Midcentury is launching Akari, a learned renderer that produces camera images from classical physics engines for teams training and evaluating visual robot policies. A policy that recognizes an object in a clean simulated scene still has to recognize it under glare, in shadow, or through a different camera, and that visual mismatch is a practical limit on deployment. Akari is the company's first learned model aimed at that gap, and it will be available only inside Matrix, Midcentury's simulation platform for designing, testing, and training policies across many parallel scenarios. Akari reads simulator G-buffers (surface color, orientation, depth, roughness, and metalness) plus the previous frame, then generates the next image. Generation does not start from an engine-rendered color frame. Objects, geometry, camera pose, and physical state stay with the physics engine, so the observation model can change without handing task control to a video model. Midcentury trained Akari in-house. Brandon Samaroo, head of research, led the work. On two warehouse sequences, the company reports mean LPIPS of 0.046 and 0.052 against Unreal Engine renders across 1,000 autoregressive frames. Warm inference averages 0.46 to 0.47 seconds per frame on one RTX 5090 with four denoising steps. Those figures measure reconstruction of engine renders, not photographs from real robots. Midcentury says the next step is learning from real observations and testing whether the images change robot behavior. The same report flags remaining gaps in small surface markings and frame-to-frame consistency.