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

A cautionary autopilot video

Do you use autopilot on secondary roads (not highways)?


  • Total voters
    208
  • Poll closed .
This site may earn commission on affiliate links.
If you mean TACC, yeah I use it a lot. If you mean auto steer, no I have not used it anywhere any time in the past couple of months. I was starting to call it auto awake cause if you weren't paying 100% attention it would pick that instant to do something real scary and now I was really awake.
 
Posting this short video (5 sec) captured by the dash cam as a cautionary note for those who use the Tesla autopilot on secondary roads. The video shows my almost hitting a pedestrian dressed in dark at night standing still in the middle of my lane. The autopilot did not detect the person, but fortunately I did.


Why did it take you so long to brake? Finally watched this on a full screen, and with the road being quite well lit (with other brake lights and stuff), the pedestrian was visible quite a long way in advance, actually (the key thing to me being the brake light of the car in front being obscured - that would mean immediate braking for most people). I can tell when you start to brake by when the camera angle changes; it is significantly after (1-2 seconds) the pedestrian becomes clearly visible in the camera, and as others have noted the actual image most people actually see with their eyes (obviously some people have night vision issues and that would explain a delayed response) at night is much more clear than what shows up on the dash cam.

Or did you see the pedestrian and judge that you had plenty of time to stop, so you waited to see what the AEB would do? Did you have your foot on the accelerator at all?

Why only 5 seconds of video? I'm curious to see about the lead up (what were other cars doing - it looked like the car in front had maybe dodged the pedestrian - or maybe it came from a side street).

Anyway, still stunning poll results. Frightening how many people use EAP in places where pedestrians/bicycles may be present (where use is specifically prohibited)! Though EAP had nothing to do with this scenario of course (from what I understand, behavior in this case should be the same for FCW/AEB from a car where EAP is not in use, since safety features are enabled on all vehicles).
 
Please send this to Tesla... they need to see it... I'm sure the response is (of course) that your supposed to disengage and your not supposed to use on secondary roads... but that is a close enough call that you should give them the benefit of seeing it.

If I were an engineer there, or an exec --I'd want to have seen it, and share it internally.

As a TM3 owner, I greatly appreciate your sharing that. It's a healthy reminder. Thank you.
 
  • Like
Reactions: XBS9
Why did it take you so long to brake? Finally watched this on a full screen, and with the road being quite well lit (with other brake lights and stuff), the pedestrian was visible quite a long way in advance, actually (the key thing to me being the brake light of the car in front being obscured - that would mean immediate braking for most people). I can tell when you start to brake by when the camera angle changes; it is significantly after (1-2 seconds) the pedestrian becomes clearly visible in the camera, and as others have noted the actual image most people actually see with their eyes (obviously some people have night vision issues and that would explain a delayed response) at night is much more clear than what shows up on the dash cam.
I had exactly the same reaction to this video. I hope people are not adopting the "Let's see what happens" mindset. If the car does anything you wouldn't do or fails to do something you would do, you should take over immediately. You are still the driver and are responsible for what the car does.

Furthermore, radar does not detect soft objects like people and animals. The ultrasonics can, but they are short range (~16 feet?). The camera is the only real chance of detecting the pedestrian, and given the poor visibility from night, rain and dark clothing, it is not surprising that the pedestrian was not detected (in time).
 
I use AP everywhere I can. But I treat it more like an experiment or test to see how or when it fails, as a result of gain experience using it on city streets I can more accurately predict where it's going to fail or have issues.

I'm always alert when using it and have only had 3 instances where it would have caused a life-threatening situation if I had not taken over. I do take over often because of unpredictable drivers near me, I think AP would have been fine, but would rather take control in those situations.

Overall I'm impressed with how well it does work.
 
As much as I agree with the fact that the person standing in the middle of the road was an utter idiot, I also know that the technology to avoid pedestrians IS already used in many other cars. I currently own one of the two cars that I had the occasion of experiencing pedestrian avoidance. My car shows warning in big red letters accompanied by a ding on the dash when it detects a pedestrian walking in a direction that intersects the car projected path. Usually the pedestrian has to be close to the road or on it. I had alerts of people running on the sidewalk towards my car and stopping at the last moment at the edge of the sidewalk.
I think Tesla still needs to do some work on this> it's not about innovating or inventing at this point. I suspect that it may have to do with my buddy Elon's stubbornness not to have Lidars. :mad:

Your car that will detect pedestrians has lidar?
 
