Realset AI raises $10M to capture real-world data the web can't provide
What's the deal? Realset AIDealroom has a profile for this one. Try Dealroom →, a data lab that produces training data for large language models, frontier models and embodied agents, has raised $10 million in a Series A alongside FlatkeyDealroom has a profile for this one. Try Dealroom →. The San Jose company will use the money to expand its capture network of real workplaces and studios, grow its pool of expert demonstrators, and fund open benchmarks that test whether AI works outside the lab.
Why now? Frontier labs and robotics firms have largely consumed the text on the internet. The next gains, Realset argues, come from data on people doing real tasks — how a worker folds laundry, packs an order or handles a return — which does not exist online and cannot be reproduced in simulation.
What's the endgame? Realset runs three products. Realset BodyDealroom has a profile for this one. Try Dealroom → captures skilled workers performing manipulation tasks in homes, kitchens and warehouses for vision-language-action training; Realset FieldDealroom has a profile for this one. Try Dealroom → builds reinforcement learning environments from real workflows like e-commerce operations and customer support; and Realset JudgeDealroom has a profile for this one. Try Dealroom → offers expert evaluation of production agents by people who do the job the agent is replacing.
All three run on the Realset Workspace, where domain experts working in six languages — English, Simplified Chinese, Japanese, Korean, Spanish and Arabic — complete tasks, attach evidence, and pass an independent quality review before a record is approved. Every approved record exports with provenance that customers can verify.
What's next? The company plans to publish open benchmarks measuring real-world performance. The first, the Realset Household Manipulation Bench, evaluates open-source policies including π0, OpenVLADealroom has a profile for this one. Try Dealroom →, GR00TDealroom has a profile for this one. Try Dealroom → and OctoDealroom has a profile for this one. Try Dealroom → on tasks such as folding and sorting, with results expected in Q4 2026. A Light Assembly Bench and a Commerce Ops Agent Bench are also planned.
The signal: As the supply of scrapeable text runs dry, value is shifting to data that must be captured, not collected. "The easy data is gone. What is left is the physical world, and you cannot scrape it," said founder Hunter Guo. "That is a capture and quality problem, not a labeling problem."
Image credit: Generated with Gemini
