FSD / AP Rewrite - turning the corner?

[Wol747]

Or its just debug mode...

Your comment brings me to another issue - in fact, the fundamental one.
Reading many posts and watching dozens of videos, the common theme is that no-one knows the principles on which the cars work nor things like failure modes etc.
Airplane drivers have to have a pretty comprehensive technical understanding of everything but especially the autopilot. Monitoring the avionics and knowing what and why they are doing what they are is essential - and it's arguable that monitoring a car driving itself through a city is far more stressful than monitoring an automatic landing when everything is behaving itself, let alone when it throws a wobbly.
There is damn-all information available about the way the thing works and its limiting envelopes, and that's not healthy IMO. The problem is, the manual would be several thousand pages and even with the present enthusiastic driver cadre no-one would be conversant with it: in five years' time there will be millions of vehicles zooming around with "drivers" who have no idea what's going on. (There are plenty now!)

 

[Yev000]

there will be millions of vehicles zooming around with "drivers" who have no idea what's going on.

You just described my wife when she's in the passenger seat and I'm driving. She doesn't know my failure modes either.

 

[Glan gluaisne]

Also worth noting that both these use HD maps and are heavily geofenced.

Certainly Waymo are relying very heavily on detailed mapping, but MobilEye not so much, much of the content of that video makes it clear that there is a lot of similarity between the MobilEye approach and that of Tesla, using live camera sensing.

Do you have a source for "you cannot get this in the UK today, and most probably won't be able to for some time"? There wasn't a major delay in rolling out (after first mass rollout in US - not at that point yet with fsd beta) stop sign control in other regions although I do suspect that V11 and seasonal break may get in the way.

I don't have FSD, but of those in the UK that do, have any got this capability today?

AFAIK, the FSD beta is only on very limited release in some areas of the US.

 

[MrBadger]

This is what surprises me about how they're going about it. As you say, on a given road I know, I'll know there's a roundabout round the corner, and start to slow in anticipation. The car has a map, and so also knows this, but does no such thing, and will barrel along until it sees something it has to react to.

What we have been seeing up until now (prior to city streets) [edited for clarity when taken out of context] does not handle specific behaviour for roundabouts (except on off ramps) and many other widely reported scenarios so AP behaviour up until now is not necessarily the same as that going forward, AP is TACC+Autosteer highways. FSD city streets is, urm, driving on city streets, so it knows how to handle many city and urban scenarios, and will understand more and get better (likely very quickly according to Brandon) over time. Up until now, behaviour AP in city and urban (non highway) environments has been a consequence of more general behaviour. Only now, with the new City Streets module, can behaviour on non highway scenarios start to be judged, and even then, imho, its still very early days and still far too early to dismiss poor behaviour as endemic failures. I think we will very quickly start getting a feel on what is down to a shortfall of learning and what is more of a fundamental issue.

As for the car not using maps to better understand its surroundings, green has this tweet on the subject.

https://twitter.com/greentheonly/status/1320604581534015488

 

[Yev000]

What we have been seeing up until now does not handle specific behaviour for roundabouts

 

[MrBadger]

Certainly Waymo are relying very heavily on detailed mapping, but MobilEye not so much, much of the content of that video makes it clear that there is a lot of similarity between the MobilEye approach and that of Tesla, using live camera sensing.

MobilEye have been crowdsourcing HD maps from recent certain BMW's for several years now. Pre Covid, the aim was to have North America and Europe pretty much mapped by now. HD maps is pretty important to their autonomous driving offerings giving them 10cm accuracy - like driving 'on rails'. Around 38:20

I don't have FSD, but of those in the UK that do, have any got this capability today?

AFAIK, the FSD beta is only on very limited release in some areas of the US.

It was the "for some time" which was the statement I was querying.

 

[Tony Hoyle]

We don't know how long it'll take.. it's the time to bake in european traffic laws into the NN, handling the left/right switch, and presumably it'll need a slow beta rollout just like north america to fine tune the data.

Regulation I'm not that worried, the NN can learn to not take corners too fast which fixes what is often (incorrectly) called the steering limit. Most of what it does isn't specifically covered, so not regulated (loopholes! Same reason NoA can indicate off a motorway). Only indicating to change lanes is problematic.. it's a headscratcher to me that there's so much resistance to allowing driver assistance systems to do that...*

So stuff like avoiding parked cars, roundabouts, turning I reckon we could get within months *if* it learns and works as well as advertised.

* That can also be overridden by national legislation - same reason we didn't adopt the 50kmh limit to ALKS.. if there's the political will.. but I'm not going to rely on that.

 

[Glan gluaisne]

Does anyone really believe that we will have safe and reliable FSD on any car driving around on normal public roads anytime soon? If I had to gaze into a crystal ball, I'd suggest that we'll probably see something like Waymo or Uber operating in limited areas in some cities within the next two to five years. We may get limited self-driving on some other roads, like motorways and dual carriageways within 5 years (although I think that may be stretching it a bit). I doubt we'll get everyday fully autonomous driving within five years, if I had to guess I'd say it'd be closer to ten years.

Just dealing with the legislative aspects to enable trials to begin is going to take a year or two. Look how long it's taken to get something really simple, like electric scooters, to gain even limited approval - Segway spent years trying to gain approval here (and elsewhere), and went bust before it happened.

 

[Mardak]

Regulation I'm not that worried, the NN can learn to not take corners too fast which fixes what is often (incorrectly) called the steering limit.

Do people have a better sense of what the regulations do allow? It looks like the "maximum value for the specified maximum lateral acceleration" is 3m/s², but practically what does that feel like? Indeed planning ahead to take a turn slower instead of a last second reaction should be doable within regulation.

Some quick math of accelerating at the full 3m/s² for 1 second is just under 7mph. But a wheels don't turn 90° anyway, so only some of the acceleration would be attributed to lateral motion.

From the earlier video of crossing the center line to pass the parked truck, it looks to take 2 seconds to go from stopped to past the truck. A constant 3m/s² lateral acceleration for 2 seconds would be about 20 feet, so well within the maximum limit. Even 1m/s² lateral acceleration over 2 seconds would be about 7 feet, so that would be enough to pass slowly. However, a faster passing maneuver over 1 second accelerating at 3m/s² would move laterally 5 feet.

 

[Yev000]

Does anyone really believe that we will have safe and reliable FSD

"Belief" has nothing to do with it. Disruption is usually exponecial and follows an S-curve. You are basically saying this tech can't do it... Because if it can, it will, within 5 years or much less.

The real question to ask is how we can prepare for that possible outcome. Even if it wont happen, (it will eventually) it would be good to hedge your bets.

 

[Yev000]

It looks like the "maximum value for the specified maximum lateral acceleration" is 3m/s², but practically what does that feel like?

Pretty uncomfortable - A taxi driver would not take turns/bends that fast and AP/FSD does not like slowing down. All videos I've seen people slow it down manually.

That is one thing that basically should not be an issue. It should just take turns/bends slower.

As an example, it can't take an AP bend going out of Guildford on the A3 where the road is 70 MPH. If you set it to 55 MPH it will. Other cars slow down for that bend

 

[goRt]

Same with corners. I know and the map knows there's a sharp corner ahead.. so why not slow down in anticipation?

I wonder sometimes if tesla's 'we're going to do it with vision only' idea is causing them to be a bit blinkered. Humans don't use only vision.. we use memory and maps (and paper maps going back into antiquity).

If you read roadcraft then the technique used on unfamiliar roads is to watch the arrowhead where the curbs meet - the further away that they start to move the less sharpe the turn is - all done with sight no memory required

 

[NewbieT]

FSD cheating? Nope.

Think about blind corners you’ve never driven before. You drive slowly on the first couple of times, learn the curve of the road and on subsequent trips often drive a little faster. Potentially dangerously because by their nature you can’t see around the corner until quite late on. There is a happy middle ground in terms of speed. Making progress & safe. FSD should use fleet speed for these corners.

Now if we’re not being pureists and allow fleet speed, why not fleet mapping? You drive a route, learn the roads, speeds marking and use this information when you drive it again, updating your knowledge of the road each time you drive it. New speed limit, new pot hole etc. I have no objection to the car updating a central map and the car using that info. Yes the car should be able to drive on a road it’s never seen (prove autonomy works), but that’s not the most common human driving equivalent so why shouldn’t FSD have the benefit of detailed prior knowledge too. The important part is the car deciding when to use, override or update its prior knowledge.

 

[Glan gluaisne]

"Belief" has nothing to do with it. Disruption is usually exponecial and follows an S-curve. You are basically saying this tech can't do it... Because if it can, it will, within 5 years or much less.

The real question to ask is how we can prepare for that possible outcome. Even if it wont happen, (it will eventually) it would be good to hedge your bets.

Selectively quoting someone to deliberately distort a quote to mean something other than that intended by the author isn't very honest, is it?

What I actually wrote (my added emphasis) was this:

Does anyone really believe that we will have safe and reliable FSD on any car driving around on normal public roads anytime soon?

The bit you chose to edit out is the key phrase in this question, which relates to the timescale. You seem to be in broad agreement with this, with your estimate

within 5 years or much less.

The timescales I quoted in the rest of that post of mine you chose to edit out, were:

If I had to gaze into a crystal ball, I'd suggest that we'll probably see something like Waymo or Uber operating in limited areas in some cities within the next two to five years. We may get limited self-driving on some other roads, like motorways and dual carriageways within 5 years (although I think that may be stretching it a bit). I doubt we'll get everyday fully autonomous driving within five years, if I had to guess I'd say it'd be closer to ten years.

Just dealing with the legislative aspects to enable trials to begin is going to take a year or two. Look how long it's taken to get something really simple, like electric scooters, to gain even limited approval - Segway spent years trying to gain approval here (and elsewhere), and went bust before it happened.

Maybe your

within 5 years or much less.

is closer to what we might see than my

I'd say it'd be closer to ten years.

Both are at some time in the future, rather than next week, next month or even next year.

I have absolutely no doubt that we will get fully autonomous vehicles sooner or later, the questions are really when this will happen and what form they may take. I'm inclined to the view that thinking of fully autonomous vehicles as if they will be cars, as we know them, might be a constraint. It seems probable that, by the time we have fully autonomous vehicle technology, the need for a vehicle to look like, or behave like, a car may not be the primary design consideration any longer.

 

[Sparkeur]

The one thing I can't seem to see on these new beta's is whether they have fixed what I believe is a fundamental problem - the fact that autopilot or FSD will hurl the car towards a speed restricted zone only slowing well after the new speed sign has been passed. I can see that the nav 'knows' when it is entering a new speed zone so why can't it slow in preparation (like humans do) rather than wait until the cameras detect the sign?

 

[Yev000]

Selectively quoting someone to deliberately distort a quote to mean something other than that intended by the author isn't very honest, is it?

What I actually wrote (my added emphasis) was this:

Does anyone really believe that we will have safe and reliable FSD on any car driving around on normal public roads anytime soon?

Yep.... I spoke and made my point in the context of public roads FSD. Motorways have already been solved so It's not really up for debate.

 

[Dilly]

The one thing I can't seem to see on these new beta's is whether they have fixed what I believe is a fundamental problem - the fact that autopilot or FSD will hurl the car towards a speed restricted zone only slowing well after the new speed sign has been passed. I can see that the nav 'knows' when it is entering a new speed zone so why can't it slow in preparation (like humans do) rather than wait until the cameras detect the sign?

Pre- speed limit recognition, the change was well past the signs. I’m finding that changes are now at the signs. It would be nice to see a slowing down as the lower limit signs are approached.

 

[Glan gluaisne]

Yep.... I spoke and made my point in the context of public roads FSD. Motorways have already been solved so It's not really up for debate.

You may think that deliberately misquoting in order to make someone out to have a view that they do not hold is acceptable. I do not, it's the sort of behaviour I associate with the grubbier elements of the media (and, perhaps, politicians).

Have motorways been "solved"?

Do we have fully autonomous cars driving along our motorways now?

Right now, AFAIK, the only fully autonomous cars driving on UK roads are test vehicles in Milton Keynes and Greenwich. IIRC, one developer has been permitted to do a trial from Oxford to London (it wasn't Tesla). I'm far from convinced that even the current state of the art vehicles from Tesla, MobilEye, Waymo, GM, Ford, Nissan etc will be permitted to drive fully autonomously (as Waymo do now in Phoenix, Arizona) on UK roads any time soon, so stating that "Motorways have already been solved" seems a wee bit premature.

The topic of FSD seems very much up for debate to me, after all, it's the key reason this thread exists.

 

[Yev000]

Ok, I'll make a point not to quote you then.

I find this "debate" amusing. Not exactly sure what the point of the "debate" actually is... To figure out if someone else will solve autonomy? To take a guess at who? To pick the right company to invest in? To figure out who to look at for a career?

All I see here is people arguing absolutes. 10 year this, 5 year that, this isnt solved that is etc....

One can argue that something isnt solved because its not quite up to that individuals' standard in perpetuaty.

Again, I don't see the use in debating these points.

My point is rather simple. *IF* FSD (in the context of city street driver assist) is following a disruption curve it will be solved in the next 1-3 years. *ELSE* it's not good enough in the current itteration and it will never be solved with this software/hardware combo.

This is what happened with AP - software component was not good enough, hence the re-write.

Once they pass that hurdle, it's on to robotaxis. Thats more of a regulatory thing

And to my other point - the only reason to debate this (internet entertainment aside), is to prepare for your next career/investment. It really wont affect peoples car purchases in the long term - because fully autonomous cars will crash personal car sales, people will stop buying cars to commute.

 

[browellm]

The one thing I can't seem to see on these new beta's is whether they have fixed what I believe is a fundamental problem - the fact that autopilot or FSD will hurl the car towards a speed restricted zone only slowing well after the new speed sign has been passed. I can see that the nav 'knows' when it is entering a new speed zone so why can't it slow in preparation (like humans do) rather than wait until the cameras detect the sign?

Even a manual option to slow down x seconds before the sign would do fine. A bit like the car lengths option for TACC. This would then let people adjust for their local market/local area conditions accordingly.