Midcentury
Midcentury builds the data and simulation infrastructure required to move physical AI systems from lab demonstrations to real-world deployment. The company offers large-scale egocentric vision datasets with rich action supervision, custom gameplay environments, and the Midcentury Matrix platform for massively parallel cloud-based training and evaluation. Its research tools include MC-EgoHands, a 3D harness that converts first-person recordings into structured robot-learning episodes.
- Robotics
- Artificial Intelligence
- Developer Tools
- Software Only
Funding
Founders
Founder details are not available yet.
Product
Problem
Physical AI models often perform well in demonstrations but rarely succeed in real-world deployment. Training capable robots requires both human action data at internet scale for pretraining and reliable simulation for evaluation and reinforcement learning — resources that are fragmented, low-quality, or unavailable to most robotics teams.
Solution
Midcentury builds the data and simulation infrastructure needed to move robots from lab demos to real-world use. The company provides large-scale egocentric vision datasets with rich action supervision and dense annotations, plus custom gameplay environments built to spec with frame-aligned inputs and telemetry. Its Midcentury Matrix platform enables agentic simulation across massively parallel cloud environments for designing, testing, and training physical AI. The company also develops research tools like MC-EgoHands, a 3D harness that converts first-person recordings into structured episodes for robot learning workflows, and MC-PhysBench, a benchmark that measures how long video world models remain physically valid.
Target Audience
Primary customers are robotics companies, AI research labs, and enterprises developing physical AI systems that need large-scale human action data and reliable simulation environments for pretraining, evaluation, and reinforcement learning.
Features
- Egocentric vision datasets spanning 2,000,000+ hours of first-person video with synchronized IMU, hand pose annotations, and task metadata across 50+ environments
- Custom gameplay environments built to spec with any camera angle, paired with player inputs, camera state, telemetry, and engine G-buffers across 100+ environment types
- Midcentury Matrix, an agentic simulation platform for massively parallel cloud-based training and evaluation of physical AI
- MC-EgoHands harness with source adapters for EgoDex and HoloAssist, exporting to LeRobot v3 episode format and Panda/LIBERO/Yam robot target layouts
- MC-PhysBench physics benchmark with seven failure checks, survival curves per model, and calibrated measurement against a physics reference
- Support for monocular and calibrated stereo processing paths with MP4, IMU JSON, and MCAP input formats