One of the three front cameras should have been a thermal camera. Hopefully Tesla changes this in the future. Watch the video below to see examples but substitute autopilot for where they say driver.

See through dust, smoke and glare. I think it's pretty dumb Tesla did not include this from day one.


I voted yes. I use it on secondary roads when traffic is backed up and in stop and go mode.
 
Last edited:
  • Informative
Reactions: Cowby
Posting this short video (5 sec) captured by the dash cam as a cautionary note for those who use the Tesla autopilot on secondary roads. The video shows my almost hitting a pedestrian dressed in dark at night standing still in the middle of my lane. The autopilot did not detect the person, but fortunately I did.

I think both the cruize control and especially the collision advoidence system should have responded, which together make up the autopilot system. Also, one of the perils (to pedestrians) of EVs is practically no noise. I must be especially careful in parking lots. Standing in the middle of a highway at night, wearing dark clothing, is either stupid or suicidal, or both. I have a M3 and I use the autopilot often, but never at night. It will be a long time before I trust it at night, and even longer when we get to beta it to destinations.
 
If the car does anything you wouldn't do or fails to do something you would do, you should take over immediately.

This is actually my number one issue when I do use autopilot - how fast can I immediately take over? Will my reaction be as fast as if I had just been fully engaged in the driving/steering/speed modulation? It's hard to "blend" my actions with what I anticipate autopilot should do. Since it drives so unlike how I do on the freeway, I am tempted to intervene...but I don't in situations where it's just not doing what I would prefer (for example, it might be driving in someone's blind spot, or might not be giving sufficient space to a vehicle in an adjacent lane). So I kind of end up in this situation where I am "letting it do its thing" while still paying attention 100%, I am not 100% in control. However, because I have to have a slight bias to NOT intervene, I do wonder whether my reaction time would be slowed in a real emergency...

Honestly, even just simple old-fashioned cruise control suffers from this same problem, it's not really unique to EAP - and TACC is definitely an improvement there over standard CC.

Overall I probably need to practice with it more, but it's tough having a system which is mostly capable but not really up to every situation. Compared to my Highlander, which has extremely subtle (totally insufficient for negotiating a corner for example) lane keeping, it's a challenge. The Highlander system is super helpful for avoiding drift on long road trips, but I still have to drive everything. It's actually a good blend because there are fewer fine adjustments - but I still have to be 100% in control. Not saying it is better than EAP or anything, I really prefer Autosteer...but it worries me.

I think both the cruise control and especially the collision avoidance system should have responded,

Why? It's not really designed to respond appropriately in this sort of situation (the manual says it is capable of detecting such things but to not rely on it to slow down adequately). The pedestrian was basically stationary. Most available evidence suggests the car will typically not stop in such a situation. It (really TACC) will do much better if it is already tracking a vehicle in front and something unexpected happens with that vehicle. That's a situation where it can really help (as I said above, the situation of coming up rapidly on untracked traffic is significantly different and it may not respond appropriately at all).

You should expect EAP & TACC to just plow over pedestrians and cyclists and animals, as it would have done in this case.
 
Posting this short video (5 sec) captured by the dash cam as a cautionary note for those who use the Tesla autopilot on secondary roads. The video shows my almost hitting a pedestrian dressed in dark at night standing still in the middle of my lane. The autopilot did not detect the person, but fortunately I did.


Automatic Emergency Braking doesn't kick in until it has to. I have only had it kick in one time and it only kicked in because I purposefully waited to see what it would do. I was just about to slam on the brakes (within 1/4 second of hitting the brake pedal) when it slammed on the brakes HARD and avoided t-boning a car that pulled right in front of us. I'm also familiar with dashcam videos and it's apparent that the manual braking was initiated early enough to completely stop well before the potential impact and it was unnecessary to initiate threshold braking to achieve that.

In other words, it's impossible to say that AEB didn't work in that situation (because you intervened before it was necessary). This is one of the strategies AEB uses to prevent false braking events, waiting until it needs to react. We would all go crazy if it activated earlier or was always scaring us with alarms. The alarm is configurable as to how early it alerts, the AEB is not. It is particularly tolerant at speeds under around 35 mph (because cars can stop so quickly at relatively slow speeds).

Thanks for posting the video because anything that makes AP users aware of potential dangers and makes us think about how we use the automatic driver aids is good. It's important to realize that it's only a driver aid and users must still pay attention to what's going on even though it appears that the car is driving itself. But to say the car would not have stopped in this instance is quite premature.
 
