Neural Networks

[mongo]

navigating traffic signs/light has nothing to do with defining level 2/3/4.

Car and drivers definition is HUGELY incorrect which is not surprising seeing almost every statement about SDC from the media is filled with misinformation. That is why you should go to the actual SAE document.

But here's a good explanation by toyota at 20 minutes

Thanks for the better source.
Getting back to my question, would stopping at a stop sign be a different level that leaving a traffic signal (possibly 2 for stopping and 3 for going)? Or is that just variations on the level of monitoring (both 3)?

 

[arg]

Thanks for the better source.
Getting back to my question, would stopping at a stop sign be a different level that leaving a traffic signal (possibly 2 for stopping and 3 for going)? Or is that just variations on the level of monitoring (both 3)?

I don't believe that specific piece of behaviour definitively puts it in any category.

Stopping at a stop sign and never continuing could even be Level 4 - for the very limited set of driving modes "does not pass stop signs": it has driven you safely up to the stop sign and left you in a safe condition even if you didn't respond.

Conversely, stopping at a stop red light and continuing when it turns green might only be level 3 depending on the confidence with which it did it - if you are still supposed to be monitoring it to see if it has got it right, then it's not level 4.

Given that all these modes (short of 5) are subject to limitations, the higher number doesn't necessarily mean it's a better, more useful or more impressive system. A system that achieves level 4 in an extremely narrow set of driving modes (eg. up to 10mph, following other traffic) may be relatively easy to do and less useful than one that achieves level 3 across a wide variety of modes.

 

[Bladerskb]

would stopping at a stop sign be a different level that leaving a traffic signal (possibly 2 for stopping and 3 for going)? Or is that just variations on the level of monitoring (both 3)?

No they are both still level 2. The lack of Driver monitoring is the only thing that makes a system level 3 and above versus just being level 2.

 

[SucreTease]

Summoning @buttershrimp, our resident @Balderskb exterminator.

 

[lunitiks]

Summoning @buttershrimp, our resident @Balderskb exterminator.

Do you REALIZE what you just DID?????!

 

[Dan Detweiler]

What is "closer than many believe"?

and why is it "closer"?

I have no empirical evidence, this is just my opinion but I think FSD is about a year away...maybe 2. I believe Tesla will indeed demonstrate the cross country drive by the end of this year and will make it available fleet wide by mid 2019. Of course legislative limitations may certainly limit the use of said technology but I think the technology will be there. I believe LIDAR based systems will be limited to specific areas or routes that can be more easily mapped but with its inability to read road signs and otherwise interpret changes in the driving environment I don't see it being used as a universal self driving platform.

Again, just my opinion, but I think the photo recognition route will be more accurate and sustainable in the long run. LIDAR systems will be the first to be used as has been recently demonstrated, but I think their limitations will be their overall downfall.

As with all things "self driving" time will tell.

Dan

 

[SucreTease]

Do you REALIZE what you just DID?????!

RELEASE THE KRAKEN!!!

 

[malcolm]

No they are both still level 2. The lack of Driver monitoring is the only thing that makes a system level 3 and above versus just being level 2.

Quick! You better inform Mobileye of this.

Autonomous Driving - Mobileye

 

[mongo]

No they are both still level 2. The lack of Driver monitoring is the only thing that makes a system level 3 and above versus just being level 2.

I would have thought that being able to make a good decision to start moving would require the system monitoring the environment (level 3). Differentiated from current systems which require the driver to press the accelerator to resume adaptive cruise (level 2) after being stopped.

@arg I suppose I'm thinking of a philosophical question of what it means to be monitoring the environment. Noticing a stop sign is monitoring, noticing cross traffic is a higher level of awareness. For 4, it needs to either exclude a situation, or monitor all aspects of it. Go/no-go decisions seem level 3. Stop and wait for input seems 2ish.

Like you said, it can be different levels in different situations. I suppose situation and response tables are better for classifying than a single number.

 

[jimmy_d]

Good luck getting the update its being sent like crazy, so if you're on wifi you will get it (or use mobile hotspot). This one is worth it. Really impressive (the superlatives come easy with this build).

Drove SF->LA today and received 18.10.4 at the halfway point (Kettleman) when I stopped to charge. I tried the hotspot trick you suggested this morning, so maybe that was what did the trick. Of course I installed it immediately. And it turned out to be a nice natural experiment since I drove 200 miles on 2018.6.1 and then immediately another 200 on 2018.10.4.

Wow. It was perfect. It worked in places that I've seen fail 100% on both AP1 and previous AP2. No hill cresting fails - zero. No lane hugging - none. It handled the full grapevine and all the nutty construction zones and the tail and did it at regular traffic speed (about 20MPH over the posted limit). It did all the interchanges at highway speed. It even handled the 5-210 interchange which is a tight, 3 lane, decreasing radius curve on a concrete ramp which has all the lane markings completely worn off - without losing it's position in the middle lane. I think I must have been cackling maniacally at that point because I was sure it couldn't do that.

I'm just flabbergasted. It's not even a contest now - for me this thing just crushes AP1. It's just one drive, of course, but it's the hardest thing I regularly drive and it's been my acid test for a couple of years now.

Man, I am so glad I upgraded to the new car.

Holy cow.

A lot more has changed here than just the vision network. The path prediction algorithm must be completely new.

 

[croman]

@jimmy_d -- Glad you got it and can see for yourself. Despite the string of constant superlatives until you experience it, its hard to describe how different it is from 2018.6.1 and before. Closest thing I can analogize to is the massive improvement we saw with .40/.42 (another Karpathy "fix" of prior code).

I have a lot less experience with AP1 than you but I also feel glad I got AP2 for the first time since I took delivery way back in Dec 2016. It is so much better but I also see how they can make it better and better and that has me even more excited.

I'm looking forward to hearing more as the NNs are refined and the repeater cams are activated with the current NN.

 

[chillaban]

Man, I am so glad I upgraded to the new car.

Holy cow.

A lot more has changed here than just the vision network. The path prediction algorithm must be completely new.

Totally agree! I switched from AP1 to AP2 exactly a year ago and this feels like a huge milestone in the journey and the first time I don’t miss something about AP1.

I saw something interesting today when it came to path prediction. Was driving on a city street in the right lane, which is super wide and allows street parked cars.

On blocks with no street parked cars, AP2 now detects a really wide lane and drives in the center. However, on another block with a couple street parked cars, it instead detected a narrow lane that drives comfortably by the parked cars. I saw it driving in wide lane mode, see the car up front, and the detected lane width shrank and the car moved over.


All this leads me to suspect that this new vision system isn’t just reading lane markers, but it’s also taking other cues like the presence of other cars to decide where the lane is.

All in all it’s a huge step up. I also found it able to take some ridiculously challenging on ramps and even navigated a cloverleaf and correctly slowed to 25mph for it!

 

[croman]

Totally agree! I switched from AP1 to AP2 exactly a year ago and this feels like a huge milestone in the journey and the first time I don’t miss something about AP1.

I saw something interesting today when it came to path prediction. Was driving on a city street in the right lane, which is super wide and allows street parked cars.

On blocks with no street parked cars, AP2 now detects a really wide lane and drives in the center. However, on another block with a couple street parked cars, it instead detected a narrow lane that drives comfortably by the parked cars. I saw it driving in wide lane mode, see the car up front, and the detected lane width shrank and the car moved over.


All this leads me to suspect that this new vision system isn’t just reading lane markers, but it’s also taking other cues like the presence of other cars to decide where the lane is.

All in all it’s a huge step up. I also found it able to take some ridiculously challenging on ramps and even navigated a cloverleaf and correctly slowed to 25mph for it!

2018.5 would render cars in lanes like AP1 and I'm sure the system is aware of things like parked cars. My theory on why this was not released with the NNs that were part of 2018.5 has to do with Tesla marketing. I think they will save that feature and deploy it with a new EAP feature and skip the whole AP1 parity talk and launch into the fact their product is the most superior system ever deployed. Tesla already declared AP1 parity, so now its not about rehashing that but about the future being now.

As an anecdote, I engaged AS at an intersection with a protected left and where I was behind a car. It followed that car through 80% of the turn and then stopped turning (drove straight at the curb). I was ready to take over but it shows the car could even do some intersection turning because it has done the hard part of the left turn and just freaked and stopped. Further, it is tracking cars up to around 50mph reducing Russian Roulette fears (its been 100% for me below 40mph but I've only had it for 2 days but I've also driven 340 miles, but still data is limited). So they are making leaps in almost all facets and I'm amped to see how Karpathy can "refine" the system for new features (as he said on Twitter). Karpathy doesn't seem to suffer from the same sort of grandiose proclamations that Elon does and so his credibility is strong.

 

[jimmy_d]

@jimmy_d -- Glad you got it and can see for yourself. Despite the string of constant superlatives until you experience it, its hard to describe how different it is from 2018.6.1 and before. Closest thing I can analogize to is the massive improvement we saw with .40/.42 (another Karpathy "fix" of prior code).

I have a lot less experience with AP1 than you but I also feel glad I got AP2 for the first time since I took delivery way back in Dec 2016. It is so much better but I also see how they can make it better and better and that has me even more excited.

I'm looking forward to hearing more as the NNs are refined and the repeater cams are activated with the current NN.

So true. You really have to drive it to feel the difference. I had pretty high expectations from seeing the videos and even so I was surprised.

 

[SucreTease]

Wow. It was perfect. It worked in places that I've seen fail 100% on both AP1 and previous AP2....

I'm just flabbergasted. It's not even a contest now - for me this thing just crushes AP1.

No, no, no, no......that's....not....possible..... Take that back you blasphemer! Thou shalt have no other gods before AP1!

 

[Reciprocity]

Summoning @buttershrimp, our resident @Balderskb exterminator.

Uhm, since when did I get fired. No one told me. I'll let @buttershrimp take this one as he is often way more entertaining.

 

[Reciprocity]

Whats most surprising to me is just how different it is. Three things consistently happened, like 100% of the time. Diving into turn outs in local roads, ping ponging on wide lanes and darting into other lanes or off the road when cresting even small hills. All of things are gone over night. It's night and day. Many if these issues plagued both ap1 and 2 with one being a bit better then other from situation to situation. This is better then both at everything I have tried. We have all seen the videos. Amazing stuff. Gives me a lot more confidence in the future. Mostly because it seems from Andrej Karpathy's tweet, that it was almost a complete rewrite.

Things should start to accelerate. Rumor is that an early access version is already seeing stop signs. This has awesome implications for safety. The car can now warn and even stop if you are about to blow through a stop sign at high speed.

 

[bhakan]

I drove on the hwy. Yes. the auto lane changes are very smooth and not jerky. I could not test the hwy entrance ramp as it was a slow due to lot of cars ahead of me and a speed trap prior to that entrance. I have to test more tomorrow when it is day time as I did not want to be out late on St Patty's day. However, my exit did not go well. My exit suddenly starts with broken line after the curve and becomes two lanes. So the car was in the middle and then started to go straight (could have gone to grass and shrubs ) and not exit. So I have to quickly intervene and take the exit. So lot of nice improvements. Still has to learn the exit well.
Also- I noticed that the car does not know the traffic rules that we learn at driver's ed, like, not to cross the solid double line. it ignores and does auto lane change when the turn indicator is on on the hwy.

 

[buttershrimp]

Uhm, since when did I get fired. No one told me. I'll let @buttershrimp take this one as he is often way more entertaining.

Man, it definitely beats AP1 in several ways now. Only place I see it behind now is speed limit sign recognition , but AP2 is so good now I prefer it over my experience last weekend, I’m really glad I didn’t spend time editing the comparison video because I’m too busy sorting through the fact that AP 2 navigated 2222 from mopac to 360 without one disengagement!!!! That’s right, the high water mark has been achieved !!!! I’ll try to put something up tonight .

Also I need to ask if Blader needs my referral code.... I mean, supercharging for life? Like how many times do I have to be right?

 

[Bladerskb]

I would have thought that being able to make a good decision to start moving would require the system monitoring the environment (level 3). Differentiated from current systems which require the driver to press the accelerator to resume adaptive cruise (level 2) after being stopped.

A level 3 system monitors the environment so the driver doesn't have to. they can read a book, watch movie instead of babysitting the system. that's the difference here.

A level 2 system can monitor the environment and perform all necessary driver tasks, but the difference again is, does it require the driver to monitor the environment and babysit the system aswell?

If no, then its a level 3 system and would prompt the driver with sufficient advanced warning when it needs a take over.
If yes, then its a level 2 system and it will hand over control to the driver IMMEDIATELY without advanced warning.

EAP according to Tesla is a level 2 system.