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
Physical AI produces impressive demos. Yet very few robots ever reach the real world.
Scaling robotics now requires:
Human action data at internet scale for pretraining.
Reliable simulation for evaluation and RL.
Midcentury builds both.
The next industrial revolution begins when robots leave the lab.
02 Datasets
High-Fidelity Data for the Real World
Large-scale robotics and simulation data, paired with rich action supervision and dense annotations for the next generation of physical AI.
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.
- 2M+
- Hours
- 50+
- Environments
- 3D Pose
- + Annotations
Custom Gameplay Environments
Custom, on-demand gameplay data from environments built to spec — any camera angle from first-person to top-down — with frame-aligned inputs, telemetry, camera state, and engine G-buffers.
- 50K+
- Hours Supported
- 100+
- Environment Types
- Engine-Level
- Signals
Conversational Voice
Natural multilingual conversation captured in full-duplex, multi-channel settings. Speaker-separated audio, transcripts, and metadata built for speech and conversational models.
- 69K+
- Hours
- 25+
- Languages
- Multi-Channel
- Audio
03 Simulation
Midcentury Matrix
Agentic simulation platform to design, test, and train physical AI across massively parallel cloud environments.
01 Design
Digital twins from real deployment conditions, combining classical simulation with learned physics.
02 Test
Massively parallel GPU evaluation across thousands of scenarios. Replay failures, catch regressions before hardware.
03 Train
Turn failures into new scenarios and training experience. Feed real-world data back into policies.
04 Research
Our contributions to physical intelligence.
MC-EgoHands. SOTA 3D motion reconstruction from egocentric video. Sub-cm hand tracking, metric SLAM, depth, and world-space trajectories from raw video. Releasing soon.
MC-Shade. Real-time, engine-agnostic photorealistic rendering for any simulation. Turns any engine's geometry, depth, and material buffers into photoreal frames in constant time, closing the visual sim-to-real gap without touching the simulator's physics. Releasing soon.
MC-PhysBench. The first long-horizon physics validity benchmark for world models. Tests every major world model against exact, MuJoCo-simulated ground truth, and catches physics failures within seconds of an eight-second rollout across five calibrated failure modes. Releasing soon.
