MIT-founded startup Tutor Intelligence has launched DF1 — the largest "robotic data factory" in the US by its own account. One hundred bimanual Sonny manipulators are learning warehouse tasks under the guidance of remote teleoperators. The company claims it can detect and correct robot behavior errors 100 times faster than traditional methods, and plans to move from training to commercial pilots by end of 2026.
Key takeaways
- DF1 in Watertown, Massachusetts is claimed to be the largest robotic data factory in the US
- 100 Sonny robots trained by 45–50 remote teleoperators from Mexico and the Philippines
- Ti0 VLA model trained with "velocity normalization" — aligning the movement styles of different operators
- Cassie, the company's mobile manipulator, is commercially available at $14–18/h with no contract
- Company raised $34M in a Series A round in December 2025
A laboratory for robot intelligence
"I'm not building models. I'm building a way to get the right data from people teaching robots," says Josh Gruenstein, co-founder and CEO of Tutor Intelligence. The analogy he uses is unexpected: he compares DF1 to the Large Hadron Collider — not as a finished product, but as an instrument of scientific discovery on the path to scaling humanoids.
Tutor Intelligence was founded in 2021 from MIT CSAIL, and in December 2025 raised $34 million in Series A funding. The company was part of the first cohort of the Physical AI Fellowship organized by MassRobotics in partnership with NVIDIA and Amazon Web Services.
DF1 is housed in a historic mill in Watertown, Massachusetts. One hundred Sonny semi-humanoid bimanual manipulators started with piece-picking tasks common to e-commerce and kitting. The early weeks were chaotic — the robots handled sponges and snack bags clumsily, like any new class of kindergartners. But that is the point: DF1 is designed to produce data about errors and corrections, not just successes.
The learning technique: speed and real-time correction
A key challenge in training a robot fleet is data inconsistency. Every teleoperator has a different movement style — different speed, rhythm, grasping approach. A model trained on such varied data may struggle to generalize.
Tutor Intelligence addresses this through "velocity normalization" — a preprocessing method that aligns the speed profiles of demonstrations across different operators. According to the company, this enables Ti0 (the Vision-Language-Action model) to learn a more consistent movement policy.
Fleet scale provides an additional advantage: "By evaluating the same policy across all 100 robots, we are able to detect and correct robot behaviors 100x faster. An edge case that may normally require 8 hours of robot operation to notice will be visible in only 5 minutes of DF1 operation," the company states.
Cassie already earns — no contract, $14–18/h
While DF1 trains future models, the company is already generating commercial revenue via Cassie — a mobile manipulator for picking and palletizing. Cassie can be deployed in two days, handles boxes up to 22.6 kg, and operates across variable SKU flows without reprogramming.
The pricing model is unconventional: $14–18 per hour, no long-term contract. "RaaS often has hidden catches. We talk in terms of usage-based pricing," says Gruenstein. "No contract means aligned incentives — we don't earn from complexity, only from performance."
Customers include BetterBody Foods (Utah and Massachusetts) and Productiv Inc. (kitting for e-commerce, medtech, and cosmetics). Paul Baker, CFO of Productiv, notes: "The robot was profitable from Day 1 and keeps up the line speed."
Why it matters
Tutor Intelligence is attacking one of physical AI's hardest problems: the lack of robot data at a scale comparable to internet data for language models. Rather than relying on simulation, the company is building a real data infrastructure — 100 robots working simultaneously, with real objects, in real conditions, with human correction in real time.
This approach has a chance to solve the problem every humanoid manufacturer faces: how to quickly teach a robot new tasks without months of data collection. If DF1 proves the "data flywheel" truly works at scale, Tutor could become a key provider of training infrastructure for the entire industry.
What's next?
- Gruenstein expects to move from training to commercial pilots with Sonny before end of 2026
- Cassie and Sonny share components — platform convergence may simplify deployments and reduce costs
- The company is hiring research and sales engineers — a signal of further expansion
Sources
- The Robot Report — Tutor Intelligence builds Data Factory to train robot AI in the real world
- Tutor Intelligence blog — Building a 100-Robot Data Factory: Toward Factory-Ready AI
