Startup Spotlight: Praxis Robotics turns factory work into robot training data

Why it matters

Robot developers need physical-world data that cannot be scraped from the internet. Praxis is betting that access to existing workers and facilities can become defensible infrastructure, provided it can turn a claimed global network into repeatable deployments, trusted consent processes and datasets that measurably improve robot performance.

Rohan Seelamsetty (@ro_seelamsetty) is building Praxis Robotics around a bottleneck that money alone cannot quickly remove: physical access to the places where people assemble, clean, pack, inspect, repair and move things.

Praxis Robotics, a San Francisco operation founded in 2026 and backed by Y Combinator, records workers performing physical tasks and converts those demonstrations into training data for robotics developers. In its Y Combinator profile, Praxis says it can reach 60,000 workers across five continents and collect data from 150-plus types of industrial, residential and commercial environments.

Those figures are Praxis's own and have not been independently audited. The underlying strategy is still clear. Seelamsetty wants to make existing businesses part of the robotics supply chain, paying them for access to the people, facilities and workflows that robot developers struggle to reproduce in a laboratory.

That makes Praxis an access business before it becomes a data business. Cameras, wrist sensors and motion-capture systems can be purchased or assembled. Permission to record skilled work inside factories, warehouses, homes, laboratories and service businesses is harder to manufacture.

Seelamsetty studied computation and cognition at the University of Pennsylvania, was an Apollo Fellow at Oxford and competed for India's national debate team, according to YC. His path to robotics passed through factory floors, political work and student investing rather than a conventional robotics doctorate. That background fits the operational premise behind Praxis: technical data requirements matter only when someone can persuade real organizations to open their doors and then run a consistent collection process inside them.

Turning work into a dataset

Robotics models need records of physical action. Text models had websites, books and software repositories available at enormous scale. A robot learning to manipulate objects needs observations tied to movement, depth, position, force, timing and the consequences of an action.

Praxis records human demonstrations from a first-person perspective. Its product site shows a four-stream capture setup built around a head-mounted ZED 2i stereo camera, left and right wrist views and live motion-capture reconstruction. Praxis says other collections can include stereo RGB video, inertial measurements, glove kinematics, haptics and dense annotations.

The output is supposed to be model-ready rather than a folder of video files. In its YC launch page, Praxis says deliveries can be calibrated, synchronized, pose-tracked, quality-checked and scrubbed of personally identifiable information.

The range matters because physical tasks contain details that ordinary video often misses. A camera may show a hand tightening a component while revealing little about the force applied, the orientation of the wrist or the moment an operator feels resistance. Multiple synchronized sensors can preserve those signals, giving robotics researchers a better chance of connecting visual observations to actions.

Praxis also claims part of its hardware operation is based in Shenzhen, allowing it to build and deploy custom capture rigs. That arrangement could shorten the path between a customer's requested modality and a device placed in the field. Praxis has not published deployment counts, captured hours or hardware costs, leaving the operational efficiency of that setup difficult to assess from the outside.

Seelamsetty described the sales pitch directly when he brought Praxis out of stealth during the summer. Robotics developers can request data samples, while businesses in industrial, hospitality or construction operations can turn their environments into what he called a "monetisable asset."

That two-sided pitch gives Praxis two separate sales jobs. It needs robotics labs willing to pay for datasets, and operating businesses willing to let workers and facilities become the source material. Each side depends on the other. More environments improve the catalog Praxis can offer data buyers, while committed buyers give participating businesses a reason to accept the disruption and compliance work involved in collection.

A team built around research and access

Seelamsetty is joined by COO Dev Karpe and CTO Tommy Li. All three founders have ties to Penn, though their experience covers different parts of the operation Praxis is attempting to assemble.

Karpe participated in Penn's Huntsman Program, according to YC, and previously worked in investment banking and supply-chain operations in the office of a Fortune 500 chief executive. Those roles map to Praxis's less visible challenge: navigating procurement, site access, worker scheduling and the internal politics of recording production work.

Li studied computer science and finance at Penn. YC lists previous work in technology private equity at Morgan Stanley, industrial-technology venture capital at Andreessen Horowitz and robotics research at NASA's Jet Propulsion Laboratory and PerceptIn Robotics. Praxis leans heavily on that research background when arguing that it can translate a laboratory's technical specification into a repeatable field collection.

The division of labor is unusually aligned with the product. Seelamsetty handles the relationship and narrative layer, Karpe brings operational and supply-chain experience, and Li brings the robotics and data background. Their task is to prove those skills produce repeatable contracts rather than bespoke data projects held together by founder involvement.

Y Combinator lists Praxis in its Summer 2026 batch with a three-person team. YC is the only confirmed investor. Praxis has not announced a valuation or total funding figure.

