FSD - rounadabouts and T-junctions

[GeorgeSymonds]

That will still get you to Level 4.

I suggest you read up on what level 4 means.

At best it might get to level 3 with the car stopping and saying "driver please step in - I'm stuck" and the driver being ready to do so - not drunk in the back of the car having a kip.

Level 4 is full autonomous wiithout intervention being required but on a subset but defined set of roads. Likely to be roads such as motorways. I read an article I can't now find that talked about Tesla prioritising badly the decision, or they're doing it because they can't achieve genuine self driving. I think the crux was if they could achieve level 4 but just on motorways in the UK (freeways or highways whatever they're called abroad) and get the driver to do the first and last mile, then that would be mega news, instead they're trying to do lots of edge case, junctions, stop signs, traffic lights etc all of which have to be perfected to move to level 3 let alone 5

Which do you think is the better transient step? Being able to do street signs, turn across traffic, merge etc but all the time the driver holding the wheel and ultimately responsible, or they trow all their efforts into a system where the driver carries on doing all that manually, but on the motorway can lean back, watch a film, have a nap for the few hours the car is trundling along a motorway? The fact Tesla aren't nailing motorways and level 4 could imply they don't believe they can get there as it would be the bigger ticket win in the next year.

The kudos will be the first company that commercially sells a car where the car is responsible for some of the driving and not the driving, and the utility benefit for the driver will be long trips. So why not go that route and not give a damn about spotting somebody with a lollypop letting kids cross the street and being a bit flacky in the detection?

 

[MrBadger]

I suggest you read up on what level 4 means.

I know exactly what Level 4 means.

 

[Alistairuk]

Let’s not forget about the thousands of humans who are incapable of dealing with roundabouts in the UK...

If all cars were autonomously driving it would be easy to solve - unfortunately having to take into account idiots not doing what they’re supposed to be doing is the real challenge!

 

[Rob R]

I am really doubtful about FSD's ability to negotiate UK roundabouts in acceptable fashion. Safe, yes, simply by not going until there are no cars visible to the right. But useable in normal conditions, I doubt it.

The other bugbear I have driving on autopilot or even on smart cruise is parked cars in the lane. My Model 3 literally has a mental breakdown in this situation. How can you teach a Tesla when a car is parked, or when it is stopped in traffic? It is subtle. Sometimes I get it wrong, after 40+ years driving on humanpilot.

 

[Yev000]

My Model 3 literally has a mental breakdown in this situation. How can you teach a Tesla when a car is parked

Once re-write is 4D it will be able to tell if the car is moving much easier. Then parked cars dont have lights on, so maybe that?

Not sure what their plans are for "city streets" but someone must be thinking about this.

 

[MrBadger]

The other bugbear I have driving on autopilot or even on smart cruise is parked cars in the lane. My Model 3 literally has a mental breakdown in this situation. How can you teach a Tesla when a car is parked, or when it is stopped in traffic? It is subtle. Sometimes I get it wrong, after 40+ years driving on humanpilot.

Check out this video by GM Cruise just after their vehicle makes its first right turn.

https://twitter.com/i/status/1307092038765215746

 

[DrJFoster]

GM Cruise use a larger sensor suite than Tesla from what I can see on their website, they certainly list Lidar which Tesla dismissed. Isn't that the point some of us have made that the sensor suite could be inadequate on the Tesla?

 

[MrBadger]

GM Cruise use a larger sensor suite than Tesla from what I can see on their website, they certainly list Lidar which Tesla dismissed. Isn't that the point some of us have made that the sensor suite could be inadequate on the Tesla?

That video shows is more about cruises driver policy than it does their vision capabilities - it was in response to comments abour how to handle parked cars. Tesla really need to work on their driver policy as it is lack of maturity in that that probably accounts for many issues than significant shortfall in the sensor suite.

Mobileye also have systems that are not dependent on Lidar and do a very good job without. This is their 12 camera driven system. No radar, no Lidar. For comparison, Tesla have 8 cameras and radar.

 

[DrJFoster]

Mobileye also have systems that are not dependent on Lidar and do a very good job without. This is their 12 camera driven system. No radar, no Lidar. For comparison, Tesla have 8 cameras and radar.

Its difficult to tell if you acknowledge the Tesla suite may be inadequate or sufficient? ("may" as in a possibility, yet to be determined) I think you've asserted they can still achieve level 4 but I'm not sure what you're basing that on other than Tesla keep promising things, and have done for 4 years to my knowledge.

I've seen the Mobileye eyeq5 system videos, and I've driven their eyeq3 (AP1) and eyeq4 (BMW) systems and eyeq4 is pretty much as capable as Tesla FSD is today (tacc, lane keep, lane change, stop sign, speed limit, traffic light recognition, traffic light obedience, emergency lane finding etc). They even have summon and a surprisingly useful automated reverse driving feature, but they (BMW) are not claiming they'll be doing self driving with it even though mobileye say its suitable for level 3 (which as you know is the first stage where the car will take some responsibility for the driving, a significant milestone requiring the legal ramifications to be sorted out). The Mobileye eyeq5 system does have more cameras to Tesla's 8, and it seems they've changed their position saying Lidar is required as part of the sensor site for "Autonomous driving"

The Evolution of EyeQ

So that's two competitor systems, both with more comprehensive sensor suites. While there's way more to it than just the sensors, if you've not got the raw data to work with, you can't solve the problem. Musk will never let go of the claim, he has too much vested in it, he'll look to regulators and others claiming its their fault, and the internet will be full of videos with people getting from A to B using it with minimal interventions, but its those interventions that need to shift to being highly predictable, and in good time to allow any autonomy with a managed hand back to the driver. You simply can't have a car suddenly go beep beep and expect the driver to step in with no notice having not been required to pay attention during the lead up to it.

