Sensei Robotics wants to teach machines by hand

A modern robot can map a room, plan a path, and narrate what it sees. Ask it to fold a bath towel, and the whole act of intelligence tends to fall apart. The reason is boring, and it is the reason Sensei Robotics exists: the machine has never watched enough people do it. Robotics does not have a thinking problem so much as a showing problem, and demonstrations - actual humans performing actual tasks - are the scarcest resource in the field.

Sensei Robotics, a company from Y Combinator's Summer 2024 batch, is built around that scarcity. Its pitch, repeated on its own launch materials, is direct: it wants to be "the Scale AI for robotics training data." Where Scale AI grew by labeling the internet's images and text, Sensei is trying to label the physical world - the grip, the reach, the small corrections a hand makes while stacking a plate. The company sells that data as a service to robotics and AI teams who need mountains of it and do not want to build the collection pipeline themselves.

The ProblemThe most expensive thing a robot has to learn

Teaching a robot a single new skill is not cheap. Standard approaches lean on teleoperation - a person driving a real robot arm through a task while sensors record every motion - and the setup for that can run well past $40,000 per robot before a single useful demonstration is captured. The rigs are expensive, they live in a lab, and they are slow. A model that needs thousands of varied examples of "load a dishwasher" cannot easily get them if each example requires studio-grade hardware and a controlled room.

That gap is what economists would call a bottleneck and what a founder would call an opening. The algorithms for imitation learning are increasingly public. The compute is rentable. The demonstrations are not for sale in any organized way. Sensei's wager is that whoever makes the demonstrations cheap and abundant sits underneath the entire industry.

The ProductA sensor sleeve instead of a robot arm

The core idea is to remove the robot from the data-collection step. Instead of paying for an expensive arm and driving it remotely, a person wears a low-cost sensorized exoskeleton - an instrumented sleeve, effectively - and simply does the task with their own body. The device, which Sensei says costs under $300 to make, carries angle, vision, and inertial sensors. Those capture the visuo-spatial information a robot needs: where the hand was, how it moved, what it saw, how it adjusted.

The hardware is only half of it. The other half is a software platform and a workforce. A customer submits a task - fold this cloth, sort this bin, load these dishes - and Sensei routes it to a distributed network of trained operators it calls "Senseis." They perform the task many times, in varied real settings rather than a single lab, and the platform returns curated, labeled recordings. The variety is part of the value: models trained on demonstrations from many hands, homes, and lighting conditions tend to generalize better than models trained on one bench.

There is a design choice worth pausing on. Teleoperation asks a person to think like a robot - to drive an arm through a screen, fighting latency and an unfamiliar body. Sensei inverts it. The operator just does the task the way a person naturally would, and the sensors translate that motion into something a robot can imitate. Lowering the skill required to produce a good demonstration is what lets the workforce grow beyond a handful of trained specialists, which is the only way the numbers reach the scale that model training demands.

The FoundersFrom simulated dogfights to dish racks

Sensei was started in 2024 by Anubhav Guha and John Piotti, two MIT engineers who met as students and kept working together afterward. At Aurora Flight Sciences they were part of a DARPA-funded program building AI for autonomous fighter-jet combat - Piotti leading reinforcement-learning efforts, Guha working across robotics, control theory, and machine learning. Guha left an MIT robotics PhD to start the company.

The jump from combat autonomy to dishwashers looks like a step down until you notice what carries over. Both problems are about getting a machine to act well in the messy physical world, and both taught the founders the same lesson from the other side: the hard part is rarely the learning algorithm. It is feeding it. Having watched autonomy stall for want of good data, they went and built the data company.

The BusinessSelling the picks, not the gold

Sensei's model is a data-as-a-service marketplace. It manufactures the cheap collection hardware, trains and pays the operator network, and charges robotics and AI companies for bespoke, labeled datasets. The margin lives between what an operator costs and what a curated dataset is worth to a team racing to ship a manipulation model. It is a picks-and-shovels position: Sensei does not need to build the winning robot, only to sell to everyone trying.

The customers are the teams downstream of that data: robotics and embodied-AI companies training manipulation models that live or die on the volume and diversity of examples they can feed. These are buyers for whom data collection is a real line item and a real delay - the kind of internal cost a company is often happy to hand to a specialist if the specialist is cheaper and faster. Sensei is early enough that the roster of paying customers is not public, but the shape of the buyer is clear: anyone trying to make a robot do a physical task it has not seen enough of.

That framing also explains the competitive landscape. The obvious comparison is Scale AI, which built a large business supplying data for the last wave of AI - but Scale's world was pixels and text, not physical motion. Sensei's more immediate rivals are the in-house approach (teams building their own teleoperation setups, ALOHA-style rigs, and internal labeling) and the possibility that well-funded robotics foundation-model builders - the Physical Intelligences and Skild AIs of the field - decide to own their data pipelines rather than buy from an outside vendor. Sensei is betting that specialization wins: that collecting physical demonstrations well is hard enough, and central enough, to be a company on its own.

The OutlookWill it matter in ten years?

Sensei is early. It is a small team - two founders and a handful of people - operating a beta, backed by Y Combinator and listed with early investors including Network.VC. Public detail on paying customers, revenue, and the true scale of the operator network is thin, and the 10x-cheaper, 2x-faster claims are the company's own. Anyone reading this should hold those numbers loosely.

What is easier to judge is the shape of the bet. If general-purpose robots arrive, they will arrive on the back of enormous quantities of physical demonstration data, and someone will have organized the labor of producing it. Sensei's answer - cheap wearables, paid operators, data on demand - is a plausible version of how that supply chain looks. The company's whole thesis rests on an unglamorous truth: the road to automation runs, for now, through a lot of human hands patiently showing machines what to do.