Two million hours of human actions to train robots

Midcentury has emerged from stealth with $15 million in funding and an infrastructure stack for Physical AI. The startup combines a large-scale egocentric video dataset with Matrix, a simulation platform designed to evaluate and post-train robot policies before deployment.

Human actions instead of web pages

Language models have access to enormous amounts of text, images, and code available online. Training a robot to physically manipulate its environment presents a different data problem: actions have to be observed along with their spatial progression and physical consequences.

That is the layer Midcentury is targeting. The startup says it has collected more than two million hours of egocentric data across over 50 environments and 20,000 tasks, with hands visible in 90% of frames.

Video is only one component of the dataset. Sequences include supervision for 3D hand and body pose, depth, tactile interactions, and motion. Subtask and cycle annotations are also frame-aligned with millisecond timestamps.

Midcentury describes it as the world's largest egocentric dataset. Without an independent comparison establishing that ranking, the claim remains the company's own. Matrix lets robots practice before entering the real world

The second part of the infrastructure focuses on simulation. Matrix is designed to test robot policies and improve them through reinforcement learning without repeatedly running experiments on physical hardware.

The platform combines classical simulation, learned physics, and real-world data. Massively parallel cloud execution is intended to run evaluations thousands of times before a policy is transferred to a robot.

The underlying problem is sim-to-real: a policy that performs well inside a simulation can fail when physical conditions differ from its virtual training environment. Midcentury uses its real-world data to bring simulated environments closer to the conditions systems encounter after deployment. Infrastructure around Physical AI

This combination of data and simulation positions Midcentury upstream from the robots themselves. The company is not building a humanoid or a new general-purpose robot policy, but infrastructure for pretraining data, evaluation environments, and reinforcement-learning loops used by teams developing those systems.

The team lists backgrounds spanning Stanford AI Lab, OpenAI, Google DeepMind, NVIDIA, Scale AI, and Invisible, with previous work involving RoboNet, GDPVal, and NVIDIA Cosmos 3. These credentials are reported by Midcentury and do not specify each team member's individual contribution to those projects.

The stealth exit comes with a $15 million Seed round. Midcentury also says it is already supporting frontier labs, although it has not publicly identified them.

Dataset samples and early access to Matrix are available through the Midcentury website.