https://light.princeton.edu/publication/neural-nano-optics/
This looks like a really interesting breakthrough -- using a mix of machine learning and computational photography with custom optics to create absurdly tiny cameras, with a single etched optical layer replacing the usual stack of lenses.
The paper doesn't address the computational requirements of the image processing or the size of the required model, so it's unclear how small an autonomous system it could be on.
However, in many applications the image data could be processed remotely.
This essentially allows the incorporation of extremely small and difficult-to-spot cameras into just about anything. It could be useful for everything from smartwatches to dragonfly-sized drones to extra-tiny endoscopes to networks of "smart sand" that could be scattered around a space to observe everything that happens within it.