  • Like
  • Informative
Reactions: Russell and Cowby
The very fact that you succeeded and it failed is proof that as of now the Neural Network of the vehicle (what's responsible for this detection) simply isn't good enough.

Wow! I don't know where these "negative Nellie" attitudes come from. I'm a critical thinker so I'm troubled by unsupported statements like "The very fact that you succeeded and it failed is proof...". Because it is not a "fact" that AEB failed.

Consider this:

You are a beginner parachutist and have just made a jump from 10,000 feet using a chute that is supposed to open at 2000 feet. But as you see the farm fields approaching you freak out and pull the manual release to deploy your chute. Once you float to the ground you chastise your instructor for giving you faulty equipment. He says "The equipment isn't faulty, you freaked out and pulled the release at 2200 feet."

To draw the conclusion that AEB failed from that video is irrational and not supported by the evidence.
 
I have only had it kick in one time and it only kicked in because I purposefully waited to see what it would do. I was just about to slam on the brakes (within 1/4 second of hitting the brake pedal) when it slammed on the brakes HARD and avoided t-boning a car that pulled right in front of us.

In this case, had you received a FCW alert?

I would think that usually a FCW would be triggered in advance of an AEB issue, though probably not all the time depending on the situation.

I ask because I would think in this situation if it had seen the pedestrian, it should have issued a warning (but @rmacey did not mention any such warning). As I said, it's not CERTAIN that it would issue an early alert (it might receive information from a plurality of sensors in very close proximity, and then decide that there is an issue and start AEB - and you might not get an early FCW). But I would think that if the vision system had decided it saw a pedestrian it would issue an alert. Of course, you don't want false alarms either, so it may not have been a strong enough "yes".

And we don't know what setting @rmacey had FCW set to.
 
Exactly! I have my Forward Collision Warning sensitivity set to the minimum and don't experience nuisance alerts. My wife's car is on a more sensitive setting and I'm always startled when I'm driving that car when I needn't be. But I think she might benefit from the earlier warning! :D

I guess what I was asking was: whether you received an FCW (even though you have it on the least sensitive setting) when the AEB engaged for a vehicle you were going to t-bone that you referenced above?
 
Wow! I don't know where these "negative Nellie" attitudes come from. I'm a critical thinker so I'm troubled by unsupported statements like "The very fact that you succeeded and it failed is proof...". Because it is not a "fact" that AEB failed.

Consider this:

You are a beginner parachutist and have just made a jump from 10,000 feet using a chute that is supposed to open at 2000 feet. But as you see the farm fields approaching you freak out and pull the manual release to deploy your chute. Once you float to the ground you chastise your instructor for giving you faulty equipment. He says "The equipment isn't faulty, you freaked out and pulled the release at 2200 feet."

To draw the conclusion that AEB failed from that video is irrational and not supported by the evidence.

What does this have to do with AEB?

The OP says autopilot did not detect the person. As in the neural network didn't tell TACC that there is an idiot in the road right in front of us. It also didn't trigger any FCW as one would expect.

Now this doesn't necessarily mean that autopilot wouldn't have seen the pedestrian eventually. The question is whether it would have seen him soon enough to brake completely. Obvious AP didn't see it soon enough to slow to a stop in the way the driver expected. It should have handled it so beautifully that neither FCW or AEB would even be part of the discussion. FCW is really for the human driver to take over, and AEB is last second emergency crash mitigation (as in not always prevention).

I'm being critical of it not because we really expect it to detect a pedestrian in the dark at this point, but the fact that it doesn't illustrates that we're pretty far away from FSD.

I do remain hopeful that the additional processing capability of HW3 will allow for a much more accurate neural network that can detect a pedestrian in these conditions. Assuming the sensors see enough detail of course.
 
Last edited:
I guess what I was asking was: whether you received an FCW (even though you have it on the least sensitive setting) when the AEB engaged for a vehicle you were going to t-bone that you referenced above?

Sorry, I missed that. I thought you were asking the OP.

Yes, but it didn't alert until the instant it actually slammed on the brakes. The reason I was able to wait so long was that the pavement was clean and dry, I was only going about 30 mph, and I know how long it takes the Model 3 to stop in those conditions. Since I could see the problem car's wheels inching forward under the frame of a big box truck that was blocking the other driver's view, I knew what might be coming. I just covered the brake pedal, checked that no one was behind me and waited. AEB (which is used by AP) performed perfectly.