AI Now Designs Camera Lenses Better Than Existing Tools

Designing a lens is a brutal puzzle. You have dozens of curved glass pieces, each with thickness, shape, and coating choices, and turning one dial ruins the others. Today it takes expert opticians with decades of intuition, plus weeks or months of trial and error, to get it right. A team of researchers has now built LensDesigner, an AI "agent" (software that takes actions on its own) that copies how those human experts actually work.

The hard part is the starting point. Most AI design tools flail because there are too many possibilities. So the team built LensLib100K, a library of 100,000 lens designs, and gave the AI a way to pull out promising, physically realistic starting shapes. From there, the AI works inside a physics simulation, getting instant feedback on how each version performs. A second AI acts like a coach, handing it progressively harder assignments. Over time, LensDesigner writes its own notes — what worked, what didn't — and reuses those tricks on the next job, much like a student moving from easy homework to hard homework.

To prove it works, the team created LensArena, a test set of 120 different lens design tasks, including extreme configurations that are notoriously difficult. The AI beat publicly available baseline algorithms on both success rate and speed. The code will be released publicly, meaning other researchers and companies can build on it.

The honest catch: all of this happens in simulation. Real lenses must survive manufacturing tolerances, glass availability, cost limits, and assembly errors. A mathematically perfect design can still be impossible or too expensive to actually build. So don't expect human opticians to disappear — expect them to spend less time guessing and more time checking AI proposals. If it holds up in the real world, this kind of automation could speed up everything from smartphone cameras to AR headsets, car sensors, and medical imaging devices.