Wiki Tracking FSD Feature Complete

[EVNow]

Oh, that won't work. They lay it in one part around here....

Right. They need a generic solution for unmarked roads - by that I mean with no center line. When there are side lines marking shoulders but no center lane, like a lot of residential streets near me, the car just drives in the middle.

 

[EVNow]

Just to be sure we're not talking different things, its stationary cars that were stationary before they came into view and not cars that have stopped while in view. I rarely come across that situation and I'd be surprised if many people did. I certainly don't trust it to act on the maybe 1 or 2 occasions over a couple of years when it has happened as for me its usually when travelling on fast roads and coming up behind an incident (like the fire truck scenarios where the cars have crashed).

I think it is a hard problem to stop if there is a stopped car in front that suddenly comes into view (like when the car infront of you changes lane). I think that is what the TACC is referring to. AP needs to swerve to avoid the stopped car.

In general I think they need to implement the 3D world they showed in the autonomy presentation to figure out the stopped car. Now that they know hackers will get to any hidden code in production, I'm afraid hey are going to not merge the code in to production branch even if ready, if they don't want that feature out. So, we'll have to wait to see what they actually have.

 

[jerry33]

I think it is a hard problem to stop if there is a stopped car in front that suddenly comes into view (like when the car infront of you changes lane). I think that is what the TACC is referring to. AP needs to swerve to avoid the stopped car.

If possible. It needs to be aware of what's around because just swerving might have bad consequences.

 

[RDoc]

If possible. It needs to be aware of what's around because just swerving might have bad consequences.

This.
IMHO, the car really needs to build a model of the world around it that includes identifying objects and tagging them with characteristics, capabilities and likely behavior. In the above scenario, if the car suddenly sees a stopped car, it needs to be aware of what's happening around it, for example, another car coming up rapidly in the next lane over.

 

[EVNow]

If possible. It needs to be aware of what's around because just swerving might have bad consequences.

Oh yes. BTW, Tesla already apparently swerves if someone cuts in and is about to hit you. Just saying they need to do the same thing.

 

[GeorgeSymonds]

Oh yes. BTW, Tesla already apparently swerves if someone cuts in and is about to hit you. Just saying they need to do the same thing.

It does that with AP1 and I don't think its unique to Tesla - probably a mobileye feature - if the parking sensors detect something entering your space it moves the car but not by much - although like all thesed things there is a difference from having a feature and the feature being reliable

Side Collision Avoidance - disappointing?

 

[EVNow]

I've updated the first post with the feature set that Tesla lists on the website officially as being part of FSD. The corresponding feature numbers are noted in brackets).

As of 7/12/19 this is what Tesla website lists as part of FSD.

Autopilot Included

• Enables your car to steer, accelerate and brake automatically for other vehicles and pedestrians within its lane. (2.1, 2.3)

Full Self-Driving Capability

• Navigate on Autopilot: automatic driving from highway on-ramp to off-ramp including interchanges and overtaking slower cars. (9)
• Auto Lane Change: automatic lane changes while driving on the highway. (3)
• Autopark: both parallel and perpendicular spaces. (10.1, 10.2)
• Summon: your parked car will come find you anywhere in a parking lot. Really. (1)

Coming later this year:

• Recognize and respond to traffic lights and stop signs. (5, 4.1)
• Automatic driving on city streets. (2, 3, 6, 7) <-- This is the ambiguous one, which could mean just a few basic features or full set of features.

 

[EVNow]

I'd really like to know more about the stop light and stop sign software, because I don't think I'd trust it for a long time. Visual recognition plus a map of every stop sign and stop light?
That will be a really big step forward. If someone gets killed, that would put a damper on sales, as well as the brand and whole idea of self driving. I'm thinking if stop light and sign recognition comes out later, it would start as a back up to humans, or at least insist that the driver stay engaged as an emergency back up - will be interesting to see how they roll this out.

I don't expect this to be any different than all other EAP/FSD features. Tesla will insist on driver paying attention and keeping hands on the wheel. Personally I use EAP/FSD carefully monitoring to make sure its doing what its supposed to be doing and intervene if it doesn't.

For eg., once stop sign recognition is live, you expect the car to start slowing down when it sees stop sign. I expect them to visualize the stop sign as well. If it doesn't slow down, I'll intervene. Same for red traffic signals. Turns will probably be indicated before the turn happens and you expect the car to slow down as well - so we can monitor and if the turn doesn't happen, can intervene.​

 

[tmoz]

I don't expect this to be any different than all other EAP/FSD features. Tesla will insist on driver paying attention and keeping hands on the wheel. Personally I use EAP/FSD carefully monitoring to make sure its doing what its supposed to be doing and intervene if it doesn't.

For eg., once stop sign recognition is live, you expect the car to start slowing down when it sees stop sign. I expect them to visualize the stop sign as well. If it doesn't slow down, I'll intervene. Same for red traffic signals. Turns will probably be indicated before the turn happens and you expect the car to slow down as well - so we can monitor and if the turn doesn't happen, can intervene.​

Good point - one needs to know that the car sees the sign or light well in advance of taking action. Just like it is reassuring to see, on your screen, the cars that AP sees.

