Tutor Intelligence CTO Alon Kosowsky-Sachs and CEO Josh Gruenstein with Alla Simoneau from the AWS Generative AI Innovation Center, at Tutor’s Digital Factory. Credit: Eugene Demaitre
To make robots more flexible and capable, they must learn the way people do—but on a massive scale, says Tutor Intelligence Inc. The company has created DF1, its “kindergarten” or Data Factory, which houses 100 bimanual robotic arms. Together with remote teleoperators, or “tutors,” these robots are used to train the company’s Ti0 vision-language-action (VLA) model.
“I’ve been constructing robots since I was 9 years old,” shared Josh Gruenstein, co-founder and CEO of Tutor Intelligence. “Unlike large language models and the Internet, there’s no robot equivalent of Wikipedia. We need to transfer knowledge from 8 billion people at scale. DF1 isn’t focused on building models—it’s about capturing the right data from humans teaching robots.”
In 2021, Gruenstein launched the company alongside Alon Kosowsky-Sachs, who now serves as chief technology officer, from MIT’s Computer Science and Artificial Intelligence Laboratory (MIT-CSAIL). In December 2025, the firm secured $34 million in Series A funding and recently hosted an open house event.
Tutor Intelligence was among the first group of companies in the Physical AI Fellowship, gaining mentorship and backing from Amazon Web Services, NVIDIA, and MassRobotics. It has maintained its collaboration with AWS for large-scale cloud computing and with NVIDIA CUDA for model development.
Editor’s note: At the 2026 Robotics Summit & Expo on May 27 and 28 in Boston, MassRobotics will feature startups from the second cohort of the Physical AI Fellowship. Registration is now open.
Data Factory uses human guidance to train physical AI
“Our goal is to develop broadly capable robot AI in collaboration with the industrial sector,” said Gruenstein. Tutor aims to create a “virtuous cycle” where robots are deployed commercially to gather data at scale and keep improving through training.
Instead of relying primarily on simulation, Tutor Intelligence built DF1 to collect real-world data at its new headquarters in a historic mill in Watertown, Mass. The startup claims that, as far as it knows, DF1 is the largest “robotic data factory” in the United States.
The 100 Sonny semi-humanoid robots began with piece-picking tasks typical in e-commerce, kitting, and other commercial settings. Like a classroom of kindergartners, they were initially clumsy with consumer packaged goods such as sponges or bags of snacks.
Within just a few weeks, though, the robots have already proven the effectiveness of Tutor’s approach to gathering real-world data, overseen by 45 to 50 tutors in Mexico and the Philippines using proprioceptive teleoperation, as well as on-site staff. They apply a similar scoring system to the one Gruenstein developed as a high school student over ten years ago.
“By testing the same policy across all 100 robots, we can identify and fix robot behaviors 100 times faster,” said Tutor Intelligence. “An edge case that might normally take eight hours of robot operation to spot becomes apparent in just five minutes of DF1 operation.”
Gruenstein compared DF1 to the Large Hadron Collider, calling it “an instrument of discovery and scientific exploration for scaling humanoid robots.”
AWS has assigned a team of technical strategists and scientists to assist startups like Tutor Intelligence, explained Alla Simoneau, physical AI technology lead at the AWS Generative AI Innovation Center.
“We provide support from initial concept through to production,” she told The Robot Report. “The Physical AI Fellowship stands out because our experts help the startup community bring products to market safely and securely.”
To minimize inconsistencies in fleet-scale data for “industrial intelligence,” Ti0 is trained using “velocity normalization,” which Tutor described as “a new preprocessing technique that synchronizes the speed patterns of demonstrations from different teleoperators.”
“Through DF1 and Ti0, our aim is to launch the world’s first commercial humanoid deployment flywheel on the Sonny platform, enabling continuous policy improvement and growing real-world value over time,” the company stated in a blog post.
Gruenstein expressed confidence that Tutor can transition from training to pilot programs by the end of this year. To accelerate scaling, the company has used the same collaborative robot hardware, sensors, and contract manufacturers for Sonny as it did for its single-armed Cassie system.
Cassie is built for rapid deployment
As Tutor Intelligence showed at MODEX 2026 last month, the newest version of Cassie can be up and running in just two days and manages boxes weighing up to 50 lb. (22.6 kg). Thanks to years of ongoing data collection and training, the case-picking and palletizing robot can process a continuous stream of SKUs with minimal changeover time and handle up to 14 cases per minute.
Cassie uses vision technology and connects with external sensors through the EVE native I/O module. Its safety features include lidar-defined zones and compliance with ISO/TS 15066 standards. The mobile manipulator can travel between workstations.
The system’s tablet interface supports multiple languages and doesn’t require programming knowledge or the configuration of waypoints, according to Tutor. The robot can also be made washdown-ready for food-handling environments.
Since Sonny and Cassie share components, will Tutor’s bimanual picking eventually merge with its mobile manipulator?
“Right now, we’re focused on training the models for tasks the robots can perform,” responded Kosowsky-Sachs. “At MODEX, we were surprised to hear people say, ‘I want this yesterday.’ Cassie isn’t just valuable for large third-party logistics providers—it’s also suited for other types of customers.”
Usage-based pricing aligns Tutor’s development with customer needs
Tutor Intelligence charges $14 to $18 per hour for Cassie’s services, which it says is competitive with the increasingly scarce manual labor market.
“Robotics as a service often comes with hidden drawbacks, so we prefer to discuss usage-based pricing,” said Gruenstein. “For us, innovation means developing technology that reaches a lower price point. We offer a ‘no contract’ or ‘no commitment’ model, which keeps our incentives aligned to build an intelligentrobot workforce and partner with businesses, rather than integrators, which make money by supporting complexity.”
BetterBody Foods has added the palletizer to its facilities in Utah and Massachusetts.
“For the deployment process, Tutor lived up to its word,” said Jeff Pulley, facility manager at the company. “The hardest part about the install was setting up the electrical and airdrop. With their team onsite, we got the robot plugged in and up and running in a few hours. Now, it can build on a skewed pallet or straighten boxes.”
Pulley told The Robot Report that he looks forward to using Sonny to run during multiple shifts. “It was cool to see the progress,” he said after touring DF1.
Productiv Inc. conducts highly variable kitting for e-commerce, medical device, and cosmetics customers. It assembles 30 million kits per year with 10 to 15 items each. The Richmond, Va.-based company looked at force- and power-limited robots and humanoids before getting started with automated palletizing using Cassie.
“We didn’t want to spend $150,000, but $14 per hour was competitive,” noted Paul Baker, chief financial officer at Productiv. “The robot was profitable from Day 1 and keeps up the line speed.”
Productiv has four metrics: SKU coverage, placement accuracy, pack time, and how long it takes to program a new SKU. Tutor met all of them, from 99.9% quality and helping to ship 350,000 kits to reducing programming time from five days to one, Baker said.
Tutor Intelligence is also working with leading brands and a major 3PL. The company is hiring for roles including research and software engineers, customer success managers, and salespeople.
