Tesla May Be Adding A Radar, Which Is The ‘Crutch’ They Definitely Need

A Tesla which hit a police cruiser on the side of the road while in Autopilot mode. These crashes … [+] are under federal investigation. (Laguna Beach Police Department via AP)

ASSOCIATED PRESS

A recent application by Tesla to the FCC hints they are planning to add a radar to their cars , or rather re-add since they removed the radar they had earlier. This would be an HD or “imaging” radar, which is quite a bit superior to the radar they removed.

When Tesla removed the radar, they stated it was because they wanted a car driven with purely computer vision. Later, they also removed the ultrasonic short-range sensors from new cars to fully commit to that. They claimed that by avoiding the challenge of sensor fusion — trying to combine different sensors looking at the same object — they would have a cleaner and more reliable perception system for their cars, which dream of “full-self-driving” some day.

Many in the industry were skeptical of this, and felt the decision to remove radar was simply because the radars were hard to source during the supply chain crisis, and in fact holding up the shipping of cars. This was supported by the fact that after Tesla removed the radars, they had to disable important functionality provided by the radar, and only much later did they restore most of it. In fact, many Tesla drivers report that the switch to vision only resulted in an increase in “phantom braking” where the vehicle suddenly slows because it thinks it sees something (that isn’t there) up ahead. When ultrasonics were removed it also removed a number of important features like park assist and automatic parking.

Tesla’s old radar was a fairly old design and not that capable. The new radar would be an “HD” radar which makes a low resolution image of the radar targets ahead of the vehicle. Like all radars, it not only gives a precise distance to the targets but detects their speed towards or away from the vehicle. On older radars, that speed measurement is crucial, because the resolution is so low you can’t tell a car from a sign next to it or above it, except for the fact the car is moving. Users of such radars tend to be forced to ignore returns from stationary obstacles because the whole world is made up of those.

An imaging radar is, in fact, a lot like the LIDAR that Elon Musk has referred to as a “crutch” that dooms the teams that use it. It’s lower resolution to boot. It has a couple of advantages, though:

Frame from Waymo imaging radar on foggy day

Waymo

Many other self-driving teams are working with imaging radars. Earlier, Waymo released some details of their use of the technology. In Waymo’s video you can see how low resolution the signal is. The best ones I have seen claim 0.5 degrees, and most have a one degree horizontal resolution. Typical LIDARS do 0.1 degrees — 10 times better.

It’s hard, though, to imagine why LIDAR would be a crutch and HD radar would not. Most teams wan’t both, and right now Tesla wants neither.

Such a radar would help them with several well known problems. Tesla is under investigation for the way that Teslas in autopilot have repeatedly crashed into emergency vehicles (and others) on the left shoulder which stick out into the lane. Regular radar sees these, but can’t tell them from the guardrail. An imaging radar (or LIDAR) would tell these more easily. Computer vision also hopes to see them, but clearly still limps enough to need that crutch.

This imaging radar (not from Tesla) claims 0.5 degree resolution with its special antenna.

Lunewave

It also would help with phantom braking. Radar gives you a much more certain reading on whether there’s a car in front of you, particularly if it’s moving at all. False negatives (which cause crashes) are the worst, and it would bring those down, but it would also reduce false positives (phantom braking) as well.

If confirmed, this would also suggest that Tesla’s new effort at a “virtual LIDAR” based on vision, which it calls its “occupancy network” is not showing enough reliability. The human brain is able to look at moving images and figure out how far away everything it sees is. Neural network tools can also do this, but they don’t have the power and reliability of the brain in doing that. LIDAR and imaging radar don’t have to guess or figure things out with all the power of a brain — they read that information directly.

This is not as big a reversal for Musk as it might seem. While he has regularly dissed LIDAR, he has been more open to the idea of imaging radar, particularly because it doesn’t use light and thus gains even more superhuman abilities than LIDAR does in many ways. While they would have to walk back their condemnation of sensor fusion, it’s otherwise a reasonable choice based on what they have said in the past. They would have to retrofit all the cars which pre-bought the “FSD” package, but such radars only cost a few hundred dollars in bulk, and there is already a mounting point in the cars. If you paid $15,000 (or even much less) they can afford that retrofit. It is also expected they will need to replace the cameras with new, higher quality ones and even replace the processor — something they already did once for many owners.