from the You’re-a-Bad-Driver-if-Tesla-Full-Self-Driving-Is-Better-Than-You dept.
Motor Trend reports on FSD in their long term test 2023 Model Y, https://www.motortrend.com/reviews/2024-tesla-model-y-long-range-yearlong-review-update-9-full-self-driving-fsd-version-13/ They weren't too impressed with the first software version that came with the car,
I'm not using FSD because it's bringing me much utility or, much less, enjoyment. I'm using it because we paid $15,000 for this software (it costs $8,000 today), and I'm going to do my job and report on how it works, dammit.
That's despite FSD giving me many, many reasons to forsake it, to decide that my safety and sanity are worth more than what it cost. Yet the Full Self Driving note I created in my phone to log the system's transgressions has an ever-increasing abundance of entries as we pile the miles onto our Model Y.
Dumb and Dangerous Decisions
For example, there was the time it failed to recognize an increased speed limit sign and continued bumbling along at a 15-mph deficit. In stark contrast, later in that same drive, it detected a 55-mph speed limit sign specific to vehicles towing (which it wasn't), decelerating from 75 mph so rapidly that traffic behind had to swerve around the Model Y. Moments later, FSD decided to change lanes to follow the navigation route toward its next turn—still some 10 miles away—cutting someone off in the process.
[...]
On a different day, FSD deviated from my navigation route because it neglected to recognize that the lane it was occupying became right turn only. After making that turn, it wanted to correct its error and resume the route by making a U-turn at the next intersection, where a "No U-Turn" sign was clearly posted. It tried to make its illegal U-turn from the right side of that double-protected left turn, such that if I hadn't intervened it would've overlapped with the vehicle turning from the left-side lane.
[...]
FSD's errors aren't always dangerous. More often, they're just asinine. Like when our Model Y didn't react to a green arrow for a protected right turn, inconveniencing me and drivers behind. Or entered a packed intersection as a yellow light expired, coming to a stop inside a crosswalk. Or braked hard after it accelerated up a freeway on-ramp because it detected an inactive traffic control signal. Or when it encountered an unexpected road closure and drove around the same block three times because each time it arrived back at the closure it failed to recalculate its route. Who knows how long FSD might've kept circling had I not turned it off.
Then the car updated itself to FSD v.13 and while the general experience was smoother there were still problems.
Twice in a week, FSD 13 completely missed freeway exits because it shifted over too late to negotiate fitting in with other traffic. Those bungles are bizarre given FSD 13's tendency to otherwise move toward exits literal miles early, often abdicating the fast-moving left lane to fall behind slow vehicles to the right—vehicles it could've passed had it stayed put for longer. When merging onto a freeway, it will stubbornly attempt to fit into a gap regardless of whether those other closest drivers seem willing to allow it, ignoring suitable openings immediately ahead or behind. Generally, its lane shifting strategy is poor; I've counted as many as eight back-and-forth lane changes within a minute as FSD 13 tries to figure out what to do. The day this article was due, it veered toward a wall as two lanes converged into one.
If simple cruise control was the genesis for any self-driving tech, then FSD 13 represents an ignominious legacy as it struggles to maintain speeds. Once, it gradually relaxed its cruise speed to 64 mph in a 65 mph zone, when the maximum speed I allowed for it was 75 mph. There was absolutely no traffic, faulty input, or other detectable reason for this slowing. At other times, it doesn't keep up with leading traffic accelerating to the speed limit, driving slower than necessary and letting a gap ahead grow even when the Hurry logic is selected.
One would also hope that software ostensibly aware of the physical dimensions and dynamic abilities of the Tesla it controls would know how to avoid hitting markers along freeway curves and keep itself centered in the lane. Not FSD 13.
I believe this long term test car is being used in LA and surrounding areas. YMMV if you live elsewhere.
(Score: 4, Insightful) by PiMuNu on Wednesday January 29, @08:57AM (8 children)
Self Driving is a hard problem - so no surprises that it is hard to solve. By my guess, probably it is impossible to solve without support from the infrastructure (e.g. active roadways communicating with the car and cars communicating with each other).
(Score: 4, Informative) by drussell on Wednesday January 29, @09:43AM (3 children)
Man, it's a good thing that Elon's cracker-jack Tesla team already solved that problem a decade ago!
Remember back in 2014 when he said that in 2015 you would be able to do 90% of your miles self-driving on autopilot. Then in 2016, when he said that the, "Model S and Model X, at this point, can drive autonomously with greater safety than a person, right now!" He then further stated that by the end of 2017, you "would be able to go cross-country from LA to New York by the end of the year, fully autonomously."
You also see all those millions of Tesla robo-taxis everywhere, since he said that by 2020 they, "for sure, will have over one million robo-taxis on the road.
