The Y Combinator startup sends workers into hotels, warehouses, and restaurants wearing cameras and grippers. What they capture becomes training data for robots and world models.
The hardest thing to teach a robot is something a person does without thinking. Fold a towel. Restock a shelf. Wipe down a counter that has three cups, a spilled napkin, and a phone charger in the way. These moments never repeat exactly, and that is precisely why they are so hard to program. Sureform, a two-person startup from Y Combinator's Spring 2025 batch, built its entire company around a single response to that problem: stop trying to invent the data, and go record it where it already happens.
Based in San Francisco, Sureform collects real-world, multimodal human data - the raw material that robotics teams need to train robots and what researchers call world models. Its pitch is unglamorous and specific. Somewhere right now, a housekeeper is making a bed, a line cook is plating a dish, a warehouse picker is sorting boxes. Each of those shifts is a demonstration of physical skill. Sureform's job is to capture that skill in a form a machine can learn from.
For years, a common way to teach robots was inside simulation - clean virtual environments where a machine could practice a task millions of times. Simulation is fast and cheap, but it has a known weakness: the real world is messier than any simulator. Lighting changes. Objects sit at odd angles. People improvise. A robot trained only on tidy data tends to fail the moment reality stops cooperating.
Sureform's argument is that the variation itself is the lesson. To learn a specialized job - pharmacy restocking, power-line repair, kitchen prep - a model needs to see that job done across many real situations, not one idealized version of it. That data is expensive and scarce, and it does not exist on the open internet. It lives inside ordinary workplaces, in the hands of people who are already good at the task.
Consider the range of things that never make it into a clean dataset. A worker adjusts their grip when a box is heavier than expected. They pause when a colleague walks past. They reach around a mop bucket that should not be there. Those small corrections are the difference between a robot that works in a demo and one that works on a Tuesday afternoon in a real building. Sureform's whole reason for existing is to capture the Tuesday afternoon.
Here is how a Sureform session works. An employee at a partner site puts on a small kit: a head-mounted RGB camera to see what they see, a stereo depth rig to measure distance and shape, and a handheld gripper that records how a task is actually grasped, not just how it looks. As the worker does their normal job, the system captures time-synced video, audio, pose, depth, and tactile signals - all lined up on the same clock.
That last part matters. Raw footage is not the product. The value is in synchronization: video matched to depth matched to the motion of the hands matched to touch. Sureform structures every session and delivers it as training-ready data, so a robotics team can feed it into a model without months of cleanup.
The design choices are practical. A head-mounted camera gives a first-person view, which is closer to what a robot's own sensors will eventually see than a fixed overhead camera would be. The stereo depth rig turns a flat image into a sense of three-dimensional space. And the handheld gripper is the quiet clever part - by having the worker act through a gripper rather than their bare hand, the recorded motion maps more directly onto how a robot arm would perform the same task. The kit is built to make the data transfer, not just to make it look impressive.
Sureform sits between two groups that rarely talk to each other. On the demand side are robotics labs and AI teams building robot foundation models and world models. They have compute and talent but not enough real-world action data. On the supply side are everyday workplaces - hotels, warehouses, restaurants, construction sites, laundromats - full of people doing exactly the tasks those labs want to learn.
The company connects the two directly. Workplaces host data-collection sessions, workers get compensated for contributing, and robotics teams receive datasets tuned to specific tasks. It is a marketplace, and the thing being traded is human expertise made machine-readable.
The business model follows the shape of the market. Sureform sources task-specific data on the supply side and sells structured, training-ready datasets on the demand side. For a partner business, hosting sessions turns a normal workday into a source of revenue. For a robotics lab, buying a Sureform dataset is cheaper and faster than standing up a data-collection operation from scratch. The company positions itself as the layer that makes both sides easier, and takes its place in the middle.
Sureform was founded in 2025 by Ananth Kashyap, who serves as CEO. His background reads like a tour of the field's harder problems: work at Google, plus earlier stints at HEBI Robotics, Carnegie Mellon's Robotics Institute, and Near Earth Autonomy, following studies at the University of Pennsylvania. The through-line is robotics, and with it a recurring frustration that many in the field share - capable machines held back by a shortage of real examples to learn from.
That frustration is the seed of the company. Rather than build another robot or another simulator, Kashyap built the layer underneath both: the data. It is a less flashy place to compete, which is part of the point. In a field where nearly everyone is racing to build the robot, far fewer are building the supply of examples the robot needs.
The expertise Sureform brings is less about any single gadget and more about the whole pipeline. Capturing five synchronized data streams in a working environment, keeping them aligned, and delivering them in a form a model can actually train on is a systems problem. It touches hardware, sensor timing, data structuring, and the logistics of getting into real workplaces and back out with usable footage. That combination - robotics fluency plus the operational grind of collection - is the part that is hard to copy.
Sureform operates in the same broad space as data-collection and labeling companies that serve AI, but its angle is narrower and more physical. Generic web data does not teach a machine to grip a wet towel or reach around an obstacle. Teleoperation and lab-based capture can produce high-quality data, but they are slow and detached from the messiness of a working shift. Sureform's counter is to embed capture into real jobs, betting that authenticity and scale beat controlled conditions.
The alternatives it competes against are as much philosophies as companies: simulation-first pipelines, in-house data efforts at robotics labs, and other real-world data vendors. Whether Sureform's approach wins depends on a question the whole industry is testing - how much does real-world variation actually matter to a robot's reliability. Sureform is wagering that it matters a lot.
For now the company is small: two people, a working capture kit, and a two-sided platform still early in its life. What it has is a clear idea and a growing consensus behind it - that embodied AI's tightest constraint is not compute or model size, but the supply of real-world data. If that consensus holds, the boring middle of the stack is a useful place to be standing.
Video: Sureform has not published a public product demo or founder interview at the time of writing. Check the Y Combinator profile and the company's X account for the latest.