autopilot issue with non-moving objects

[tslaEbay]

Hi All...there was one post on Tesla forums few days back on autopilot having issue with non-moving objects which resulted eventually in some accident. That post was mysteriously deleted (may be somebody reported as spam) after few days.

I do know that AP1 cars have this limitation. I read somewhere that the current radar based system doesn't differentiate a static sign board and a stalled car on your lane. Tesla autopilot can track a moving object very well once it is locked in.
In the case of some stalled vehicle in your lane...what will happen eventually?
Doesn't at least AEB kick-in in the last minute? But depending upon the speed it might not avoid the collision fully?

What about current AP2? do we still have the same limitation?
I'm sure EAP will solve this eventually. Does anybody know how exactly solve this issue?

 

[Tam]

...In the case of some stalled vehicle in your lane.....

Tesla has fine-tuned its radar software after the Florida fatal autopilot accident and guessed that the system would likely be able to deal with that scenario.

However, I believe the current radar technology and its software have not solved the problem of stationary vehicles including those that partially intruding into your lane.

That disclaimer about stationary vehicles is still in the owner's manual and has not been taken off just yet.

Of course Tesla is still working on this or otherwise how can its driverless program will deal with this kind of scenarios?

...Doesn't at least AEB kick-in in the last minute? But depending upon the speed it might not avoid the collision fully?

You seem to misunderstood the function of Tesla AEB because you see other companies' ads that show clearly their cars automatically stop to avoid a collision.

Those ads are meant for low city speed, not highway high speed.

On the other hand, Tesla owners' manual clearly states that its AEB designed is not to avoid a collision. It is designed to allow a collision to happen. But with a reduced speed and reduced force of collision.

...What about current AP2? do we still have the same limitation?...

Yes, AP2 still has the same problem. My guess is it's a little better than AP1 for now. I have driven almost 4,000 miles for the past 3 months and I had to manual brake for stationary vehicles about 3 times.

...Does anybody know how exactly solve this issue?...

1) Cameras:
AP1 has only 1 camera that is monochrome so it had a problem of differentiating a Florida accident bright background and the white trailer.

Some suggest that if you introduce full spectrum color to your camera, that may solve the Florida accident contrast confusion problem.

Some suggest more cameras so if one angle misses the obstacle, another one can pick that up: Thus, AP2 has tri-focal cameras.

2) Tesla Vision:
Image recognition software will have to be improved so your car can avoid various obstacles including traffic cones...

3) Radar:
Radar can pick up all obstacles but its resolution is so poor so it may see a very small harmless concave soda cap that magnifies its signal as a gigantic obstacle. That has been a problem since in WW 2 when aluminized mylar (plastic) Chaff were recognized by radar as countless air planes flying in the sky.

Tesla thinks it can use software to deal with that problem: Ignore a stationary soda cap and not a stationary vehicle.

4) Lidar:

Pro-Lidar says Radar resolution is so poor so you need to supplement it with Lidar. Its resolution is so high that it can see a bicyclist arm signal as well.

5) Where are we?

I love AP2 because something is better than nothing but it is still at its infancy so it still has lots of work to do.

Google/Waymo has demonstrated that it has not gotten into any serious/fatal accidents or traffic violations so far so I would trust its safety more than Tesla's.

My problem with it are:

1) private beta: only its employees can use its beta program while Tesla allow you to use theirs for a fee.

2) no independent verification of accident avoidance testings: It has not released a test that could deal with a Florida fatal Autopilot accident.

3) Launch date is still around 2021.

In summary, it's great to be enthusiastic, but it is prudent to realize current limitations and progress status.

 

[BrettS]

Hi All...there was one post on Tesla forums few days back on autopilot having issue with non-moving objects which resulted eventually in some accident. That post was mysteriously deleted (may be somebody reported as spam) after few days.

I do know that AP1 cars have this limitation. I read somewhere that the current radar based system doesn't differentiate a static sign board and a stalled car on your lane. Tesla autopilot can track a moving object very well once it is locked in.
In the case of some stalled vehicle in your lane...what will happen eventually?
Doesn't at least AEB kick-in in the last minute? But depending upon the speed it might not avoid the collision fully?

What about current AP2? do we still have the same limitation?
I'm sure EAP will solve this eventually. Does anybody know how exactly solve this issue?