Yep! All done!! 🙄
Musk, the carnival barker: "Step right up, folks!!" 🤡
(Score: 3, Informative) by PiMuNu on Wednesday January 29, @10:33AM (1 child)
> you would be able to do 90% of your miles self-driving on autopilot.
It's true, you can do the 90 % where the car doesn't crash on autopilot. [joke]
(Score: 3, Informative) by Freeman on Wednesday January 29, @03:20PM
Like jumping out of an Airplane without a parachute. Generally, it's not the fall that kills you, it's the sudden stop at the end.
Joshua 1:9 "Be strong and of a good courage; be not afraid, neither be thou dismayed: for the Lord thy God is with thee"
(Score: 2) by epitaxial on Wednesday January 29, @04:55PM
Still waiting on my $30k cybertruck!
(Score: 1, Insightful) by Anonymous Coward on Wednesday January 29, @01:30PM (1 child)
What do we do with all the other drivers on the road who aren't listening, taking cues, and constantly communicating intents?
What I see coming is a carefully choreographed traffic dance, where the tech cars know the moves...but there are those out there who do not have the technology and will always seem to be in the wrong place at the wrong time.... Like a loose nut in the crankcase.
Is this a plan to force us into electric cars with remote control? This to me is just as worrisome as being corralled into computers with backdoors that only supposedly "trusted" "authorities" have access to.
Secrets are very fleeting
Personally, I am leery of centralization, as it has a single point of failure...or that it would be used as a central point of monitoring everyone's comings and goings with the ability of selective immobilization at the click of a mouse. I fear giving anybody that much control.
The way things seem to be falling into place, soon a surprisingly small cadre of "elites" will be able to impose their will on the rest of humanity without having to rely on other humans to carry out orders, as other humans may not obey if they smell a rat.
I believe the people who are planning to rule the roost are well aware of the practice of being "fragged" by their own men.
(Score: 1, Interesting) by Anonymous Coward on Wednesday January 29, @01:38PM
Are we reliving this bit of history? https://en.wikipedia.org/wiki/William_McKinley [wikipedia.org]
Perhaps unfortunately, it doesn't look like we have anyone like Teddy Roosevelt and his square deal in line as a successor... https://en.wikipedia.org/wiki/Theodore_Roosevelt [wikipedia.org]
(Score: 2) by hendrikboom on Saturday February 01, @04:56AM (1 child)
You mean a railway?
(Score: 2) by PiMuNu on Saturday February 01, @09:33PM
Railway is great but in my experience has problems:
* "Last mile" issue - time and cost of travelling to the station is often prohibitive
* Scheduling/routes - for many journeys the scheduling and routes make travel from A to B impractical - for example if A and B are not major cities one often has to change several times to make a journey.
* Cost - railway works out more expensive than car even for single journeys. I understand road may be more heavily subsidised than rail.
I don't live in a major city which makes travelling by rail often impractical.
(Score: 3, Insightful) by pTamok on Wednesday January 29, @10:11AM (16 children)
My car (non-Tesla) is meant to recognise speed-limit signs.
It's not very good at it.
Not only does it miss signs, it misreads some, and confuses some non-speed limit signs (such as 'no-overtaking') as a speed limit sign. It also does not have a complete database of signs, so that in foreign countries where the format of a speed-limit sign is unusual, it simply doesn't recognise them.
The function to set the car to automatically abide by posted limits is available to me. I have it disabled because it is so poor at it. In the EU that might not be an option in the near future: or might have insurance implications if you do.
A previous car had road speed limits incorporated in the mapping software data, which, though sometimes out of date, acted as a good backup (where I live, large zones can have a speed limit signs at the entrances to the zone and at the exits, but none in the zone itself. It can be difficult to remember that you are in a low speed limit zone rather than a standard urban speed limit zone, so having the map as a reminder is good. It also helps where signs are obscured by foliage in summer and snow in winter.)
What is irritating is that I have no means of giving feedback to the software to get it to recognise unfamiliar signs, unlike, say, a human navigator. Or to explain that that particular sign is a 'no-overtaking' sign, not an 80kph speed limit sign; or that the information under the sign tells you that the speed limit only applies at certain times; or that the speed limit is different for vehicle towing trailers or caravans. The software does not learn, and will continue to do a bad job indefinitely.
I really don't know what the solution is. I've thought of using QR-codes on signs in addition to human-readable symbols, as QR-codes can be more recognisable for automated systems. In principle, one could also use balises [wikipedia.org], but I suspect that could be quite challenging as the road/highway environment is significantly different to railways - it is less controlled and less predictable.
The authorities responsible for road/highways must have a database of what speed limits (and other restrictions/warnings) apply to each stretch of road/highway, so what could be done is have a common (open) map and update protocol that (a) keeps a common mapping resource up-to-date and (b) allows for a common (open) map update protocol for the mapping software in vehicles. It's not technically difficult, but there are too many political and financial interests involved to make it a reality. It is the kind of thing I could expect the EU to mandate, as it would be a 'common good'. I can dream.
Meanwhile, I get irritated about the shortcomings of the current implementation on nearly every. single. car. journey.
Sigh.
And this is a simple problem to solve compared to 'Full Self Driving'.
(Score: 2, Insightful) by shrewdsheep on Wednesday January 29, @10:40AM (2 children)
Certainly depends a bit on the situation. However, speed limits expire automatically at the intersection of larger roads (or ramps on freeways). You could have come from the other road or could have entered from a ramp and haven't seen the sign yet. For your example of the fenced area, you could have ordered your car by helicopter drop and would have not passed a speed sign at all until you would leave the area. This would be a safe defense in court.
(Score: 2, Interesting) by pTamok on Wednesday January 29, @01:45PM (1 child)
That depends on the country, and, I guess, in the USA, by state.
https://www.gov.uk/government/publications/setting-local-speed-limits/setting-local-speed-limits#framework [www.gov.uk]
Note that some roads with a speed limit of 30mph have no explicit signs for this: you need to measure the distance between lamp-posts.
(Score: 2) by PiMuNu on Wednesday January 29, @02:03PM
[whinging] And with the introduction of 20 mph limits, speed limit is now completely random. I now frequently find I have been driving at 30 in a 20 zone and vice versa.
(Score: 2, Insightful) by Anonymous Coward on Wednesday January 29, @11:42AM (6 children)
> using QR-codes on signs in addition to human-readable symbols
In a same world that might make sense. At least around here, QR codes would be targets--if not for shotguns, then for spray paint. The weather can also cover signs, lately we've had high winds and snow which often sticks to vertical surfaces.
> a database of what speed limits (and other restrictions/warnings) apply to each stretch of road/highway,
You would think so, but here roads can be owned/maintained by federal gov't (most likely to have good data), states, counties, cities, townships and even villages (small areas within towns that may have separate taxation). My guess is that when there is emergency work being done and local speed restrictions (eg. water-main leak-repair), that isn't available in any database.
What I believe Tesla claims to do is review masses of customer data (vide0, speed, etc) and fix their own database. If their cars are routinely going a different speed than what their database says for that bit of road, maybe that is reviewed and updated? It's whack-a-mole with continuing changes on the road, so the database will always be behind.
(Score: 0) by Anonymous Coward on Wednesday January 29, @01:12PM (1 child)
> In a same world --> In a sane world
(Score: 3, Insightful) by Dr Spin on Wednesday January 29, @01:34PM
Best of luck finding one of those!
Warning: Opening your mouth may invalidate your brain!
(Score: 2) by PiMuNu on Wednesday January 29, @02:05PM (1 child)
> review masses of customer data (vide0, speed, etc) and fix their own database
Presumably they can run Tesla equivalent of google car around every street in the USA. I assume this is what Self Driving really means in a lot of cases (hence municipally limited robotaxis exist).
(Score: 1, Insightful) by Anonymous Coward on Wednesday January 29, @03:21PM
> Presumably they can run Tesla equivalent of google car around every street in the USA.
There are enough Teslas on the road to have pretty good coverage for most roads, at least in all the places that count(grin...meaning poor areas may lack coverage).
But consider this thought experiment: Add up all the hype, effort, and money that has been poured into self-driving over, what, the last 10 years? What if that treasure was put into driver training, where every driver:
+ rides in a skid car so they learn the rudiments of car control on slippery surfaces and when a tire goes flat
+ gets some supervised training to experience what hard braking and avoidance maneuvers can achieve on good road surfaces
+ has the effects of lack of attention/distractions demonstrated first person in a car (not in a lecture hall)
+ (optional) gets to drive drunk in a supervised area to witness the effects on judgement and coordination first hand
This could be a relaxed version of pilot licensing where some amount of flying and/or simulator time is required to "stay current".
(Score: 2, Interesting) by pTamok on Wednesday January 29, @02:27PM
Yes, that is a common problem. My thought would be to mitigate it slightly by using paint that is transparent in visible light, but opaque in IR or UV.
I am familiar with that problem, which I why I began to think of balises. Possibly something like a Low-Energy Bluetooth beacon, which could be powered by a battery topped up by a solar cell (or possibly a small windmill). Unfortunately not vandal-proof.
(Score: 2) by hendrikboom on Saturday February 01, @02:11PM
But they can collect this mass of driver data only when enough of their drivers do their own driving. Otherwise its the AI learning from its own output.
(Score: 4, Insightful) by ElizabethGreene on Wednesday January 29, @05:10PM
You raise a good point. Any system that's working towards autonomy needs a feedback mechanism, if not to the local instance, at least back to the development team.
(Score: 2) by mcgrew on Friday January 31, @02:10AM (1 child)
Springfield changed a few streets from one way to two way. So now my car complains noisily when I go the "wrong way" down those streets. But it usually gets the speed limit signs right.
A Russian operative has infiltrated the highest level of our government. Where's Joe McCarthy when we need him?
(Score: 0) by Anonymous Coward on Friday January 31, @07:48PM
> Springfield ...
Wait, your neighbors are the Simpsons? That explains a lot(grin)...
(Score: 2) by aafcac on Friday January 31, @06:07AM (2 children)
It's been my feeling for a while that these systems should be able to predict what a competent driver is going to do before they do it based on whatever sensor data it has available. Take the data from the harder bits of driving and make the AI work out what to do to avoid having a crash. Which, admittedly requires a large amount of data and it requires some method of filtering out the incompetent drivers, but if the AI is programmed to enforce a proper following speed in later stages, the competence of the driver's being modeled can probably be somewhat relaxed.
But, really, Teslas should be using more sensors than they are, and they should be training them more than they are. Even just enforcing a duty cycle like tools would probably help as well. 5 minutes of AI driving followed by 5 minutes of person driving would probably go a long ways to building up the necessary library of interactions with various eventualities with enough drivers while drivers could still be expected to monitor what's going on.
(Score: 0) by Anonymous Coward on Friday January 31, @07:53PM (1 child)
> ... predict what a competent driver is going to do before they do it based on whatever sensor data it has available
That sounds within the realm of possibility.
Now try this one -- have you driven at dawn or dusk in areas with herds of urban deer? Their behavior is so random that I strongly doubt that anyone can model it. Once I saw a small herd running in a circle that took up the width of a 4 lane major road (at dawn).
(Score: 2) by aafcac on Sunday February 02, @12:23AM
Clearly, it would take an incredibly long period of time for some less common situations to be encountered enough to build a proper library of situations. But, it's several levels more appropriate than what Musk is doing, refuse to have an appropriate set of sensors to save money and just get people killed. In an ideal world, this would be tested separate from other drivers, but there are so many issues with that, not the least of which is that other drivers are a large part of what these systems need to be able to deal with until and unless all the other cars are also being driven by computers.
In the case of deer, the focus would really have to be in identifying that there are or could be deer present and slowing enough to react as they are deer, they react unpredictably for evolutionary reasons and self-driving does not get you out of what physics limits the cars motion to.
(Score: 3, Informative) by VLM on Wednesday January 29, @04:49PM (4 children)
My experience with AI in general is the doomers are invariably authoritarians who STILL believe the lies the media tells, and then you get a "tell all" review from someone who actually uses it without a NDA or being paid by the marketing department, and inevitably it turns out to be a plate of hot garbage.
The biggest problem we have economically right now, is too many people for FAR too few jobs; if you want a driver for your car, hire about 21 poor people to drive it for you over the internet, and do some additive analysis stuff such that if 11 of them vote to step on the gas then it steps on the gas. It'll be cheaper and more effective than writing an AI that won't work well.
(Score: 3, Interesting) by ElizabethGreene on Wednesday January 29, @05:20PM (3 children)
I don't know that I'd trust Motor Trend to be an unbiased source. It's a car magazine written by and for car guys that as a general rule dislike EVs specifically and anything that further separates the driver from the road generally.
They squawked relentlessly for drive-by-wire systems and the slow death of manual transmissions until EVs started getting popular enough to give them a new target.
As anecdata*, the feedback I've gotten on FSD in the beta from real human beings driving on Tennessee and Georgia roads is a net positive. They have remarked that the newest version doesn't lock-in and center the car in the lane with religiously fervor like the old version did, but they're getting miles-between-interventions high enough that getting bored is an issue. The one person I know with it in Indianapolis, IN does NOT like it because it doesn't dodge potholes.
* anecdata is my new favorite word, and I strongly encourage its use. :D
(Score: 4, Funny) by VLM on Wednesday January 29, @05:46PM (2 children)
There's a financial motivation for a self driving car to damage the car resulting in higher profit.
(Score: 3, Touché) by Tork on Thursday January 30, @03:30PM (1 child)
"Standard Pothole Avoidance: $5.99/mo. -- Premium Pothole Avoidance powered by DOGE-AI: $19.99/mo."
🏳️🌈 Proud Ally 🏳️🌈
(Score: 2) by aafcac on Sunday February 02, @12:25AM
It would be cool if these cars could report the position of these potholes and similar road defects for authorities to fix. Even cooler would be if the local authorities had money do do so and would fix the damage.
(Score: 2) by PastTense on Wednesday January 29, @06:15PM
It's not clear when the particular incidents happened (although the article was published very recently), as I have been reading that very recent updates to the software have resulted in significant improvements in self-driving) .