“Drive on Nav” feature in Enhanced Autopilot v9.0

[strangecosmos]

Electrek claims (citing an unnamed source) that the feature called “Drive on Nav” — previously shown in leaked screenshots of the developer build of Enhanced Autopilot v9.0 — will suggest lane changes to the driver, requiring them to confirm by pulling the stalk before actually changing lanes.

This is far from driving onramp to offramp on highways with no driver input, but it is an incremental step in that direction. (Caveat: Electrek’s information may be wrong.)

This feature would provide labelled data to Tesla. Every time a driver pulls the stalk (or doesn’t), that’s an example of Autopilot’s suggestion being correct or incorrect. If a deep neural network is making the suggestions to change lanes, then Tesla can use the labelled data to train and improve the neural network.

It is disappointing if indeed onramp to offramp is still several updates away. But pulling in lots of human labelling from drivers (if that indeed is what will occur) will accelerate progress.

 

[Derek Kessler]

Elon said during the last conference call that it'd be part of v9:

So right now, a lot of the focus is on Autopilot version 9, which is our sort of an on-ramp to off-ramp solution that’s going to automatically attempt to change lanes, understand what lane the car is in, understand the route the user wants to travel and take that route for the user and ultimately hand back control to that user which is kind of safe and controlled.

Though it's worth noting that he never stated outright that it would be feature-complete on release. So right now it looks like step one is "understand what lane the car is in" and "understand the route the user wants to travel".

 

[strangecosmos]

That’s actually a quote from Stuart Bowers, but Elon has said similar things, including at the shareholder meeting.

It’s never been exactly clear to me if these capabilities are planned for v9.0 or, say, v9.1, but Elon said in June the release might be a “couple months” away, which implied to me v9.0.

Keeping human supervision (and human labelling for deep supervised learning) in the loop as long as possible is a good idea, so I don’t begrudge them if driver input is still required for a while yet.

 

[J1mbo]

This feature would provide labelled data to Tesla. Every time a driver pulls the stalk (or doesn’t), that’s an example of Autopilot’s suggestion being correct or incorrect. If a deep neural network is making the suggestions to change lanes, then Tesla can use the labelled data to train and improve the neural network.

It's not a very clean way of labeling data. There are plenty of reasons why the human will delay or even ignore the suggestion: traffic; poor Nav route; not paying attention...

I am also very sceptical about this report, because I cannot see any difference between this and using Nav & AP today:

Road splits ahead, nav tells me to keep left or keep right; I indicate if I need to change lanes.
Nav tells me to exit the highway ahead, I use AP to change lanes and it exits...

"Drive on Nav" must be more than this.

 

[Bladerskb]

This seems like a huge disappointment but not to me. I always expected this. The feature has been hyped to hell but it will only be a glorified lane change on during empty freeway not in dense traffic.

Saying they are using human supervision for human labeling and deep learning is quite reaching.

If that was the case, they would be gathering data today when automatic lane change is requested by pulling the turn signal.

 

[strangecosmos]

I am also very sceptical about this report, because I cannot see any difference between this and using Nav & AP today:

Road splits ahead, nav tells me to keep left or keep right; I indicate if I need to change lanes.
Nav tells me to exit the highway ahead, I use AP to change lanes and it exits...

Right now, you, the human driver, decide when a lane change is necessary. With “Drive on Nav”, supposedly Autopilot will decide when a lane change is necessary. But it will wait for you to either confirm or deny that Autopilot has made the right call.

It adds a slight amount of convenience.

It's not a very clean way of labeling data. There are plenty of reasons why the human will delay or even ignore the suggestion: traffic; poor Nav route; not paying attention...

Good point, but the data uploaded to Tesla HQ can be manually annotated afterward. You’re right to suggest it would probably not be best to just train the (hypothetical) neural network on raw correct/incorrect labels. But the correct/incorrect or accepted/rejected suggestion labels can just be the start of the labelling process.

If that was the case, they would be gathering data today when automatic lane change is requested by pulling the turn signal.

They may indeed already be doing this. I don’t know— I’m just speculating. For all I know, they aren’t using deep learning for lane changes at all! I’m just trying to think about what they might be doing based on what I know.

 

[strangecosmos]

I think going forward, it may be a recurring theme that new autonomy features require human supervision or confirmation. You can test a feature for billions of miles in simulation and millions of miles in dev cars. But nothing substitutes for tens of millions or hundreds of millions of miles of testing, which is only possible in customer cars.

You need something like 65 million miles of testing to get good statistical evidence that a feature improves safety. So for something like Autopilot actuating lane changes autonomously, you would want customers to use the feature in suggestion-only mode for 65 million miles of highway driving first. If it never suggests an unsafe lane change in 65 million miles, then maybe you can think about letting Autopilot change lanes autonomously.

