Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

How will Full Self Driving deal with turning into perpendicular fast-moving traffic?

This site may earn commission on affiliate links.

JoshG

Member
Jun 23, 2011
440
443
Shouldn't Tesla have two more cameras that point directly left and right, ideally from somewhere close to the front bumper?

With the current set of cameras, there doesn't seem to be a view perpendicular to the direction of travel, pointing left and right of the vehicle.

A FSD car approaching a stop sign where the traffic coming from left and right doesn't stop and has a high speed limit, say 45mph or higher has to be well aware of traffic coming from left (if turning right) or both sides (if turning left). There aren't cameras for this, as far as I know. So how will the car know it has enough time/space to safely make it's turn and accelerate to the speed of the traffic?

This image, if it's accurate, backs up my worry: https://tr1.cbsistatic.com/hub/i/r/...b7eb2ea3266a5d1dd7828da3/autopilotsensors.png . The wide front cameras have a range of 60m and a field of view that looks to be about 120 degrees wide, so there's no view directly left or right. Rearward-looking side cameras (B-pillar) clearly can't see left/right cross traffic.

Near my home, there are several T-junctions where the cross traffic is very fast and doesn't stop. It's dead simple for us humans to look left and right to gauge the cross traffic. I'm just struggling to see how an autonomous or mostly-autonomous car can do this without 90 degree cameras pointing left and right, ideally from as far forward (and as high) on the car as possible to get the best view. Obviously it's such a common situation that Tesla has found a way to address it, but it doesn't seem the forward camera or B-pillar mounted cameras provide the coverage, nor do the side sonar sensors.

Thanks for any insight.
 
  • Like
Reactions: beachmiles
Forward looking side cameras 80M/260'. Is that not far enough?

EDIT: Also all cameras overlap significantly.

Screen Shot 2019-04-29 at 3.13.48 PM.png
 
Last edited:
  • Like
Reactions: M109Rider
Apologies, i misread that chart - it does seem the "forward looking side cameras" can cover the 90 degree sideways view. I thought that pale yellow was (not to scale) ultrasonic. Excuse the question... should have done more research before posting.
 
Forward looking side cameras 80M/260'. Is that not far enough?

EDIT: Also all cameras overlap significantly.

View attachment 402193

Let's do some math:
60 mph (fast traffic) is 88 feet/sec. That means that 80m range gives the car just under 3 seconds of time before the cross-traffic reaches it. That seems pretty marginal. (60 mph is admittedly fast, but I've seen some cases where this exact scenario would have to be supported).

My question would be where the 80m comes from - is that minimum visibility, or a real maximum. In general these fast merge environments have pretty good sight lines, so maybe the system can exceed these specs under good conditions?
 
  • Like
Reactions: beachmiles
To be honest I share these concerns somewhat. While I assume that cameras will be fine for FSD, the coverage seems weak in some angles. Imagine the car creeping slowly into a fast moving traffic street while cars are parked on either side of the Tesla. It then has to creep quite far into the street for the B-Pillar cams to see perpendicular traffic. Human heads are further forward than the B-Pillar cams and can gimbal to either side.

Incidentally that is a situation that I face every morning, and even with eyes I have to creep dangerously far into the residential street from my garage to make out perpendicular traffic.

In these situations it might make sense to have side cameras at the corners of the car frame.
 
  • Like
Reactions: beachmiles
Let's do some math:
60 mph (fast traffic) is 88 feet/sec. That means that 80m range gives the car just under 3 seconds of time before the cross-traffic reaches it. That seems pretty marginal. (60 mph is admittedly fast, but I've seen some cases where this exact scenario would have to be supported).

My question would be where the 80m comes from - is that minimum visibility, or a real maximum. In general these fast merge environments have pretty good sight lines, so maybe the system can exceed these specs under good conditions?


To be fair- I'd like to see the intersection a model 3 couldn't get across in under 3 seconds if it had to.
 
I have this situation several times every week. Divided cross-traffic flowing at 55-65 mph and I have no merging lane available. It's a long stretch turning your head enough to see over your right shoulder and through the back-side windows to look for traffic.

Wickenburg.jpg
 
  • Like
Reactions: beachmiles
To be fair- I'd like to see the intersection a model 3 couldn't get across in under 3 seconds if it had to.
It's not about how fast you can make it across it's about being able to see all the vehicles and not get hit in the process. In instances of merging into traffic you're not trying to cross anything, you're trying to get into on without getting asspacked.
 
I have this situation several times every week. Divided cross-traffic flowing at 55-65 mph and I have no merging lane available. It's a long stretch turning your head enough to see over your right shoulder and through the back-side windows to look for traffic.

View attachment 402215

At that angle of entry, FSD would be utilizing the Rearward-looking side cameras. But it has to be smart enough to know about that angle and to use those cameras.

That is also horrible road design. Doesn't minimize the point about FSD being ready to handle strange situations though.
 
To be honest I share these concerns somewhat. While I assume that cameras will be fine for FSD, the coverage seems weak in some angles. Imagine the car creeping slowly into a fast moving traffic street while cars are parked on either side of the Tesla. It then has to creep quite far into the street for the B-Pillar cams to see perpendicular traffic. Human heads are further forward than the B-Pillar cams and can gimbal to either side.

Incidentally that is a situation that I face every morning, and even with eyes I have to creep dangerously far into the residential street from my garage to make out perpendicular traffic.

In these situations it might make sense to have side cameras at the corners of the car frame.

This is a great point, I also share the concern that the B-Pillar cams might not be able to see the perpendicular traffic if there are parked cars along the side (a very common scenario).

I am very curious as to how Tesla is going to handle this.
 
The car takes like 3 seconds just to start to SLOWLY move forward while on TACC in heavy traffic. I don't expect it to confidently dart across an intersection anytime soon but I guess we'll see. The OP raises a great point about the camera limitations in this scenario.
Keep in mind this is APH2.5/standard rules based code. APH3 with neural code is a completely different animal (just looks the same on the outside :D ) and should put the FSD in the "big leagues" of performance and "decision making".
 
Last edited:
Shouldn't Tesla have two more cameras that point directly left and right, ideally from somewhere close to the front bumper?
I also share the concern that the B-Pillar cams might not be able to see the perpendicular traffic
if there are parked cars along the side (a very common scenario).
I'm exactly in this situation when I exit from my garage and turn right to enter into the street.
I have to literally put my face almost against the windshield to try looking at the coming cars.

The front of my car in already inside the moving lane, which can be an issue with some trucks
or bicycles and sometime I need to backup.

I'm considering installing some side cameras under the front license mount connected to a display mounted in front of me.

This is a very difficult manoeuver, and I really wonder how I could make my car working as a robot taxi.
 
All good questions and comments. I think the camera coverage is adequate but the range concerns me a bit because that doesn't give the car much time to "decide" and go.

Of course, if the neural network is designed to make our cars "think" like humans, maybe our cars will just pull into traffic and expect the other cars to slow down or dodge us. I see a lot of humans that drive that way! :)
 
While we are on the subject, how is FSD going to handle:

1) Ambulances and fire trucks that need right of way
2) School zones that have restricted speeds while children are present
3) A police officer flashing lights behind it indicating pull over if it violated some law.