Akari is a learned neural rendering model that produces photorealistic frames directly from physics-simulator state buffers. It closes the sim-to-real gap for robotics by delivering 0.046 LPIPS fidelity at 0.5 seconds per frame across long sequences, without relying on traditional ray tracing.
Akari is Midcentury’s learned renderer for classical physics engines. It takes structured simulator render passes, including surface color, orientation, depth, roughness, and metalness, and uses the preceding image as context to generate subsequent frames. Midcentury reports warehouse sequence evaluations spanning 1,000 frames and inference measurements of about 0.46–0.47 seconds per frame on one RTX 5090. The reported image comparisons are against Unreal Engine renders, not photographs. Akari is intended for use through Midcentury’s Matrix simulation platform.