 

[jerry33]

I don't expect this to be any different than all other EAP/FSD features. Tesla will insist on driver paying attention and keeping hands on the wheel. Personally I use EAP/FSD carefully monitoring to make sure its doing what its supposed to be doing and intervene if it doesn't.

For eg., once stop sign recognition is live, you expect the car to start slowing down when it sees stop sign. I expect them to visualize the stop sign as well. If it doesn't slow down, I'll intervene. Same for red traffic signals. Turns will probably be indicated before the turn happens and you expect the car to slow down as well - so we can monitor and if the turn doesn't happen, can intervene.​

I hope it uses mostly regeneration the way a human would rather than mostly friction brakes the way current AP is said to do.

 

[EVNow]

I hope it uses mostly regeneration the way a human would rather than mostly friction brakes the way current AP is said to do.

Good point.

 

[tmoz]

Good point.

I suspected it did that but wasn't 100% sure, so I nearly always take it off AP (first gen) when coming to a stopped set of cars.
Perhaps AP1 can't see the cars far enough in advance to smoothly use regen in time.

 

[powertoold]

Good progress here:

 

[EVNow]

Good progress here:

Yes - but its not even clear under what condition this happened !

But I'm fairly sure they have figured out how to make turns - like we saw in the demo.

 

[EVNow]

EM just tweeted today/yesterday - they are targeting release of V10 in August/Sept. That includes stop sign / traffic light recognition (I think they will at least start showing the visualization, if not stopping the vehicle).
​

@elonmusk Replying to @DMC_Ryan

Yes, V10 will include several games & infotainment features, improved highway Autopilot, better traffic light & stop sign recognition & Smart Summon​

​

@elonmusk Replying to @redmor11 @schristakos and 2 others

Will recognize & read all signs for FSD. Focused on traffic lights & stop signs right now. Most are very easy. Difficulty is millions of corner cases.​

 

[tmoz]

EM just tweeted today/yesterday - they are targeting release of V10 in August/Sept. That includes stop sign / traffic light recognition (I think they will at least start showing the visualization, if not stopping the vehicle).
​

@elonmusk Replying to @DMC_Ryan

Yes, V10 will include several games & infotainment features, improved highway Autopilot, better traffic light & stop sign recognition & Smart Summon​

​

@elonmusk Replying to @redmor11 @schristakos and 2 others

Will recognize & read all signs for FSD. Focused on traffic lights & stop signs right now. Most are very easy. Difficulty is millions of corner cases.​

difficulty in millions of corner cases is what AI and machine learning was all about, wasn't it?

 

[EVNow]

difficulty in millions of corner cases is what AI and machine learning was all about, wasn't it?

NN allows "millions of corner cases" to be considered - but to handle millions of corner cases, you need billions of items of training data - and that takes years

Anyway, "millions of corner cases" is Musk's frustration coming through.

Did Musk really believe FSD is possible by next year and is now finding out there are millions of corner cases, so it can't actually be solved in a year ?

 

[Slarty1080]

It’s interesting to see how all the FSD features are developing and I’m planning to have FSD on my model 3, but I fear that it the UK at least there will be extreme regulatory resistance to driverless cars for a long time to come. Despite often claiming to use evidence based policy the UK Government does nothing of the sort. It is often driven by expediency, public option and newspaper headlines as could be seen a few years back when the Government's chief scientific adviser on drugs said the taking ecstasy was as dangerous as horse riding. A true statement but was an uncomfortable fact for an evidence based policy so they fired him.

There will also be opposition from groups like taxi drivers who will be made redundant. An example of the trouble that can be caused can be seen in trains. How much easier would it be to have an automated train with or without a driver? Are we anywhere near that in the UK? No. In fact we have massive industrial disputes about guards on trains and whether it’s safe to have a train with just a driver on board. According to the unions it’s because having a guard on board makes the train safer. The real truth is the guards don’t want to lose their jobs.

 

[ReflexFunds]

NN allows "millions of corner cases" to be considered - but to handle millions of corner cases, you need billions of items of training data - and that takes years

Anyway, "millions of corner cases" is Musk's frustration coming through.

Did Musk really believe FSD is possible by next year and is now finding out there are millions of corner cases, so it can't actually be solved in a year ?

The millions of corner cases is exactly why Tesla chose its data heavy, hardware light approach to self driving years ago. Elon is just saying the corner cases is the most difficult part, which it is (however corner cases including predicting actions of other road users is much more difficult than corner cases for signs). He obviously isn't going to be surprised by the corner cases as it was the basis of his entire strategy.

Tesla AP2+ vehicles should now be driving over 0.5 billion miles per month. As Tesla start focussing on the rarer and rarer edge cases they just need to scale up their fleet data collection rate and scale up their data annotation partners in India.

 

[tmoz]

NN allows "millions of corner cases" to be considered - but to handle millions of corner cases, you need billions of items of training data - and that takes years

Anyway, "millions of corner cases" is Musk's frustration coming through.

Did Musk really believe FSD is possible by next year and is now finding out there are millions of corner cases, so it can't actually be solved in a year ?

Those billions of miles driven were supposed to be the sufficient "training data set", or so we all thought.