New New York startup, Midcentury, has emerged from stealth mode, announcing the raising of a $15 million seed round. The company aims to solve the problem of data scarcity in robotics. The company has presented two main products: an extensive egocentric dataset containing over two million hours of human behavior, and a cloud simulation platform called Matrix.
The raised funding round places Midcentury among the top 1% of all seed deals in the big data category. For comparison, the median size of a seed round in the second quarter of 2026 was about $4.5 million. However, the startup did not disclose its valuation or information about investors.
The company plans to use first-person human action data to train robots, similar to how web scraping trained large language models. Its proprietary dataset covers over 50 environments and 20,000 tasks. Each hour of recording is annotated, including 3D hand pose tracking, depth maps, and point trajectories. This scale significantly exceeds existing public research options, such as Ego4D v2, which contains approximately 3,600 hours.
In addition to recordings of human actions, Midcentury possesses about 50,000 hours of gaming data with engine-level signals, as well as approximately 69,000 hours of conversational voice data in 25 languages. Since this dataset is strictly proprietary, researchers will not receive free access to it.
Midcentury also released Matrix—a cloud simulation platform that allows teams to create digital twins of real-world scenarios. Instead of relying on hard-coded rules, the platform derives physics directly from real data. Engineering teams can run thousands of parallel tests on GPU clusters, turning failures into immediate examples for training.
The company actively leans on Richard Sutton's 'brute force' concept, according to which raw computational power and massive amounts of data ultimately surpass manual feature engineering. CEO Chetan Kulhari previously worked at the AI coding startup Magic. According to SEC filings from January 2026, the company sold about $8.9 million to five backers, indicating a staged closing of the round.
There are many startups in the market competing for control over the data layer for embodied AI. Competitors such as Rerun and Vision Lab have also attracted venture funding in the last year. If scaling laws successfully transition from language models to physical robots, market positions will likely be determined by the entities controlling the training data.
