A discussion about the benefits of radar should account for how modern AI based on neural networks makes decisions. Understanding this clarifies Elon's comment about not needing radar and thinking probabilistically.
The NN AIs in Teslas are basically big probability calculators. The decision making in NNs is probability based, not binary (e.g. not check under the car in front and see if another car is slowing down). NNs have weights (coefficients) applied to each input (pixel) of each sensor (camera, radar, etc) which generate outputs that are fed into more layers that have weights that effectively use input to predict what the proper objects or behavior should be.
How to calculate the coefficients and create the models is where the magic of machine learning lies, but the point is that Tesla can assess the contribution provided by the radar to the accuracy of the system's predictions. So it's all about how much improvement there is in the accuracy of predictions, rather than specific situations because predictions are a spectrum of probabilities.
Arguing that radar handles certain cases is binary thinking because it implies that a "case" is always recognizable (TRUE) to begin with, but NNs assign probabilities to even recognizing the scenario (e.g. 82.35% probability). Also, this argument does not include ROI.
Let's look at the ROI argument with completely made up numbers. If radar ($100) improves NN predictive accuracy by 5%, but adding two more cameras ($20) improves NN predictions by 10%, Tesla is better off installing two more cameras instead of using an expensive radar. The ROI argument immediately says that radar is not worth it and extra cameras are a better choice.
Now let's look at the probability argument for 2 cameras and whether a 10% improvement makes sense for $20. If prediction errors are reduced by only 10% with two cameras, maybe accidents would drop from 10,000 to 9,000. That sounds like a lot, but autonomy requires NNs reduce errors by an order of magnitude for 5 iterations. In other words, reducing errors from 10,000 -> 1,000, then 1,000 -> 100, etc until it's .1%. It's clear that 10% doesn't make any difference in getting to autonomy, because the rate of improvement through better NN models and training needs to improve so much more that a one time 10% improvement would not even be noticed compared to the 100,000% improvement through the NN.
I'm not saying my made up numbers are right. However, Tesla does know what the numbers are. Given the priority Tesla has placed on safety, I'm pretty sure that they would not drop radar if it added material improvement relative to what FSD software is achieving.