Cortex AI Builds the Real-World Dataset Embodied AI Has Been Missing

Cortex AI, a Y Combinator Fall 2025 company, is racing to build that missing dataset and become the data backbone for the labs training the next generation of robots. The AI revolution of the past five years was built on the back of internet-scale data. Trillions of words and billions of images turned out to be exactly what large neural networks needed to develop genuinely useful capabilities. But the next frontier — embodied AI, the kind that powers humanoids, manipulators, and general-purpose robots — has no equivalent dataset. The physical world is messy, expensive to capture, and almost entirely untranscribed.

Founded in 2025 and based in San Francisco, Cortex AI is building what it calls “the world’s most diverse and large-scale real-world workplace robot and egocentric dataset” — turning the physical world into the next training and evaluation set for embodied AI. The company powers frontier labs developing robotics foundation models and general-purpose robots by providing three categories of data they cannot easily collect themselves: egocentric data (real-workplace human video with hand and body pose, depth, and subtask labels); robot data (trajectories collected from manipulators and humanoids in real industrial settings); and human-in-the-loop rollouts and evaluations, in which remote operators recover robots when they fail. That last category is critical — the failure-and-recovery data feeds back into model training, allowing labs to systematically improve performance in exactly the situations where their robots break.

The company is led by Lucas Ngoo, founder and CEO. Prior to Cortex AI, Ngoo was the co-founder and CTO of Carousell, the consumer-to-consumer marketplace that scaled to a $1B+ valuation. That marketplace background is more than résumé garnish — it’s central to Cortex AI’s business model. Through the Cortex Marketplace, real workplaces get paid to host data-collection and evaluation sessions, while AI labs gain access to the in-the-wild data that actually matters. The two-sided marketplace structure is a direct inheritance of Ngoo’s experience scaling Carousell, applied to a market where supply (workplaces with interesting tasks) and demand (frontier labs hungry for embodied data) have until now lacked any liquid connection.

The market opportunity is enormous and growing fast. Robotics foundation models are now the most expensive, most ambitious bets at major AI labs, and their progress is bottlenecked precisely by the kind of diverse real-world data Cortex AI specializes in. Egocentric video with rich annotation is rare; manipulator trajectories from real industrial sites are even rarer; and structured evaluation data with human-in-the-loop recovery is essentially absent from the public ecosystem. The company is positioning itself as the indispensable data layer for an industry that is collectively spending billions on training robots that need exactly this fuel.

Cortex AI’s focus areas — artificial intelligence, reinforcement learning, and robotics — reflect its position at the center of one of the most strategically important areas of contemporary AI. With a focused three-person team, the company is moving fast to lock in marketplace supply (workplaces willing to host data collection) and customer relationships (frontier labs and humanoid companies) at the same time.

Researchers, robotics labs, and prospective workplace partners can learn more at the Cortex AI website, the Y Combinator profile, or follow the team on LinkedIn, X, and GitHub.

If Cortex AI succeeds, the company won’t just supply data — it will define the substrate on which the next decade of robotics is trained, evaluated, and improved.