> ## Content Index > Fetch the complete content index at: https://www.tao.media/llms.txt > Use this file to discover other available public pages before exploring further. # Midcentury Raises $15M Seed to Build Physical AI Data and Simulation Infrastructure - URL: https://www.tao.media/midcentury-raises-15m-seed-to-build-physical-ai-data-and-simulation-infrastructure/ - Published: 2026-09-23T14:31:20.000Z - Updated: 2026-09-23T14:31:20.000Z - Description: The stealth exit pairs a large egocentric robotics dataset with Matrix, a simulation platform for evaluating and improving robot policies before deployment. - Author: Bart Hillerich - Tags: Midcentury, Robotics, News [Midcentury](https://www.midcentury.xyz/?ref=tao.media) has emerged from stealth with a $15 million Series Seed to build data and simulation infrastructure for physical AI. It is targeting two of robotics' hardest scaling problems: real-world action data and reliable policy evaluation. The company [announced the funding and launch](https://x.com/MidcenturyAI/status/2102412610071339414?ref=tao.media) on Tuesday, saying it is already supporting frontier labs with what it calls the world's largest egocentric robotics dataset and a simulation platform called Matrix. Midcentury’s dataset includes more than 2 million hours of egocentric data across more than 50 environments and 20,000 tasks. > "Human action data has finally unlocked a scaling law for Physical AI. It’s time to join LLM researchers in taking the bitter lesson pill," the company wrote in its launch post. "Scaling robotics now requires internet-scale pretraining data and reliable simulation for evaluation + RL. We build the core infra for both." The announcement arrives as robotics companies and AI labs increasingly look beyond language and internet-scale text data toward embodied systems that can perceive, act, and learn in physical environments. Midcentury's pitch is that robot learning needs a similar scaling layer: large volumes of high-quality human action data paired with simulation environments where policies can be tested and improved before deployment on hardware. 0:00 /1:22 1× Midcentury launch video ## Midcentury Is Building Around Egocentric Robotics Data Midcentury's core dataset centers on egocentric video, meaning footage captured from a first-person perspective as people perform tasks in real environments. That perspective helps robot learning because it shows hands, objects, body motion, timing, and task context from the point of view of an acting agent rather than a fixed external camera. The company’s dataset includes hands visible in most frames, along with action supervision such as 3D hand and body pose, depth, tactile signals, motion information, and task annotations aligned to fine-grained timestamps. That turns raw video into structured training material that helps models learn how physical actions unfold over time, not only what a task looks like. Midcentury's egocentric data is rights-cleared robotics and world-model data captured with a "full signal stack" for embodied agents. There are additional data categories, including engine-level gameplay data aligned with player inputs and telemetry, as well as multilingual conversational voice data. The bigger goal here is for the data to help models generalize beyond controlled demonstrations. Physical AI systems often perform well in scripted videos but struggle when lighting, object placement, human behavior, or environment geometry changes. Midcentury presents its dataset as part of the answer: real-world variation at a scale large enough to support pretraining and downstream policy improvement. ## Matrix Targets Policy Evaluation and Post-Training The second part of Midcentury's platform is Matrix, an agentic simulation system designed to evaluate and improve robot policies at scale. Matrix combines classical simulation, learned physics, and real-world data so teams can test policies thousands of times before sending them back to physical robots. The platform is a massively parallel cloud simulation environment for evaluation and reinforcement learning. Real-world robotics testing is slow, expensive, and operationally constrained. A robot policy that needs thousands of attempts to improve cannot rely only on physical trials, especially if failures could damage hardware, interrupt a customer site, or create safety risks. Simulation gives developers a place to replay failures, change conditions, compare model versions, and expose policies to edge cases that rarely appear in a real deployment. As such, Matrix is oriented around three functions: designing digital twins and scenario distributions around deployment conditions, testing policies across closed-loop scenarios, and turning failures into additional training experience. With Matrix, Midcentury wants to build a loop where real-world data improves simulation and simulation feeds back into policy training. ## Midcentury's Team Frames It as Robotics Scaling Infrastructure Midcentury’s team has backgrounds at Stanford AI Lab, OpenAI, DeepMind, NVIDIA, Scale AI, and Invisible, with prior work connected to RoboNet, GDPVal, and NVIDIA Cosmos 3. The seed round is a big milestone for the group; it funds their platform built on the idea that robot learning will scale through the combination of real-world action data and simulation environments large enough to turn failures into training signals. Midcentury is now offering early access to Matrix and dataset samples through its website.