Tesla also need to do more about their dodgy safety stats. I'm sure AP has fewer accidents per thousand miles driven, but I also know that UK motorways are 4x safer than other roads overall, so given I'm most like to use AP on a motorway in ideal conditions and turn it off when the world gets tricky (and therefore higher risk), I'd expect the figures to be a minimum of 4x safer if not higher, just down to the nature of the road I use active AP on. But that's another story and I only mention because Musk at one point was claiming he just needed to be twice as good as a human driver and those stats were used to try and demonstrate thats where they were.

 

[MrBadger]

Its difficult to tell if you acknowledge the Tesla suite may be inadequate or sufficient? ("may" as in a possibility, yet to be determined) I think you've asserted they can still achieve level 4 but I'm not sure what you're basing that on other than Tesla keep promising things, and have done for 4 years to my knowledge.

I'm not totally green on the subject and can make my own assessments of the state of play without working on hype and false promises. For part of my recent previous life I use to work in research and development sector, part of which covered autonomous vehicle technology and mobility solutions. Whilst not directly involved, there was a high level of information exchange between at both official and unofficial level. The level of our involvement was high, as in, we were involved in many significant projects that were absolutely cutting edge and possibly yet to see the light of day. We were running fully autonomous vehicles carrying the public in public spaces and I believe they still do. As part of the health and safety committee, I also got insight into risks involved. Add full research access to top academic online lectures and resources, I got a pretty good idea of what, how and when things were going on. Certainly enough to make my own judgement calls based upon the evidence at hand.

I think there's a fair chance that while the sensors may do a good job (perhaps more so on Elon's route to work), it's likely they'll need another overhaul or three before we get to autonomous all the time every time. Not least to work around obstructions and weather.

That will still get you to Level 4.

No assertion here that Tesla will get to level 4, but they may well do. Level 4 does not define what the operational design domain (ODD) of the autonomous capabilities are. Whilst the car is capable of driving itself and no human intervention or focus is required during that period, the human driver may still be required to take control under certain circumstances outside the ODD. This may be types of roads or climatic conditions. For example, a vehicle can still be Level 4 if it can only fully autonomously drive on motorways in good weather conditions. That was the point of my post. A level 4 vehicle does not need to "work around obstructions and weather" - it can gracefully hand back control as required.

As for Teslas sensor suite. Who knows whether it will be good enough? At present I have seen no evidence that the sensor suite in its current form will preclude Level 4. Yes, it is affected by bad weather for example, but Level 4 does not mean that the car must operate autonomously in all weather conditions. Navigate on Autopilot already reduces its capability in poor weather and notifies the driver when it is leaving its ODD. Much the same may occur if the vehicle was Level 4 capable.

From my experience, the lack of a coherent visual view is really holding the current capabilities back and I am surprised that 1/ Tesla got this far without having it and 2/ have only just got around to implementing it as a coherent view tracking objects over time is critical to the success of accurate self driving. I have said on many occasions that this autopilot rewrite is the last chance saloon for Tesla and their FSD capabilities. imho They will have nowhere to hide and will have to get their path planning.

Until that happens, I think it is too early to speculate on whether their current sensor suite is up to the task and how many 9's (as in 99.9% or 99.99999% accuracy) it will achieve and whether that will be acceptable. I personally think that V2I (I wont go as far as saying V2X at present due to lack of other vehicles having V2V capability) and possibly rear radar are two tools that Tesla should add to their armoury. But I can also see an argument for higher resolution cameras should the AP rewrite not give the special accuracy that it needs. But too early for that call. I'm not a fan of Lidar and static high resolution mapping that many also think are important to full autonomy - I've done a fair bit of large scale mapping for city wide transport modelling and from experience even at that small scale, problems are sure to creep in over time.

Personally, I think Tesla will achieve Level 4 driving for some scenarios. But I'm not going to speculate how big their ODD will be.

imho.

 

[phil4]

Sounds like far more of an educated take on it than I offered

 

[Wol747]

The basic features are still a way off too, interesting read about the NHTSA list of requirements and an assessment of where Tesla are against them. Quite a few make you stop and thinkTesla FSD and Feature Complete

An interesting piece!

It's long been my opinion that FSD is not going to come in my lifetime on normal roads.

It's obviously immensely difficult to make a car drive itself on motorways but in all weathers on every driveable route is orders of magnitude more difficult. The "low hanging fruit" is still not 100% and the chances of what is called AI (which in Tesla's case is actually nothing of the sort) being up to self driving taxis is wishful thinking.

I'd like to be wrong - I sunk $8500 into FSD just for the fun part - but I would take a largeish bet on it!