Data and Simulation Infra for Physical AI
Real-world egocentric data and high-fidelity simulation at scale for training, evaluating, and deploying robots.
01 Mission
An applied research lab scaling robotics.
The next industrial revolution begins when robots leave the lab.
Physical AI produces impressive demos. Yet very few robots ever reach the real world.
Scaling robotics requires two things:
Data that captures the diversity and complexity of the real world.
Simulation where robots can be evaluated and learn through experience.
Midcentury builds both.
02 Datasets
High-Fidelity Data for the Real World
Rights-cleared robotics and world model data, captured in motion with the full signal stack for the next generation of embodied agents.
Egocentric Vision
One of the largest unscripted egocentric datasets across industrial and everyday environments. Video, IMU, and audio enriched with 3D hand pose, point tracks, depth, and task annotations.
- 1.9M+
- Hours
- 15+
- Environments
- 3D Pose
- + Annotations
Engine-Level Gameplay
Rights-cleared gameplay from AA/AAA studios and custom environments, aligned frame-by-frame with player inputs, telemetry, camera state, and engine G-buffers.
- 50K+
- Hours
- 20+
- Environments
- Engine-Level
- Signals
03 Simulation
Midcentury Matrix
Agentic simulation platform to design, test, and train physical AI across massively parallel cloud environments.
01 Design
Build digital twins, environments, and scenario distributions around real deployment conditions. Combine classical simulation with learned rendering and manipulation physics for higher sim-to-real fidelity.
02 Test
Massively parallel GPU evaluation across thousands of closed-loop scenarios. Sweep conditions, compare builds, replay failures, and isolate regressions before returning to hardware.
03 Train
Turn failures into new scenarios and training experience. Feed simulation and real-world data back into policies so systems continuously improve through interaction.
04 Research
Our contributions to physical intelligence.
Steam. Closed-loop evaluation for spatial intelligence. A purpose-built benchmark of interactive 3D environments that measures navigation, object interaction, spatial reasoning, environmental understanding, and planning directly from simulator state. Releasing soon.
Giacometti. Action supervision from human video. A data engine that turns raw egocentric video into training-grade physical interaction data, including metric 3D hand trajectories, camera motion, depth, and task structure. Releasing soon.
Neural Renderer. High-fidelity rendering from engine-level supervision. A learned renderer trained on gameplay data paired with depth, normals, albedo, motion vectors, and camera state to map structured simulator state into realistic visual observations. Releasing soon.