This is less glamorous and magical than the car doing things by itself right out of the box, but it seems like the best way. The problem with the approach of Waymo, Cruise, and others is that the system might be significantly more dangerous than a human driver at launch. Waymo is only at 9 million miles total. It will have to scale up testing 5.5x to get to a rate of 65 million miles per year, from a run rate of around 12 million miles per year currently.

 

[Tam]

...will suggest lane changes to the driver, requiring them to confirm by pulling the stalk before actually changing lanes...

There's no going around it because current Auto-Lane-Change is unable to detect cars far enough from behind to safely change lane on its own.

To remedy this current driver-dependent Auto-Lane-Change, cameras need to be turned on to see cars from behind and software to recognize speeding cars from behind and react accordingly.

This additional camera activation is not for cheaper EAP and only for more expensive FSD.

 

[J1mbo]

Right now, you, the human driver, decide when a lane change is necessary. With “Drive on Nav”, supposedly Autopilot will decide when a lane change is necessary. But it will wait for you to either confirm or deny that Autopilot has made the right call.

So instead of Nav telling me to change lanes, AP does?

Sorry, this is purely semantics.

 

[strangecosmos]

So instead of Nav telling me to change lanes, AP does?

GPS nav doesn’t tell you the exact moment to change lanes, right? Autopilot will try to.

 

[SSonnentag]

There's no going around it because current Auto-Lane-Change is unable to detect cars far enough from behind to safely change lane on its own.

To remedy this current driver-dependent Auto-Lane-Change, cameras need to be turned on to see cars from behind and software to recognize speeding cars from behind and react accordingly.

This additional camera activation is not for cheaper EAP and only for more expensive FSD.

Perhaps I'm misunderstanding, but the current lane-change will wait for an opening before executing the lane change. Whether it is using cameras or other sensors is unclear to me, but it certainly can "see" cars to the side and rear and will wait for an opening before moving over.

 

[J1mbo]

GPS nav doesn’t tell you the exact moment to change lanes, right? Autopilot will try to.

Yes it does. Nav tells you in advance when the next turn is. You can choose to trigger auto lane change at exactly that moment.

 

[animorph]

There's no going around it because current Auto-Lane-Change is unable to detect cars far enough from behind to safely change lane on its own.

To remedy this current driver-dependent Auto-Lane-Change, cameras need to be turned on to see cars from behind and software to recognize speeding cars from behind and react accordingly.

This additional camera activation is not for cheaper EAP and only for more expensive FSD.

EAP includes the two side repeater cameras, which are rear-facing and perfect for watching the blind spot and farther back. At least that's how they originally described it.

 

[Tam]

...will wait for an opening...

If you can follow the youtube below. There are 3 recordings:

1) Upper right is the instrument cluster showing autopilot was on and when Auto-lane-change was activated.

2) Lower right is rear view where you could see a very slowly approaching car from the rear that Auto-lane-change ignored it and still performed an unsafe a lane change but was honked by that car and I had to manually hold the steering wheel back to prevent it from turning into the left lane and thus, automatically canceled the Autopilot.

3) Front view on the left side of the screen. I waited for the car in front and side to clear BUT not cars from behind.

 

[Tam]

EAP includes the two side repeater cameras, which are rear-facing and perfect for watching the blind spot and farther back. At least that's how they originally described it.

I was confused. Thanks for the correction. Yes EAP does include 2 rearward looking side cameras.

 

[arathald]

Yes it does. Nav tells you in advance when the next turn is. You can choose to trigger auto lane change at exactly that moment.

It tells you which lanes are valid options for following the route. It doesn't specifically suggest one of those lanes or suggest to move to a faster lane to pass when there's no upcoming turn/exit. These are completely different features.

 

[d21mike]

I have been playing with AP on trips with Navigation Active. I know based on Navigation that there is a freeway change up ahead. But I have to manually get the the correct lane to make the change. I am not sure Navigation will always tell me exactly which lane to be in early enough. Sometimes I think I am in the correct lane but I am not and have to get over 1 more either right or left depending on the freeway. And if there is a lot of traffic and I do not make the change early enough (when there is an opening) I will simply take over because I can not make it happen just be using auto lane change. This is really a GAME for me. If the New Feature makes this GAME more fun then great. No way is my car able to make all of the auto lane change accurately at this point. Maybe when the rear facing camera and detect cars coming up "fast" like say +20mph then no way can it safely make the auto lane change without my input.

 

[Alketi]

This is relatively straightforward I think. Being prompted to initiate the lane-change is a logical first step.

1. It allows for a first-step integration with maps.
2. It provides feedback on navigation prediction/accuracy.
3. It can be implemented without rear-facing camera functionality and object identification.
4. It absolves Tesla of full responsibility.

 

[Darmie]

Fascinating stuff. Now when will we see V9?

 

[strangecosmos]

Fascinating stuff. Now when will we see V9?

Towards the end of the month, according to Elon. But the early access release is already late.