-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
-
Tesla's Supercharger Team was recently laid off. We discuss what this means for the company on today's TMC Podcast streaming live at 1PM PDT. You can watch on X or on YouTube where you can participate in the live chat.
Accident while on EAP...
- Thread starter Freewilly
- Start date
EdisonDrvr
Member
I think you’re exactly right. Foot on accelerator overrides TACC braking. I had an issue the other day that I was ready to blame on the car but it was my operator error. I was in AS and the car came to a stop behind traffic at a light. After the stop the AP went wonky with an error alert so I flipped the right control stop up to cancel AP. Light turns green...traffic starts moving forward but my car went reverse....LMAO. Foot on brake then into drive and all was good...it’s a good thing I didn’t get on the accelerator pulling away because cars were behind me. Car only moved in reverse a foot or so before I simply stopped...then to FWD ...and all ops normal. Careful disengaging automation using the control stalk slightly up because you can get Reverse accidentally. Cheers.
JeffK
Well-Known Member
I always try for the half press. I've done the full press accidentally once myself.
Certain things are definitely going to have to change between what we have today with EAP and initial FSD features. The response to the system disengaging is one of them. Today, it's perfectly acceptable for NoA to alert the driver to take over immediately and disengage. For true FSD, it will need to disengage safely, pulling into a breakdown lane if possible or simply stopping safely in a travel lane with the hazards engaged if not.
There's a ton of speculation about how EAP or AEB failed in this situation and nobody without access to the sensor data is actually going to determine the true reason. The OP has been very upfront about accepting responsibility for the accident and ultimately at the current level of automation that's where the responsibility sits. Speculating on AP bugs to explain this specific instance is silly because there's absolutely no evidence to back them up. If the driver was awakened by the collision, then there's no insight into what warnings were or were not issued by the car prior to collision.
AP isn't perfect, but it's also never going to be. At its current level, with an attentive driver it is safer than a vehicle without it. With an inattentive driver, it's still safer than a vehicle with both an inattentive driver and no AP. I expect a lot more scrutiny once we start seeing topics titled 'Accident while on FSD', but even then realize that if every car in the USA went FSD and killed 15,000 people a year, that would be a huge success.
There's a ton of speculation about how EAP or AEB failed in this situation and nobody without access to the sensor data is actually going to determine the true reason. The OP has been very upfront about accepting responsibility for the accident and ultimately at the current level of automation that's where the responsibility sits. Speculating on AP bugs to explain this specific instance is silly because there's absolutely no evidence to back them up. If the driver was awakened by the collision, then there's no insight into what warnings were or were not issued by the car prior to collision.
AP isn't perfect, but it's also never going to be. At its current level, with an attentive driver it is safer than a vehicle without it. With an inattentive driver, it's still safer than a vehicle with both an inattentive driver and no AP. I expect a lot more scrutiny once we start seeing topics titled 'Accident while on FSD', but even then realize that if every car in the USA went FSD and killed 15,000 people a year, that would be a huge success.
Sounds high, but since it's a new car, you're obviously getting a new deck lid and since it'll require a repaint, it means blending paint on all the surrounding panels. I assume the quote provides a cost breakdown.
This is a fantastic post and exactly what I think as well with a similar background.
For me the biggest issue at the moment is the phantom braking which really feels solvable in the very short term.
I also think if they used rear camera to verify it's safe to brake (lowering the "score" to do it or not, or how much is safe) I'd feel much better for when it does glitch. We may need HW3 before this phantom braking is gone. They must know the issue is there and there seems no progress on it. I suspect they know they need to "process more neural net cases" (what ever they call them) to rule out mistakes. More cases processed per second the more precise "scoring" will be. I think an order of magnitude more processing may solve a lot of issues.
The erratic lane changes I don't mind because I'm ready for it, when I confirm change lane. Not good of course but the braking out of the blue can happy any time. You could go 2 hours with out a glitch and wham.
I'm really shocked they are considering no confirmation NoA without HW3. It seems unnecessary risk. It better be optional, and defaulted off. Or those one in 10 million spectacular failures will become 1 in a million (which is a lot). It may damage their reputation and harm the eventual FSD regulation process to come.
Does anyone know of a phantom brake that has caused an accident? I'd be shocked if it hasn't.
It is very difficult (and therefore inaccurate) to predict distance from a single visual image source, like that from a single rear camera. If you've ever known anyone who has vision problems or lacks one of their two eyes, you might be aware that they've got spacial and depth perception challenges. Taking into account what is occurring behind you is of course very important, but I think Tesla is going to be challenged to utilize this information accurately without the presence of additional radar sensors. There is so much to take into account in a situation like this, including the reaction and the stopping power of the car behind you.
I regularly accelerate through AP directed speed reductions, less to avoid an accident, and more so to prevent other drivers from becoming upset.
I don't personally know of a phantom braking incident causing an accident, but I do have knowledge of an AP initiated lane-change that was made by a Tesla Model X that caused an accident. The vehicle should have continued to travel in it its current lane, but for reasons yet unexplained, the Tesla moved over one lane and struck a slower moving vehicle in the left lane while merging. The driver did not signal a lane change, and the vehicle did not indicate or suggest a lane change.
t could be that Tesla has determined that lane change confirmation does not require the processing power of HW3 and that most of the requirements are solved with their deep-learning models that can be computed with their older generation of hardware.
I feel its a bit early to introduce - not so much because of the danger involved, but more because it's never really done a great job of suggesting lane changes. Tesla probably feels that it's worth making some poor decisions now concerning lane changes and gather the data so that they can improve on it as their data-set size increases substantially.
Daniel in SD
(supervised)
I don't think this is true. Human stereopsis is useless beyond 10 meters. Your brain determines distance by looking at many other clues.
https://human-factors.arc.nasa.gov/publications/AIAA.2011.DepthPerceptionCueCntrl.pdf
EdisonDrvr
Member
I haven’t experienced any phantom braking for the last 3 OTA updates. Pretty solid. I do occasionally still get the nervous Nellie second though lane change abort. 8.3 didn’t seem as solid in general as the prior OS. Just received 8.4 but I haven’t driven with it yet. The 5% HP boost in 8.3 seemed like a nothing burger. Can’t discern any difference whatsoever ... I guess it would depend on what part of the motors powerband was improved but from a stop I can’t feel any difference at all....acceleration 60 to 80mph felt the same as well. Kind of a bummer but I was hoping for a little more pop from say 50 mph or so...nothing noted. Guess 23 HP more isn’t really noticeable anyway from a seat of the pant perspective. Interesting to see P3D dyno to see if there really was a HP bump.
My uncle was killed by one of the sleepy drivers, as was the other driver. Please pull over and use dog mode for a quick nap.
If the driver didn't initiate a lane change and the car didn't signal a lane change, then AP didn't initiate a lane change. Now, it might have followed the wrong lane markings, inexplicably drifted out of lane or what seems more likely in the case you've described tried to center itself in an apparently wider lane caused by an unmarked merge. I'm especially paranoid of using AP past unmarked merges because of that behavior, it should keep tracking the left (or right) line if the other line suddenly terminates and the lane widens.
NoA suggests lane changes for route keeping or passing. You have control over how aggressively it suggest speed based lane changes and it's gotten better about moving you back over after passing. Route based lane changes are still a bit iffy, especially on 2 lane interstates in my area.
In any case, an explanation of NoA lane change confirmation has already been posted by Tesla. You need to actively disable lane change confirmation if you choose and you're given an opportunity to cancel a proposed lane change before it executes.
Introducing a More Seamless Navigate on Autopilot
Good point about the rear not having "Stereo" vision like the front does. Perhaps it really should. Or Radar.
I wonder if the two sides with rear could form "most" of a "depth view". Maybe even the inside rear facing as well
Sure, you can drive through the deceleration, if you catch it early enough. I'm usually in WTF mode for a second or two, trying to assess if it's right. I personally have less issue with lane change issues. But that is certainly #2 on my list.
I have not had the car change lanes on it's own, which is what I think you said. I understand that it could (under the collision avoidance umbrella). But not for NoA. As far as I understood.
Good point about the rear not having "Stereo" vision like the front does. Perhaps it really should. Or Radar.
I wonder if the two sides with rear could form "most" of a "depth view". Maybe even the inside rear facing as well
Sure, you can drive through the deceleration, if you catch it early enough. I'm usually in WTF mode for a second or two, trying to assess if it's right. I personally have less issue with lane change issues. But that is certainly #2 on my list.
I have not had the car change lanes on it's own, which is what I think you said. I understand that it could (under the collision avoidance umbrella). But not for NoA. As far as I understood.
Stereo imaging for the rear is better than the front, at the moment - though with the challenge of different focal lengths. The fender cameras see directly to the rear, so any object will be in the sight from at least one fender camera and the rear fisheyebl unless it's very close.
That gives a minimum of a three foot separation between cameras, compared to the couple inches forward.
Have a look at the video.
The lane markings appear to be clear. It does not seem to represent the case you have described (and for which I encounter daily on my drive home of a widening lane but it is possible that it got confused).
I'm very interested in your opinion, and the others in the forum.
Here is the narrative of what we think happened, after speaking in more detail with the driver this afternoon, and reviewing the footage in slow-motion.
- Model X was driving in the center lane and passing an adjacent car positioned in the left lane.
- The Adjacent vehicle was starting to drift slightly into the center lane.
- As the Model X was passing this vehicle, it also drifted slightly over the lane marker to the left as if it had an intention to occupy the lane.
- The Model X did not appear to be fully aware of the adjacent vehicle, or its distance.
- At this point, the two vehicles collided - the rear left wheel of the Tesla came into contact with the right front wheel of the adjacent vehicle.
- Driver corrected for a moment after the impact by briefly steering right, and then manually took the Tesla into the left lane, and stopped.
Tesla gathered all of the data, escalated it internally, made some calls back to the driver, and promised to be in touch within 7-21 days later. There was no additional meaningful follow up from Tesla other than to say that their AP team reviewed the footage and are currently prioritizing other AP related investigations and would eventually respond with more detail. There have been weekly communication since then (initiated by the driver) but no answers. I think we are hoping to better understand AP's intention just prior to, and during the impact.
The driver took immediate responsibility for the event and settled the damage. He did not end up involving insurance and was able to repair the damage to his model X for relatively little cost. Fortunately, no one was injured in the event.
Here are the images post impact:
Last edited:
Daniel in SD
(supervised)
How sure is your friend that he was actually using autosteer? I know that if I was next to someone driving erratically like the driver in the left lane I would have disabled autopilot. Maybe he disabled autosteer by putting pressure on the wheel and then let go of the steering wheel after he thought he was clear of the car in the left lane?
Affirmative on the question of whether or not autosteer was still enabled. At this point in his AP driving experience, he was still extremely trusting and let it do its thing. I'm a bit more cautious and have a tendency to disable AP on any questionable driving condition. I'm fairly certain I would have disengaged in this scenario.
Daniel in SD
(supervised)
Dunno. Maybe the car driving over the lane line repeatedly made it hard for AP to see the lane line. I bet neural nets are hard to debug! That's why it's beta software I guess.
One way to increase depth perception is to move each image source further away from one another. If our eyes were a foot apart or more (like in the case of a Hammer-Head Shark) for example, we'd have even greater depth perception than we do today. Correct me if I am wrong, but I've reviewed the footage recorded by my TeslaCam, and don't see evidence that the fender cameras have any visibility of cars directly behind the vehicle.
For stereo vision to work, there is a need for overlapping frames, and it doesn't appear like there is. Without which, we really can't estimate distance. I should also point out that there is quite a bit of visual processing that goes into our assumptions of distance (since we don't have radar or sonar), and we can easily throw it off by presenting it with images that we assume to be of a specific size, but really are not. This is also one reason why the moon appears to be of different sizes depending on where it is relative to the horizon (the closer to the horizon, the larger it appears to be). It also explains why radar is so important for measuring distance since its more of an authoritative (vs. interpreted) source.
Similar threads
