Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Elon: "Already testing traffic lights, stop signs & roundabouts"

This site may earn commission on affiliate links.
Isn’t Tesla claiming they have a generic solution that does not rely on detailed maps?

I think it means to talk about drivable space such as when the snow covers up all the lane markings.

High Definition Maps dependent companies don't need to see the lane markings at that time but still can place the car to the correct lane thanks to its database reference.

Tesla can just use live cameras to figure out where are drivable space rather than HD map.

However, Tesla does use some kind of generic maps such as Global Navigation Satellite Systems (GPS) to figure out when the next exit / freeway interchange would be.

High Definition Maps dependent system like current GM Cadillac Cruise Automation is turned off when construction zone is detected because it cannot risk placing the car in an obsolete "correct" lane according to the database but that lane might be not correct anymore during construction. That system would frustrate me very much because I spend lots of time in construction zones in Bakersfield and Los Angeles metropolitan.

In the meantime, Tesla drivable space system works very well in the construction zone as I was driving southward on a Northbound I-5N lane near Six Flags Magic Mountain.

You can see the road sign is facing toward the correct opposite Northbound traffic. The lane I drove in the picture was physically Northbound but it was temporarily allowing southbound traffic during construction:

 
Can anyone please tell, how the Model 3 with the latest Software handles a roundabout, if you’re steering towards one?
If there are generous lane markings through the roundabout, and it is only a gentle curve large roundabout, there is a small possibility it will make it straight through the roundabout unhindered... But the vast majority of the time it will likely drive you up onto the roundabout or into a curb or straight into a signpost or tree. In other words, it can't remotely handle roundabouts. The currently released autopilot features are only driver assist autosteer and lane change features designed for highways. Do not assume anything beyond that.
 
You are a much braver man than I using autosteer that close to a jersey barrier.

One good thing about Bakersfield and Los Angeles Metropolitan construction is: Their lane markers don't lead Autopilot slamming into a median concrete SO FAR!

In future, if they forgot to update the lane marking away from a median concrete, I might be in trouble too because Autopilot can definitely follow a misdirected lane marker to slam into a median concrete.
 
One good thing about Bakersfield and Los Angeles Metropolitan construction is: Their lane markers don't lead Autopilot slamming into a median concrete SO FAR!

In future, if they forgot to update the lane marking away from a median concrete, I might be in trouble too because Autopilot can definitely follow a misdirected lane marker to slam into a median concrete.
As long as the autopilot vision can recognize a non-driveover obstacle and drivable space, it's pretty easy software 1.0 problem to see that driving the marked path will lead to a crash, and pick an alternate path, eventually reduce speed.

I suspect reason they have not activated that code yet is that they don't fully trust their vision NNs yet. So enabling such code would lead to more phantom braking and add an unpredictable swerving problem (shadows interpreted as objects mm... dangerous stuff).
 
Correction. Los Angeles did guide their yellow line into the concrete median but I was lucky that Autopilot didn't swipe it although if I lowered my window down, I could have touched the concrete barrier because it's just too close:

Autopilot is VERY bad at concrete barriers near the road. I would guess that at least 95% of my manual interventions involve my Model X getting too close to the concrete center barriers or edge barriers on CA-17. I really don't understand why Tesla still hasn't gotten something this conceptually simple even remotely close to correct. After all, apart from training the neural net to identify concrete barriers, it's just a matter of implementing three rules:
  1. If there is a car or concrete barrier to the left of you, stay towards the right edge of your lane.
  2. If there is a car or concrete barrier to the right of you, stay towards the left edge of your lane.
  3. If there is a car or concrete barrier on both sides, slow down to the speed of the vehicle to avoid passing, and hug the vehicle side, so that if Autosteer makes a mistake, you'll end up behind the danger instead of crashing into the side of it.
That's it. o_O
 
  • Like
Reactions: arnolddeleon
...implementing three rules...

It's easy and quick for human programmers to write the code for a specific route, specific situation, specific location...

If I understand it correctly, Tesla doesn't prefer human programmers to write the autonomous driving algorithm. It prefers to supply the machine the enormous data and let the NeuroNet figures it out and does its own code.

