FSD Edge cases

[gaspi101]

After testing FSD Beta for a little while, I’ve encountered a few edge cases where I am not sure whether a solution is possible at all. For example, if you are driving on a one lane road, and you encounter a stopped vehicle, FSD beta invariably chooses to leave the lane and go onto the opposite lane in order to go around the stopped vehicle. however, it’s often the case that the stopped vehicle is only the last in a long string of stopped vehicles in gridlock traffic. Or perhaps traffic on a one lane road that is turning. on my commute to and from work, it’s decisions in this regard or approximately 90% incorrect, and I have narrowly avoided a few accidents after the car chose to go around at full speed.

The decision to go around a stopped car on a one lane road is purely contextual. If there is a traffic light a quarter-mile ahead, and you can see that the number of cars is constant bumper-to-bumper, clearly that is not a case where you go around. But other situations are much more subtle, where there could be a stopped car that wants to either make a U-turn (no turn signal, but you see that its wheels are turned all the way to the left, ready to make that U-turn) or maybe the stopped car wants to turn left and is waiting for the opportunity to do so since there is oncoming traffic… perhaps that traffic is a good deal away, but the driver is elderly, so you don’t want to blow past them and scare them, or worse, cause an accident ….or maybe the stopped car is waiting to allow a pedestrian or a cyclist or a school bus to clear the way up ahead. Human drivers notice this through contextual clues, maybe by looking at the driver him/herself— contextual clues are nearly infinite, and must be nearly impossible to quantify.

reminds me of this excellent Tom Scott video where he discusses how computers may be unable to solve human language because of their inability to resolve contextual clues. I feel it must be the same with autonomous driving.

how can Tesla solve this with as much certainty (or more!) than a human? Is that even possible? The human ability for pattern recognition based on an almost infinite number of contextual clues really seems like an unsolvable problem to me… and that’s just for the simple situation of going around a stopped car! What do you folks think? Have you found similar cases like this? What other edge cases do you think are unsolvable, or perhaps I am wrong in my thoughts here?

 

[nvx1977]

i posted an intersection where humans commonly fail at. so if FSD can solve this some day, I'll be impressed.

Post locations where you think FSD (beta or not) will fail spectacularly

As I drive these days, I frequently encounter really poor road designs that sometimes bamboozle humans. And I wonder how the heck is FSD possibly going to make the right call. Here's an example: https://www.google.com/maps/@42.8058138,-71.2452106,183m/data=!3m1!1e3 If you're driving...

teslamotorsclub.com

 

[gaspi101]

i posted an intersection where humans commonly fail at. so if FSD can solve this some day, I'll be impressed.

Post locations where you think FSD (beta or not) will fail spectacularly

As I drive these days, I frequently encounter really poor road designs that sometimes bamboozle humans. And I wonder how the heck is FSD possibly going to make the right call. Here's an example: https://www.google.com/maps/@42.8058138,-71.2452106,183m/data=!3m1!1e3 If you're driving...

teslamotorsclub.com

Thanks, maybe the mods can move my post as a comment to yours, but what I’m really interested in is not so much a collection of edge cases, but rather to see if anyone has any insight on how resolving the problem of context-dependent problems can possibly be solved… if I could understand the process by which we can achieve this, then I can believe that we will have actual full self driving within my lifetime. But it seems to me that the only solution is either to create an environment that’s much more predictable, such as driving on tracks, or the full self driving computer has to be semi self-aware (!!!). I watched and re-watched the lectures at the Tesla AI Days, and despite all the technical jargon discussed, I have not seen them discuss this problem well enough.

 

[Daniel in SD]

Not really an edge case if it's something that every driver encounters weekly. To me an edge case is something that you encounter once in a lifetime (which if you consider the variety of bizarre things that can happen isn't all that unlikely).
I agree that cars stopped in traffic lanes seems like the hardest common situation to handle. It's a case that other self-driving car companies use remote assistance for so I bet no one else has solved it reliably enough either.

 

[gaspi101]

Not really an edge case if it's something that every driver encounters weekly. To me an edge case is something that you encounter once in a lifetime (which if you consider the variety of bizarre things that can happen isn't all that unlikely).
I agree that cars stopped in traffic lanes seems like the hardest common situation to handle. It's a case that other self-driving car companies use remote assistance for so I bet no one else has solved it reliably enough either.

I suppose the question I’m trying to get to his weather L4/L5 autonomy is ultimately a pipe dream, and far far beyond our current technological capabilities. I’m thinking 2058, Elon announces HW15.0 retrofit, which is a repurposed human brain in a jar with cables sticking out.

 

[EVNow]

I suppose the question I’m trying to get to his weather L4/L5 autonomy is ultimately a pipe dream, and far far beyond our current technological capabilities. I’m thinking 2058, Elon announces HW15.0 retrofit, which is a repurposed human brain in a jar with cables sticking out.

We tend to overestimate what can be done in the short term and highly underestimate long term.

 

[Daniel in SD]

I suppose the question I’m trying to get to his weather L4/L5 autonomy is ultimately a pipe dream, and far far beyond our current technological capabilities. I’m thinking 2058, Elon announces HW15.0 retrofit, which is a repurposed human brain in a jar with cables sticking out.

My opinion is that L5 is not happening without a fundamental breakthrough in AI.
Very limited L4 already exists and will continue to slowly improve but it doesn't seem like it will be practical for consumer owned vehicles for a very long time. Personally I think Tesla should focus on highway only L3 as that may be achievable and there is a risk that competitors will get there first.

 

[DaveG_NJ]

reminds me of this excellent Tom Scott video where he discusses how computers may be unable to solve human language because of their inability to resolve contextual clues. I feel it must be the same with autonomous driving.

I used to do contract IT work and, being up there in years, never knew if, at the end of the contract, I was unemployed or retired. I did some gig work and also worked as an Amazon Mechanical Turk, which, for those who are unfamiliar, is a program run by Amazon to perform crowdsourced tasks (for pennies) that cannot be done by a computer. Much of it has to do with image recognition and understanding the context of the scenes. I think I made about $250/month. Know those annoying Captcha images asking you to identify stoplights or crosswalks? There you go.

So, in my testing of FSD, I see two main flaws, or "opportunities for improvement" as HR would say. One is having a long view of the road. It it entirely too focused on what is immediately in front of it. I can see and react to panic braking up ahead on the highway or an adjacent lane - or a complete stoppage of all lanes with emergency vehicles, but the Tesla will barrel blissfully toward a jam until something is a danger, then it will slam on the brakes.

What I believe will be harder, and I'd say, at this point, not solvable, is understanding the context of the car's environment. I don't think it will ever have the situational awareness to make decisions that involve highly contextual inputs - the behavior of other cars (driven by emotional humans of various skill levels). I can spot vehicles on the road that I know, by their behavior, are going to cause trouble. Or why a vehicle is stopped on the side of the road or blocking a lane. Backed up traffic? Delivery? Mechanical breakdown? Turtle in the road?

I will admit I'm inconsistent on my feelings about FSD. I love my Tesla (and actually driving it) and the technology is fascinating, but I cannot see the value in it. Don't get me wrong, I think active safety measures are very important. But FSD is like those Rube Goldberg device contests - fun for the challenge of making a machine that can do something that really doesn't need to be done by a machine, like a room-sized apparatus to slice an apple. And just try explaining FSD to my wife. She thinks its the most absurd thing she's ever seen. LOL.

 

[TessP100D]

After testing FSD Beta for a little while, I’ve encountered a few edge cases where I am not sure whether a solution is possible at all. For example, if you are driving on a one lane road, and you encounter a stopped vehicle, FSD beta invariably chooses to leave the lane and go onto the opposite lane in order to go around the stopped vehicle. however, it’s often the case that the stopped vehicle is only the last in a long string of stopped vehicles in gridlock traffic. Or perhaps traffic on a one lane road that is turning. on my commute to and from work, it’s decisions in this regard or approximately 90% incorrect, and I have narrowly avoided a few accidents after the car chose to go around at full speed.

The decision to go around a stopped car on a one lane road is purely contextual. If there is a traffic light a quarter-mile ahead, and you can see that the number of cars is constant bumper-to-bumper, clearly that is not a case where you go around. But other situations are much more subtle, where there could be a stopped car that wants to either make a U-turn (no turn signal, but you see that its wheels are turned all the way to the left, ready to make that U-turn) or maybe the stopped car wants to turn left and is waiting for the opportunity to do so since there is oncoming traffic… perhaps that traffic is a good deal away, but the driver is elderly, so you don’t want to blow past them and scare them, or worse, cause an accident ….or maybe the stopped car is waiting to allow a pedestrian or a cyclist or a school bus to clear the way up ahead. Human drivers notice this through contextual clues, maybe by looking at the driver him/herself— contextual clues are nearly infinite, and must be nearly impossible to quantify.

reminds me of this excellent Tom Scott video where he discusses how computers may be unable to solve human language because of their inability to resolve contextual clues. I feel it must be the same with autonomous driving.

how can Tesla solve this with as much certainty (or more!) than a human? Is that even possible? The human ability for pattern recognition based on an almost infinite number of contextual clues really seems like an unsolvable problem to me… and that’s just for the simple situation of going around a stopped car! What do you folks think? Have you found similar cases like this? What other edge cases do you think are unsolvable, or perhaps I am wrong in my thoughts here?

They can’t solve it. FSD is a complete fraud.

 

[EVNow]

I don't think it will ever have the situational awareness to make decisions that involve highly contextual inputs - the behavior of other cars (driven by emotional humans of various skill levels).

As I keep saying, ever is a long time.

 

[gaspi101]

They can’t solve it. FSD is a complete fraud.

You might be right. But I will tell you this: I have been humbled several times in my life by believing something was outside the realm of feasibility or even possibility, only to be proven dead wrong. When I saw the first iPhone, I thought it was absolute magic. When SpaceX started landing boost rockets on earth, I was utterly stunned. I‘d like to think I’m a pretty smart person…but I have no illusions—I’m a complete knuckle-dragger in comparison to people like Elon Musk. I just wish there were someone around much smarter than me that could dumb it down to my level so I could understand how the hell Elon plans to get there…

 

[hugh_jassol]

Saying it can't be done is just ridiculous. Flying was once considered impossible. As was space travel. So... "can't get there" is silly. Of course it will happen. Can it happen with the hardware currently in the cars? Can it happen in the next decade? or two? So "when" is the only real question.

 

[jakimo]

..which is a repurposed human brain in a jar with cables sticking out..

The real purpose of Neuralink, eh? =)

 

[ChimpledPot]

Of course it’s going to be possible. Our brains can do it and our brains are of this world just like anything else you can touch.

 

[idriveacar]

My opinion is that L5 is not happening without a fundamental breakthrough in AI.
Very limited L4 already exists and will continue to slowly improve but it doesn't seem like it will be practical for consumer owned vehicles for a very long time. Personally I think Tesla should focus on highway only L3 as that may be achievable and there is a risk that competitors will get there first.

Yeah. L4 systems will improve, but the ones that become commercially viable in the short term, IMO, rely on things that do not apply to consumer vehicles: namely, remote assistance, and purpose built hardware.

I think FSD could be a good L3 system one day. The only concern I have is the research into L3 that indicates it will not reach higher than human safety due to the difficulty of immediate mode switch. If a person is allowed to not pay attention in a scenario, how much warning do you think the car needs to be able to give the human in order to resume control safely? To my understanding, L3 does not give a set time limit; yet I think it would be entirely unfair if a car gives control back to a human 2 seconds before an imminent crash. That's effectively useless.

If the L3 system is as good at preventing crashes as the L4 robo-taxi systems, and yeilds control back to the human in the same way that robo-taxi services will query a human operator, this could work. I think that the lag introduced by remote operators is a good analog to a "fair" amount of lead time of an L3 requesting the driver make a decision. The L3 system, like the L4 robo-taxies, will have to only get "stuck" at high level, low speed policy decisions, and never get "stuck" on a decision that is life or death over the course of the next few seconds.

I don't think that this is an unsolvable problem, for what it's worth. But when discussing L3 I think it's one of the larger unsolved ones. I would be happy with a door to door L2 system in my MY, and even happier if it proves igself to be safe enough that Tesla enables it to be L3 (thus assuming responsibility for it when it's active). But they have to solve the handoff problem first.

To the point of the thread, I really do think we're far off from consumer level, operatorless, L4 and especially L5 technology. The self driving car detractors do have valid points. What they lack, however, is the vision to see what SDC technology could do today if only a few limits are imposed on it.

 

[idriveacar]

Of course it’s going to be possible. Our brains can do it and our brains are of this world just like anything else you can touch.

The way I take your comment to mean is that if we can build a computer that behaves like a brain, it could be harnessed to solve the same problems humans can, except it will not ever make mistakes and be infinitely faster in decision making.

This is a very common, likely false, line of reasoning. It's one of the things that bothers me about how androids and AI are represented in movies.

Computers and humans think in fundamentally different ways. Computers are nothing more than very, very, fast calculators. Brains are more like... super sophisticated, unfathomably complex pattern matching machines, that self mutate and rearrange based off of rules that aren't entirely understood.

If you build a brain out of a computer, you don't end up with the best of both worlds. As the artificial brain gains that fuzzy contextualization and pattern matching capabilities, its ability of executing rote logic and calculations is effectively sacrificed.

You could then try to teach that artificial brain how to drive, but now you are working with a (probably much, much, worse!) brain like intelligence that is prone to making mistakes and getting distracted, just like a human. If you let that brain self mutate and learn during operation, now it's prone to going insane!

Maybe humans will discover the secrets of the brain one day, and learn to make one that has the attributes we want without the ones we don't. But that's not happening in my lifetime, so I feel free to speculate. My speculation is that the things that make humans great at solving highly contextual situations on the road are fundamentally the exact reason why humans suck at solving the rote, boring, tasks. Pattern matching and precise calculation are mutually exclusive, and perfect mechanical driving at all time is mutually exclusive to highly intelligent, context aware, driving.

tldr; if we ever can create a brain that behaves like a human brain, it will suffer the same issues that make human brains unsuited to the task of driving. The real solution is a mix. But the proof that such a system can exist (which may or may not itself exist!) does not follow from the fact that brains are physical.

 

[gaspi101]

Yeah. L4 systems will improve, but the ones that become commercially viable in the short term, IMO, rely on things that do not apply to consumer vehicles: namely, remote assistance, and purpose built hardware.

I think FSD could be a good L3 system one day. The only concern I have is the research into L3 that indicates it will not reach higher than human safety due to the difficulty of immediate mode switch. If a person is allowed to not pay attention in a scenario, how much warning do you think the car needs to be able to give the human in order to resume control safely? To my understanding, L3 does not give a set time limit; yet I think it would be entirely unfair if a car gives control back to a human 2 seconds before an imminent crash. That's effectively useless.

If the L3 system is as good at preventing crashes as the L4 robo-taxi systems, and yeilds control back to the human in the same way that robo-taxi services will query a human operator, this could work. I think that the lag introduced by remote operators is a good analog to a "fair" amount of lead time of an L3 requesting the driver make a decision. The L3 system, like the L4 robo-taxies, will have to only get "stuck" at high level, low speed policy decisions, and never get "stuck" on a decision that is life or death over the course of the next few seconds.

I don't think that this is an unsolvable problem, for what it's worth. But when discussing L3 I think it's one of the larger unsolved ones. I would be happy with a door to door L2 system in my MY, and even happier if it proves igself to be safe enough that Tesla enables it to be L3 (thus assuming responsibility for it when it's active). But they have to solve the handoff problem first.

To the point of the thread, I really do think we're far off from consumer level, operatorless, L4 and especially L5 technology. The self driving car detractors do have valid points. What they lack, however, is the vision to see what SDC technology could do today if only a few limits are imposed on it.

Do you think that the full self driving system Tesla has been promising for the last 6 to 7 years will be feature complete an out of beta at L3 what do you think that Tesla needs/will get to L4?

 

[idriveacar]

Do you think that the full self driving system Tesla has been promising for the last 6 to 7 years will be feature complete an out of beta at L3 what do you think that Tesla needs/will get to L4?

So there are three promises:
1. Elon's different verbal promises
2. purchase page circa 2017 After what time has passed would you consider an FSD class action lawsuit?
3. purchase page today (doesn't seem to have much specificity, just "autosteer on city streets")

For 1, Elon clearly promises L5 autonomous driving. I do not believe that this is happening until we see a breakthrough in generalized artificial intelligence. Are these statements legally binding? I'm not sure. But his claims are grossly exaggerated.

Elon also verbally promised robo-taxies. I don't see this happening before they say FSD is feature complete. I believe they will also need a network of remote operators to make any system that allows the driver to leave the driver's seat.

Personally, I see FSD as being marked as feature complete when door to door L2 is deployed to the fleet. Tesla never used levels in their marketing that I'm aware of. I'm only aware of a single verbal comment by Elon that referred to L5. Also all of the marketing material I posted above does refer to somebody in the driver's seat. That eliminates L4 and L5 anyway.

Now L3 is interesting. I think L3 is possible if they demonstrate that their system can always give "enough" of a warning before requiring human input. What "enough" is needs to be the topic of research. They will have to prove this over the course of many millions of miles. Then Tesla themselves can decide if they want to accept the responsibility of letting the drivers not pay attention. Since the L2->L3 distinction comes down to liability and not technology, I think FSD will be marked "complete" before this stage.

I don't believe assistanceless L4 will be possible for a long while because I don't belive that SDCs can get parked in all the situations where human judgement is required. I think that this is the big challenge. As long as a car can get stuck in a situation where it isn't possible to change drivers, L4 will simply not be possible without remote assistance. Maybe that assistance could come from a user who's taking a nap in the back of the car.

Another question I won't go into (it's been beaten to death elsewhere) is that of hardware. Is HW3 sufficient for the L2 door to door that (I believe) will fulfill their promise? I'm not sure.

 

[eli_]

I think the biggest issue is they have this strict separation between perception and planning. Like they don't go straight from perception to planning/control. They go from perception to this hand picked, human-understandable list of objects that AP engineers have decided are important for driving, i.e. cars, pedestrians, bikes, lanes, signs, garbage cans, lane markers, cones, etc. and then they build a plan based on that. But all the richness in the visual field that humans use to drive isn't present anymore, it gets lost in this abstract vector space of objects and bounding boxes and lines. Sure you can drive decently well based on that alone, but there are probably lots of cues that humans use when driving that are being lost because there's no way to comprehend what they are. Like you see a bush obscuring something, you know something may come out from behind it, or you see a moving shadow cast by an object out of view, or simply sizing up the whole situation as needing extra caution, that sort of thing is really hard/impossible to explicitly encode into some vision model.

I think at some point they will throw out what they have for more end-to-end models, I don't think you can ever be truly end-to-end, but maybe some aspects of the planning will go that way.

 

[dcesq06]

The human ability for pattern recognition based on an almost infinite number of contextual clues really seems like an unsolvable problem to me…

Reminds me of the documentary AlphaGo, where the software improved over time and got increasingly better at learning and decision-making - to the point where it was able to master a game in which there are 10 to the power of 170 possible board configurations.

AlphaGo

Novel AI system mastered the ancient game of Go, defeated a Go world champion, and inspired a new era of AI.

deepmind.com

AlphaGo (2017) ⭐ 7.8 | Documentary, Sport

1h 30m

www.imdb.com