Funny, insurance companies seem to feel differently about whether "testing" human driving behavior *while driving* is interesting or useful. Once built, the car engine/computer should be much more deterministic over the range of driving behaviors/variables that you want to map exhaust to, versus those silly, unpredictable humans.
Is simple: test while driving, and just sample enough cars in enough areas. Testing thingy clamps onto exhaust, stores data, and measures whatever you want to measure, along with other interesting things like, gear, RPM, temperature, location (for congestion mapping or whatever), etc. Maybe you can get some of that info directly off CAN, maybe you have to plug into a diagnostic port and sync time. Whatever.
> Testing emissions outside of laboratory settings is like testing the levelness of a table on a boat at sea. It's much less precise and reproducible.
Which is, well, kinda useless. The levelness of the table is something that's determined by how it's built - an innate property it has. Whether or not it slides around and, say, damages the floor on a boat at sea under different speeds of that boat, wave conditions, and the like is really more like what we're trying to consider with a car on the road.