🌏 中文版
The Round
| Item | Value |
|---|---|
| Company | Tutor Intelligence (Watertown, Massachusetts, USA) |
| Round | Series B |
| Amount | $75M |
| Lead | Andreessen Horowitz (a16z) |
| Participants | Union Square Ventures, Neo |
| Valuation | Undisclosed |
| Total raised | $117M |
| Founded | 2021 (spun out of MIT CSAIL) |
| Headcount | ~19–31 (third-party estimate) |
What the company does
Tutor Intelligence sells robot labor the way businesses already buy labor: by the hour. Customers don't buy a robot — they pay for the work, like renting a worker.
Its products are robot workers named Sonny and Cassie, guided by remote human "tutors" who help with picking, moving, and handling tasks on factory and warehouse floors. Tutor's logic runs backward from how most robotics companies operate: deploy robots into real production work first to earn hourly fees, then feed the real-world operating data generated by every task back into training its frontier robot foundation models — letting the robots get smarter as deployment scales. It's the same loop that lets software LLMs learn from the internet's text, except the physical world never had its own "Wikipedia" to train on. Tutor is building one from scratch.
Tutor's robots are now running in nearly four times as many facilities as a year ago, and almost a third of customers that received robots in the past 12 months came back for more. The company is targeting more than 1,000 robots under contract — all built and deployed in the US — by the end of 2027, working inside major manufacturing and logistics networks that handle everyday goods like cereal, snacks, candy, cosmetics, household items, and toys.
What this round signals
What it means for the agent ecosystem
The "AI agent" narrative is stretching from pure software into physical labor. Tutor treats deployment itself as a data-collection pipeline: every completed task generates real-world operating data that didn't exist before, which then trains better frontier models — forming a flywheel of deploy → real data → stronger model → more deployment. The $75M is earmarked mainly for frontier robot foundation model training and for manufacturing and deploying more Sonny and Cassie units, not for plugging a research gap.
- Robot workers deploy to factory and warehouse floors
- Tasks like picking and handling generate real-world operating data, written into the robot foundation model
- Accumulated real-world data trains and improves the frontier robot foundation model
- The model gets stronger, robots take on more kinds of work, deployment expands, and the next loop starts from a stronger model
Simplified for illustration — in practice, data still needs filtering and labeling before it can be used for training.
What investors are betting on
Lead investor Raghu Raghuram, a16z general partner and former VMware CEO, puts it plainly: robotics research has long been stuck waiting to perfect the model before deploying it. Tutor flips that — real-world deployment demand drives the research agenda, which then works backward through distribution and hardware. That's the same training path software agents have validated in recent years: improve continuously from real usage signals rather than relying purely on synthetic or simulated data. For a16z, this is an extension bet — taking the real-world feedback loop already proven in software agents and applying it to robot hardware.
Numbers worth watching
- Total funding of $117M, with existing investors from both the seed round (led by Neo) and Series A (led by Union Square Ventures) returning to participate in this round — a sign of sustained insider conviction
- Facilities running Tutor robots have grown nearly 4x year over year, and almost a third of customers that received robots in the past 12 months came back for more — growth from both new and returning customers
- The round ranks among the top 10% of US deep-tech Series B deals by size, in a sample of 2,808 rounds, reflecting growing capital concentration in physical AI
Watchlist status
Tutor Intelligence is not yet on the watchlist. Recommend adding it to section D9 (Industrial / Manufacturing / Supply Chain), with a caveat: the existing D9 roster (Trimble, Plataine, Eyelit, Infor) is factory-scheduling and asset-management software, while Tutor is a rare combination of robot hardware plus a continually learning model. Its tracking focus — hourly-billed robot deployment and training frontier models on real-world data — differs from the rest of the list, and a dedicated physical-AI subcategory may be worth splitting out if more such companies show up.
Today's takeaway
The software-agent story of "improving continuously from real usage signals" has now been carried over to robots, unchanged, by a16z. Tutor doesn't perfect its models before deploying them — it deliberately reverses the order: deploy first to earn hourly fees, then feed real operating data back into training. That turns physical AI's training bottleneck directly into part of the business model itself.