, and 9 out of 10 U.S. factories and warehouses have none, so almost all of the market is greenfield. Buyers know the old guard of automation and assume a robot does not work for them. The opening pitch skips AI, models, and tutors.
The average time from the first call to a signed contract is 40 to 60 days. Robots reach the site and start doing work about 40 days after the contract. Production starts within a day, often within the first couple of hours on site. Each machine ships from th
-
How Do Human Tutors Train Factory Robots?
Josh Gruenstein says Tutor Intelligence trains factory robots by connecting them to human teleoperators, called tutors, who teach the robots new tasks. The internet already created a dataset for general-purpose models because online activity was recorded as it happened. Work in the physical world is not networked the same way. Someone can pick up a coffee cup and leave no record of fingertip pressures or poses.
The team set aside the search for an algorithm that would build intelligence from no data. Computer vision did not work that way, and language models did not either. A no-data start looked rigged against them. The alternative is a network of robots out in the real world, tied to tuto
-
Why Is Robotics a Systems Problem, Not Just an AI Problem?
Josh Gruenstein tells Martin Casado and Raghu Raghuram that infinite robot intelligence still leaves the problem unsolved, because the robots still have to be manufactured at scale. Eight billion people already have the physical intelligence robots need. Software that could connect any person to any robot would be unlimited robot intelligence, and manufacturing in the United States would still be a cottage industry. Robots are a rounding error even in manufacturing and logistics, the software does not exist, and the customers do not exist, because the robots cannot do the jobs customers want done.
A working platform has to put intelligence, hardware, and the customer together. Forward progr
-
How Many Products Have Tutor Intelligence Robots Helped Make?
Josh Gruenstein says Claude estimated that Tutor robots have helped manufacture and ship over a quarter of a billion products to American consumers, and he presumes most American consumers are a Tutor customer a couple of layers down. He started Tutor Intelligence about five and a half years ago so deploying robots into the physical economy would build the data sets, the models, and the supply chain.
The company is about 160 people, based primarily in Boston. It trains frontier robot AI models, manufactures robots at scale, and deploys them from small mom-and-pop shops to Fortune 50 and Fortune 100 businesses.
Robot intelligence and robot software were the bottleneck. For 60 years robots w