This is more relevant than you might think. Data is never perfect and you have to program assumptions to work around it's short comings, but as you expand the range of assessments you want a system to make you can have scenarios where those assumptions start to conflict.I see people pondering over 'sensor disagreement' as if Tesla are struggling with something completely new and unexplored. Fact is it's nothing new at all. Sure an autonomous car for public roads is a new(ish) application, but complex autonomous systems with multiple sensors have been around for decades.
Here's an example.. You're driving down the road following a vehicle. Your radar sensor is happily reporting that vehicle is moving at 70mph. Suddenly you get a reading that says the vehicle is stopped. Do you trust that and slam on the brakes or do you filter it out? After all, vehicles can't literally go from 70mph to stationary instantly, can they? What you saw is probably a bridge over the road (traditional radar has very limited vertical resolution, remember). 99% of the time you'll be correct.
Now lets say you're driving along behind a 70mph vehicle and a tree falls in to the road due to freak wind. Your radar sensor reports that a vehicle decelerated in a hugely improbable fashion so you throw the sensor reading out, and as a result hit the tree.
The point is not that the specific scenario I made up is a problem, but the general case that the more nuanced you expect your system to be the less you can afford to smooth the data because you start wiping out genuine data points. It could be that every emergency braking radar on cars today would fail the 'tree' test above but we never find out because such an occurrence is freakishly rare (and 99% of cars don't have forward facing cameras recording all the time for internet dissection), and other manufacturers are maybe happy to live with these limitations because they don't also intend to build a self driving system that SHOULD be able to spot an object has fallen in to the carriageway and navigate around it.
Andrej Karpathy has a segment on this in this video: