Featured Tesla Autopilot recall probed by safety regulator following new crashes

I interpreted it as FSD turned off for that pit stop, so the corner was done manually.

For a 90 degree turn, with 15 foot wide lanes, and both cars accelerating the same, the outside car's nose would end up 23.5 feet behind inside car's nose, or 8 feet behind the tail.

At least in my state, that isn't legal spacing for lane changing. Common maybe, but not even remotely close to the legal minimum unless moving at just walking speed.

 

It's a false economy to use faster acceleration when you can't process the location of traffic around you fast enough to avoid an accident.

A better, and safer, technique is to let the other traffic out accelerate you, then you can just slide in behind the other traffic. That's what FSD would do, I'm sure.

 

Tesla is about to remove steering wheel nag with new Full Self-Driving update | Electrek

Well that's not going to piss off NHTSA. .

 

NHTSA is specifically asking about Autopilot, Enhanced Autopilot and Traffic Aware Cruise Control, not FSD. Might still piss them off though.

 

where do you come up with this ... generally - If a car is in front - the car behind's responsibility is to back off. Whether behind you or during a merge.
If You Rear End Someone is It Always Your Fault? | Corena Law.
.

 

If you rear end someone because you're following too closely yes. If you rear end someone because they cut you off and make a right hand turn across your lane then no?

Or is this another instance of "FSD and Tesla are fine, you're just a Tesla hater so I'm gonna argue you're wrong" moment?

Because I'm sorry but you don't turn right across a lane of traffic and then claim the person who hit you wasn't providing appropriate following distance when they were never actually following you to begin with...

If you look at the path the cars followed and the short distance covered and where the damage seems to have occurred from the photos (off centre to the passenger side of VW, and passenger side rear bumper cover of Tesla), the Tesla's bumper was at some sort of angle when it was hit. So it was halfway through its turn when it was rear ended. And if that angled corner of the bumper hit the middle of the VWs bumper, then it was offset in the lane compared to where a car normally would be if it was turning after a fully completed lane change. If it was hit mid lane change, it wouldn't have been hit at such an acute angle either. As last I checked lane changes are pretty smooth, shallow turns that won't have the rear bumper at a 45 degree angle for example to be hit as shallow as the paint transfer suggests.

 

A hit offset to the passenger side on both vehicles could well mean the Tesla was turning right in the entrance (showing its passenger side the most ) and the VW veered left to avoid hitting (so its passenger side was showing the most toward the Tesla) it but still couldn't clear it. Could well be the VW driver had plenty of room. was distracted and didn't see the Tesla slowing down and turning. Too bad there is no Sentry footage. It could have put a lid on this discussion.

 

That is certainly the type of defensive driving I learned. Slowing is nearly always safer than flooring it, outside of things like passing on two-way roads in an older low-powered car, or outrunning the volcanic ash cloud of Mt. Saint Helens' eruption.

 

Note that the left turn lane's stop line is about a car length further back than the right's in this case. The distance of the turn is about the same for each lane.

That's how it behaves in the two other brands I've driven. Using the turn signal temporarily disables the LKA, and LTA if in use. It comes back once the car has spotted the lines after the lane change. Don't touch the cruise control controls or brake pedal, and the ACC keeps doing what it was doing. Sounds like Tesla was charging thousands for an ADAS functionality that was standard elsewhere. Though there isn't an EAP option when ordering a US Model 3 now.

PS: Wish Tesla used the terms adopting by the others; LKA=Lane assist, LTA=Autosteer.

 

Enhanced Autopilot does way more than just stay active on lane change, it's like self-driving but only for highways where the car self drive from a highway’s on-ramp to the destination's off-ramp while also suggesting lane changes, automatically turning on your turn signal, navigating interchanges and taking the correct exit.

And yes, they took it off sales as they lowered the cost of FSD and I do hope they'll replace AP with EAP. As castrating AP like NHTSA wants to do would piss off lots of user. If they do make EAP standard (all Teslas manufactured since October 2016 are compatible and just need an OTA update, not just some higher trims like most manufacturers do) , they'll probably take off Summon and other bells and whistles not to piss off (too much) those that bought EAP.

Wiki definition of "LCA"

Lane centering

In road-transport terminology, lane centering, also known as auto steer or autosteer

Common enough to be the mentionned in the first sentence of the definition of "Lane Centering"

 

The other day, I was in FSD on the right lane of a three lanes Highway. The vehicle ahead of me in the center lane put his flasher to merge into my lane but he didn't have the space to merge in safely. Noticing its flasher, my car slowed down all by itself to give him room to merge ahead of me.

Another defensive maneuver it did was to move to the left lane when it encountered two cars stopped (no flashers) on the shoulder of a double lane regional road (that was during the 98 km trip I mentioned last week).

Beside accelerating and braking fast (something that Musk said is fixed in 12.4), it already drives defensively.

 

No offense meant, but my understanding is that it doesn't drive defensively. That would require a set of rules that defined what defensive driving was and how to tell if an action matches those parameters. I recall a video by Musk describing the system that Tesla uses and how it depends on a bunch of amateurs who identify errors by flagging them as they are driving around as "beta" testers. If the drivers exhibit poor driving techniques you get an AI that does not drive well, and that is especially not driving defensively.

 

Is that video refering specifically to V12 and how do you explain the two examples I stated and i could also point to different encounters, one where it slows down to pedestrians standing by the side of the road, leaving room when passing a semi, or when passing a car with an opened door. There are a bunches of examples like that of the car behaving with caution.

 

I believe the ire is misplaced. The NHTSA doesn't dictate the specifics of a required fix to a safety issue. It is Tesla that designed and installed the fix. It is them pissing off users with it. NHTSA didn't tell them to turn off cruise control when making lane changes.

That can also mean Lane Change Assist. A Handy List of Automotive ADAS Acronyms - Edge AI and Vision Alliance Someone named their blind spot warning that.

The AAA has suggested a set of universal names for functions, as the manufacturers, NHTSA, and SAE all have their own terminology, in the hopes of confusing drivers less. A Handy List of Automotive ADAS Acronyms - Edge AI and Vision Alliance
Lane Keeping Assist is what they settled on for systems that steer the car in lane, and it is what Toyota uses. AAA has a lesser Lane Departure Warning that does what it says, but Subaru and Toyota can include some degree of auto steer in theirs. Lane Trace Assist is Toyota's name for it, and is what I've come to use for it here. Despite owning one, I couldn't say what Subaru calls the individual functions of Eyesight, but they work as I need.

 

It's understandable that you interpret the program's reaction to obstacles as if the car is being cautious. I'd be very, very surprised to find that there is an "acceptable level of danger" value stored somewhere in the system. Just because the car is acting as if it's being careful does not mean that being careful is the objective. It's more likely to be a rule in the neural network that says "hitting open car doors is a bad thing that slows down the progress of the vehicle."

Remember all of those "beta" test drivers that have been driving tesla's around for years? They have not been critiquing the software. They have been supplying video samples to be fed to the FSD so that it knows what is expected of the software as it drives.

Remember how cautious the car is when there are pedestrians nearby? All of those behaviors are easily explained by examining the learning dataset that was fed to the AI. If the FSD slows when there are pedestrians standing by the side of the road, it's quite likely that the majority of the beta test drivers slowed for pedestrians.

One of the features that I dislike about Neural Network learning is that it makes it hard to predict what associations will be created the next time that the system is re-trained. If the data which is fed to the system to retrain it lacks the video clips that contained open doors, will it know what to do when it approaches the first car parked at the curb with the door open? Does it have reasonable responses as a fallback when it encounters a new situation? Or will it plow through the first few hundred doors that it comes across? Time will tell.

 

This is a general issue of autonomous driving cars and why we are way behind any initial hopes. It will work 99,999% and then there will be some weird situation that was not learned yet BUT humans could somehow react given they have much more intelligence and general knowledge. Thats why some people say real Autonomous Driving is an issue that needs an AGI (Artifical General Intelligence) to be fully solved. Even then there will be a tiny % amount for error but thats something the liability/insurrance department of the car manufactor has to deal with.

Only way to manage it at the moment (while taking liability) is to reduce the complexity of the scenario while overdoing it with safety precautions and extra sensors - in varying degrees (Waymo Level 4 and Mercedes Level 3). And go forward in baby steps...
Tesla so far tries hard to circumvent this process.. or cheat around it somehow (minimal set of sensors, no liability)

 

Musk did say finding clips with drivers doing complete stops was a challenge. Of course the "defensive behavior" is learnt through videos fed, it's not like an automatic response that human have when faced to some perceived danger. Nevertheless, it does react defensively.

And for the 99.9999% of the time, when do people react correctly 100% of the time in collision avoidance? How many times people simply froze in place instead of taking evasive action? How many drivers simply lock the brakes and turned the wheel hoping the car would go in a different direction that physic dictates?

Yesterday while going to the cottage, on a city road, an unleashed dog started running towards and to the side of the car (driver side), the car moved to the curb as it showed a dog on the visualisation display. Was it trained to react to animals? Not sure but it did. That was my first encounter with a large animal (ie, not a squirrel) and I liked that it took action to mitigate the encounter.

 

"Ladies and Gentlemen of the jury."

• Accident free - 717 mi on FSD
• Minor rear accident - ~300 ft manually
  • Some difference in opinion about fault . . . (IT WAS THE HUMANS)

"In summary"

• Both drivers reported no injury.
• Both cars drove off as lights where not broken nor other safety equipment deployed.
• Bumper covers of both car revealed impact effects.

Worse case, I may sell some TSLA stock tp buy his front bumper replacement. My cost will be my labor reseating the bumper cover

Bob Wilson

 

Given Musk says alot the whole day and especially goes wild during earning reports and when talking to shareholders I really would not take his words too serious. His preditions regarding autonomous driving are pretty infamous at this point. And hes not the only one with experise in the field.

Exactly. People make alot of errors - thats why cars have more and more asistance/safety systems in the first place. Many now mandatory per regulations. I think were all on the same page here with an exception of the mandatory part.

Yeah thats a cool and advanced example of ADAS. Unless youre a machine learning engineer with alot expertise in ADAS and vision systems, the speculations how it was learned or not are probably not going anywhere. For us its important it works reliable and at some point we come back again to the liability part and that everything must be explainable in an audit or similar - especially if it doesnt work the one time.

 

Hence why for now it's supervised. Like my ride yesterday, it was a 101 km from home to the cottage. I only deactivated about 2 km within a 10 km dirt road because it took a curve too fast (about 50 km/h) to my liking and started sliding on loose gravel. I stayed in manual driving for the remainder of the gravel road. Once back on asphalt, I reactivated and stay that way until my private road. I guess it needs more training on how to drive on dirt roads. It has no problem on my private dirt road though as it only goes to about 20 km/h.

Edit: as for Musk, yes, he says lots of bullsh"t but seeing how most (including me) do their stops, I would agree with him that finding people doing stops the "right way" must not have been easy.

As for Waymo, it's not infaillible, like the two recent videos of it going the wrong way. The first one was "explained" as the uniwheeler were blocking its way but the second one? Instead of waiting for the traffic to clear up, it turned left and drove in the opposite direction!