I’m not sure this is totally accurate. Autopilot doesn’t have an issue with stationary objects, but rather there are two things at play here. First, autopilot and the radar system is designed to look for cars (and pedestrians) and not necessarily signs and traffic cones and such. There have been a number of examples where autopilot has driven cars into traffic cones in construction zones, but this isn’t because the traffic cones are stationary, but it’s rather that the lines on the road were poor due to the construction and they guided the car into the cones, which the radar does not see. This is another example of why you need to keep paying attention as autopilot is engaged.

The other issue is the limited range of the radar... it’s possible that if you are traveling at highway speed and come up on a car stopped in your lane that even though the car will attempt to stop it may not have enough time. Autopilot will see a stationary car as well as it sees a moving car, but it may not see it in time to stop.

I often drive with autopilot engaged and I have tested the system when I see traffic and stopped cars coming up ahead of me. At times I will allow autopilot to stop the car if I’m going 45 or 55 miles an hour and it is definitely able to do so, but it stops much later and therefore much harder than I would do, simply because it doesn’t see the stopped car as soon as I do.

 

[Tam]

...First, autopilot and the radar system is designed to look for cars (and pedestrians) and not necessarily signs and traffic cones and such. There have been a number of examples where autopilot has driven cars into traffic cones in construction zones, but this isn’t because the traffic cones are stationary, but it’s rather that the lines on the road were poor due to the construction and they guided the car into the cones, which the radar does not see...

Traffic cones are not painted with metallic radar-signal-reflecting materials so radar does not detect them. So it is true that radar is not designed to detect traffic cones and not because they are stationary obstacles.

However, that does not excuse Tesla from neglecting designing a basic driver education capability of avoiding traffic cones.

nVidia demo below shows that in the beginning, its system would hit traffic cones over and over again but it learned as time went by and it successfully avoided traffic cones afterward, all on its own:

In this case, the system didn't use radar, it only used camera just like the same way as Tesla driverless video demo does (without radar).

So current excuse that radar has limitations is not valid because AP2 has competent tri-focal camera (Narrow Focal can see 250 m, 820 ft or more than 2 football fields).
.

AP2 radar range is 160 m (525 ft) which is almost 2 football fields (2x 109.1 m or 360 ft) so advanced early warning data is there but how to timely make sense of it is the problem.

In summary, the hardware is there but how to make use of it is still a problem.

 

[tslaEbay]

@Tam, @BrettS..thanks a lot for the info. Really appreciate it. Yes. Elon is betting on vision as opposed to Google using Lidar.
Time will tell.

 

[TOBASH]

Bottom line... Keep yer hands on da wheel!

I just swapped from a BMW X6 M package to the Model X. The adaptive cruise control on the BMW just plain SUCKS. It accelerates toward other cars, it doesn't slow down when cars merge ahead of you, it turns off suddenly at speed if you hit a small bump...

The Model X is vastly superior. It seems to understand merging traffic and keeps a comfortable distance without trying to accelerate into the rear end of other cars.

I would guess that additional cameras and lidar and radar are being developed. That being said, these vehicles are not sentient beings, although I'm super happy with the system as compared to my former vehicle.

Best,

T

 

[Nerdy_Engineer]

I think part of the reason why Autopilot doesn't seem to react to stopped cars until very late (sometimes not at all if partially blocking a lane) is that the software hasn't been trained for city driving yet. I think the autopilot engineers have been focusing on interstate driving where you don't expect to come up to completely stopped cars. Usually on the interstate if traffic stops, you'll have cars in front of you slowing down to a stop. Autopilot handles that situation perfectly. It only seems to have trouble when you don't have a car in front of you and you come up on a stopped car. I'm sure with AP2 (hopefully AP1), this will improve with time as the autopilot engineers train the system for non-interstate driving. That's my theory at least.

 

[adwang]

With 2018.26 updates recently, my S90D seems to be able to "SEE" stopped car in front while in autopilot/cruise mode. Though not 100% sure this is always true, but, on my regular route, at traffic lights, with new software, the car reacts to stopped car in front much earlier, similar to what a person would do.

 

[skip8jj]

Yes, 2018.26 has improved driving manners. Reaction to cars changing lanes in front of the Tesla are much smother. It senses if the intruding car is moving faster and suppresses a knee-jerk reaction now. Much smoother all round in traffic. Hope we are able to download the coming version before we leave for NM in September.