Praxis attracted attention at the batch's Demo Day. In a September 13th report, TechCrunch included Praxis among nine businesses cited by at least two early-stage investors. The report repeated Praxis's claim that it works with publicly traded businesses and has collected video in 150-plus environments, without identifying those partners.

Robot data is becoming its own venture category

Praxis is arriving as investors put significant capital behind businesses that gather and process physical-world data. The competition ranges from factory networks and gig-worker recording programs to dedicated teleoperation facilities running fleets of robotic arms.

Human Archive, another YC-backed operation, said in May that it had raised $8.2 million from Wing Venture Capital, NVP Capital and others. Human Archive equips service workers with cameras and other sensors, using India's home-services, hotel and restaurant sectors as a collection network. It told TechCrunch that it had deployed 1,000-plus headsets, alongside dozens of devices for collecting motion and tactile signals.

XDOF has taken a broader and far more capitalized approach. XDOF raised $70 million from Thrive Capital, Spark Capital, Andreessen Horowitz, Lux Capital and WndrCo to build collection tools, annotation systems and data operations. XDOF told TechCrunch in June that it served 20 customers and employed about 60 people. Its work spans teleoperation on target robots, general robot datasets and planned human-worn sensors.

Vision Lab is pursuing industrial access through a claimed network of 2,000-plus factories. It announced a $6 million seed round in June led by Race Capital, with backing from YC, Foothill Ventures, 500 Global and others. Vision Lab says it collects first-person and external video inside active production environments.

Config, founded in 2025, raised a $27 million seed round led by Samsung Venture Investment at a valuation above $200 million, TechCrunch reported in May. Config combines field collection with controlled studios and software intended to translate human-motion data into forms better suited to robot movement.

Those rounds show how quickly robot data has moved from a research constraint into a funded infrastructure market. They also raise the standard Praxis must meet. Access to many environments is useful, though buyers will ultimately judge datasets by whether models trained on them perform better on robots.

Praxis's distinction is its attempt to make existing organizations into recurring data suppliers across a wide range of settings. Human Archive has emphasized service workers. Vision Lab has emphasized factories. XDOF is building a vertically integrated data operation with substantial capital and dedicated facilities. Praxis is presenting a wider access network that stretches from smaller businesses to conglomerates and from industrial sites to homes.

That breadth can become an advantage if Praxis can standardize collection without erasing the diversity buyers want. It can also become an operational burden. Every new environment introduces different lighting, equipment, worker behavior, safety rules, privacy requirements and data formats. A refinery, a pharmacy and a residential kitchen may all generate useful demonstrations, but each requires its own collection protocol.

The claims that still need proof

Praxis's public materials name no data buyer, paying customer or host business. YC describes Praxis as embedded within publicly listed and unicorn-scale conglomerates, including one organization that Praxis says accounts for 3% of a country's gross domestic product. The identities and commercial terms behind those relationships remain private.

That leaves the 60,000-worker figure open to interpretation. A reachable network can describe workers employed by partner organizations rather than people actively collecting data. Praxis has also given no public count of recording devices, participating workers, completed sessions or hours delivered to customers.

The distinction matters. Access agreements can create a pipeline, while deployed equipment and accepted datasets produce revenue. Praxis's early evidence establishes a broad claimed distribution network. It does not yet establish how much of that network has been converted into recurring collection.

Worker consent and privacy will also shape the business. First-person recording can capture faces, screens, documents, conversations, homes and proprietary industrial processes. Praxis says it removes personally identifiable information, though it has not published details about consent procedures, worker compensation, retention periods or how host organizations control downstream use.

These questions follow the product wherever it goes. Human Archive has already faced public scrutiny over consent and worker pay in India. Industrial partners may be more willing to participate when contracts keep data within defined technical and commercial boundaries. Residential collection carries a higher burden because cameras can record bystanders and intimate spaces that were never designed as data-production sites.

Praxis's strongest opportunity comes from a simple mismatch. Robotics developers want varied physical-world data, while millions of workers already perform the relevant tasks every day. Recreating that work in dedicated labs is expensive and can produce sterile demonstrations that fail to capture mistakes, interruptions and environmental variation.

Seelamsetty's bet is that Praxis can organize those scattered demonstrations into a dependable supply chain. The hardware is visible. The founders' backgrounds fit the work. The next proof will come from conversion: named or independently confirmed buyers, repeat deployments and evidence that models trained on Praxis data perform better in the physical world.

If Praxis reaches that point, participating businesses gain a second product hiding inside their normal operations. Every repair, inspection and assembly process becomes potential training material. Praxis then owns the connective layer between human experience and machines trying to imitate it.