Data and simulation infrastructure for physical AI
Physical AI produces impressive demos, yet few robots reach the real world. Scaling robotics requires human action data for pretraining and reliable simulation for evaluation and reinforcement learning. Midcentury builds both.
Data
Egocentric VisionOver 2 million hours of first-person video of real work across industrial and everyday environments. Video, IMU, and audio are enriched with signals including 3D hand pose, point tracks, depth, and task annotations. |
UMI DataHead and wrist recordings of gripper tasks across a range of environments. The dataset includes RGB and stereo camera views, IMU, fused head and tool-center-point pose, and gripper state. |
Research
SortexA data quality platform for screening, annotating, and filtering egocentric video for specific training needs. |
AkariA learned renderer for robotics simulation, trained to bring photorealistic rendering to classical physics engines. |
MC-PhysBenchA benchmark for measuring how long video world model predictions remain physically valid. |
MC-EgoHandsA 3D harness that turns first-person recordings into structured episodes for robot learning workflows. |