Tesla will probably return radars to cars for Full Self-Driving starting next year

A great deal of debate broke out about removing radar from cars. We all know cases where Teslas have crashed into white semi-trailers or cars with beacons. And here we come to a paradoxical situation. Radars stopped being fitted to Tesla cars in May 2021, but most of these methods happened before that date, so they happened to cars equipped with radars (and the remaining ones just after that transition). It can be seen that the mere presence of a radar does not guarantee that a car will “see” a car with a beacon. On the contrary, it doesn’t seem that the number of these accidents has increased dramatically recently, when all new (and today, after the updates, probably also old) Teslas on the roads only drive with a camera system.

When Musk announced the removal of radar from cars, he made an argument that may not sound intuitive. According to him, the radar made the whole system more error prone. However, it is well known that the more measurements you have, even if they are imprecise, the more accurate the result you will usually get by averaging them. You know this system, for example, from cameras and telephones, where noise is reduced by multi-exposure. Multiple noisy photos are taken, they are overlapped, and the noise is largely eliminated. Something like that should intuitively work here as well. A few inaccurate (eg noisy) measurements from the cameras, a few inaccurate measurements from the radar, and in theory we should have a great idea of ​​what’s in front of the car.

And here Elon Musk suddenly told us that it’s worse? This doesn’t make much sense at first glance. However, from my own experience of working on 3D reconstruction from stereo cameras, I can say that pairing classic cameras with depth cameras, which should theoretically improve it, does not necessarily lead to a more accurate result. Exactly opposite. An additional sensor can work in such a way that it improves detection in some situations, but at the same time quite often “breaks” something that already worked without it. It often happens that the next type of sensors will bring correct data where the original solution did not work (where the sensor did not see it correctly), but on the contrary, it reports wrong data where the original system did not have a problem. And now it’s a question of whether the benefit is greater than the downside. And it looks like that was the reason Musk had the radars removed. While this certainly helped some situations, according to him, the implication was that it actually made many situations worse (it worsened the signal-to-noise ratio). After all, it also corresponds to my experience with adding sensors, which did not always lead to what we would intuitively expect.

While Musk has fully condemned lidar in the past, he has not been so critical of radar, saying even then that while existing radar makes the situation worse with its noise, cameras combined with high-resolution radar should be better than a purely camera solution. The problem was that, according to him, such a sufficiently high-quality radar did not exist, so he preferred a solution without radars, which was more reliable than with it. Now, however, it seems that Tesla finally has a radar with high resolution and low noise available, so we will probably see it already in cars assembled from January of this year. However, you should not use the new component in practice yet. This will come with the next software versions that will be ready for this sensor. In any case, it does not change the fact that autonomous driving depends primarily on the computing part, i.e. understanding what the sensor actually sees.