That's why Elon said that he could "game" it by writing the code to drive autonomously from LAX to NYC but that won't work for other routes because someone will have to write new codes for new routes.
 
Last edited:
I think it means to talk about drivable space such as when the snow covers up all the lane markings.

High Definition Maps dependent companies don't need to see the lane markings at that time but still can place the car to the correct lane thanks to its database reference.

Tesla can just use live cameras to figure out where are drivable space rather than HD map.

However, Tesla does use some kind of generic maps such as Global Navigation Satellite Systems (GPS) to figure out when the next exit / freeway interchange would be.

High Definition Maps dependent system like current GM Cadillac Cruise Automation is turned off when construction zone is detected because it cannot risk placing the car in an obsolete "correct" lane according to the database but that lane might be not correct anymore during construction. That system would frustrate me very much because I spend lots of time in construction zones in Bakersfield and Los Angeles metropolitan.

In the meantime, Tesla drivable space system works very well in the construction zone as I was driving southward on a Northbound I-5N lane near Six Flags Magic Mountain.

You can see the road sign is facing toward the correct opposite Northbound traffic. The lane I drove in the picture was physically Northbound but it was temporarily allowing southbound traffic during construction:

The problem with this assessment is that everyone, absolutely everyone, is doing visual drivable space and were doing it before Tesla. Just Google semantic free space. The decision to switch of features based on map location is thus — compared to Tesla — simply one of safety and differing feature priorities, not one of technical features.
 
It's easy and quick for human programmers to write the code for a specific route, specific situation, specific location...

If I understand it correctly, Tesla doesn't prefer human programmers to write the autonomous driving algorithm. It prefers to supply the machine the enormous data and let the NeuroNet figures it out and does its own code.

The thing is, though, we have seen no evidence that Tesla is using neural nets to drive the car. The neural nets we’ve seen are visual networks just like everyone else is doing. In fact we know of more competitor’s driving NNs than we know of Tesla’s who has been using ”software 1.0” algorithms for Autopilot’s actual driving.

That's why Elon said that he could "game" it by writing the code to drive autonomously from LAX to NYC but that won't work for other routes because someone will have to write new codes for new routes.

There were some claimed leaks saying — these were before Elon came out with that comment — that Tesla had been trying to game the coast to coast for months but in the end the system was too brittle and they gave up. The claimed leaks said they had been re-painting Supercharger lane markings etc to try and make it work. I personally think Elon’s comment thus was more a Freudian slip.

But mostly my point is, we have so far seen no evidence that Tesla actually is using deep learning on the actual driving question — only on the visual part and are actually behind many other players in using NNs for visual detection too. So it comes back down to wishful thinking more than any solid current analysis of what Tesla is doing.

Then again, isn’t that the whole Autopilot 2 story right there. :)
 
  • Like
Reactions: Richt
If there are generous lane markings through the roundabout, and it is only a gentle curve large roundabout, there is a small possibility it will make it straight through the roundabout unhindered... But the vast majority of the time it will likely drive you up onto the roundabout or into a curb or straight into a signpost or tree. In other words, it can't remotely handle roundabouts. The currently released autopilot features are only driver assist autosteer and lane change features designed for highways. Do not assume anything beyond that.

Thank you for the detailed answer.

I already drove AP1 for 4 years and with that hardware a roundabout was suicide.

But since i'm happy to get a Model 3 next week, I wanted to know how the progress is on that topic, but it seems it is not really better.

Where I live, they changed all traffic light crossroads to roundabouts. So this really matters, since Tesla talks about automatic driving on city streets later this year (text form the Tesla Model 3 Order Page):

"Coming later this year: Recognize and respond to traffic lights and stop signs. Automatic driving on city streets."

If they can't handle roundabouts till then, they better implement a safety net, like: A loud warning to take over and stopping the car before entering a roundabout, if no response from the driver!

The actual behavior, even what you described, would be very, very dangerous.

I'll hope for the best, but since they don't talk about roundabouts at Tesla, it probably is not on their radar.
 
Last edited: