Blog Musk Touts ‘Quantum Leap” in Full Self-Driving Performance

A “quantum leap” improvement is coming to Tesla’s Autopilot software in six to 10 weeks, Chief Executive Elon Musk said a tweet.

Musk called the new software a “fundamental architectural rewrite, not an incremental tweak.”

https://twitter.com/x/status/1294374864657162240

Musk said his personal car is running a “bleeding edge alpha build” of the software, which he also mentioned during Tesla’s Q2 earnings. 

“So it’s almost getting to the point where I can go from my house to work with no interventions, despite going through construction and widely varying situations,” Musk said on the earnings call. “So this is why I am very confident about full self-driving functionality being complete by the end of this year, is because I’m literally driving it.”

Tesla’s Full Self-Driving software has been slow to roll out against the company’s promises. Musk previously said a Tesla would drive from Los Angeles to New York using the Full Self Driving feature by the end of 2019. The company didn’t meet that goal. So, it will be interesting to see the state of Autopilot at the end of 2020.

 

[Daniel in SD]

Please do tell me what Tesla has and hasn't started working on! With some actuall facts / sources to back it up.

Elon said on July 23rd that the car can almost do his commute. You are correct that I have no idea how much progress they've made since then!
There has been talk about how good the "secret builds" are at Tesla since 2016 so you'll have to excuse my skepticism.

 

[Daniel in SD]

I agree it's useless as a means to accomplish robotaxis - quite a different problem scope. But I'd say extremely good supervised self-driving is quite useful for long trips and has substantial market value - unless of course someone else has a bunch of robotaxis or a better system at the same price. I'd like my car to be much better at supervised self-driving!

Well I think supervised self-driving on city streets is unsafe without very good driver monitoring (see the Uber self-driving pedestrian fatality).

 

[Snuffysasa]

A single percentage is easy to communicate but it hides a lot of details especially when considering Level 4 vs Level 5. For example, that 90% number could have various meanings:

• 90% of worldwide roundabouts are successful 100% of the time (and remaining 10% of roundabouts always fail)
  
• 95% of worldwide roundabouts are successful 95% of the time
• 100% of worldwide roundabouts are successful 90% of the time

In the context of Level 4, the first item is desired as one would geofence deployment to the regions that have 100% success rate. The types of failures could be static, e.g., is the shape of roundabout and lane entrances/exits understood, or dynamic, e.g., other vehicles moving at certain speeds/angles or lighting conditions. (And separately, there's always the correctness vs natural/smooth-ness, e.g., yielding too long.)

For a California Tesla owner, having Autopilot successfully navigate the few roundabouts (if any) that one might encounter on a daily basis is practically full self driving in that aspect for that person's commute. And similarly for all other aspects of full self driving, a given person's commute requiring no intervention is effectively Level 4 for many of the miles the vehicle would see.

As you point out, successfully handling some number of commutes without intervention is far from the many potential routes that Level 5 would need to handle for "True self driving." But Tesla's approach allows them to release things as Level 2 and actively measure their self-driving progress within a commute, city, region, country, worldwide. And ideally other people's commutes making progress towards no intervention makes it more likely to be successful when traveling not-your-usual routes.

I think we are in agreement here. And I agree with what you say, it is far more complex, and yes I was simplifying it.

"For a California Tesla owner, having Autopilot successfully navigate the few roundabouts (if any) that one might encounter on a daily basis is practically full self driving in that aspect for that person's commute"

No, that is practically fully assisted driving from that persons perspective. That is what ADAS is. Even if a person goes to and from every work every single day without touching the wheel for a few years. That still doesn't make it self driving...

Now if that person could take a nap as the Tesla goes through roundabouts in California... it would need to be 99.99999% successful at roundabouts in that area... and that would be fully self driving.

 

[AlanSubie4Life]

Well I think supervised self-driving on city streets is unsafe without very good driver monitoring (see the Uber self-driving pedestrian fatality).

I agree with that. But I'm just saying to get to the point where that particular safety issue is a big issue, there would have to be a lot of progress from where they are at currently, and it would be impressive in its own right. At that point, that specific safety issue with supervised self-driving would definitely become a concern as the capability is rolled out to the fleet.

And yes, to be clear, people misusing the current system with its very limited capabilities also presents a safety hazard. I just think it's likely to be more and more hazardous the more capable the system gets. How the overall safety numbers will work out, I don't know.

 

[Snuffysasa]

Well I think supervised self-driving on city streets is unsafe without very good driver monitoring (see the Uber self-driving pedestrian fatality).

Agreed, a driver monitoring system should be added. However, even with the Uber Accident companies testing SDC on city streets, over the last 10 years, have proved a better mile per accident rate than normal driving in those areas.

 

[Daniel in SD]

Agreed, a driver monitoring system should be added. However, even with the Uber Accident companies testing SDC on city streets, over the last 10 years, have proved a better mile per accident rate than normal driving in those areas.

Uber added driver monitoring! Also, safety drivers probably get fired if they don't monitor the vehicle which also helps.

 

[Cheburashka]

Yeah, Tesla demos are BS. I don't remember last time Tesla did a big demo and it resulted in something useful for us owners.

You know if it's a "demo" it probably won't get released any time soon.

Full self-driving in 2016.

Then car parking itself in the parking lot.

Autonomy day.

etc

 

[30seconds]

Has anyone noticed a substantial improvement in Autopilot over last 18-24 months? To me there seems to be a bit better stability, but that's about it.

 

[Daniel in SD]

Has anyone noticed a substantial improvement in Autopilot over last 18-24 months? To me there seems to be a bit better stability, but that's about it.

It's way smoother for cut-ins and when you have driver who uses the accelerator as an on-off switch in front of you. Those were my main complaints so I'm pretty happy. Lane changing is also much more refined with fewer aborts.

 

[powertoold]

AP accelerates and decelerates much smoother, especially in cut-in situations as you say. I also enjoy the gradual and human-like stopping at traffic controls. From time to time, I still get a "pulling" sensation from the lead car accelerating. Other than that, APs been great (I use it at least once in 95% of my drives) and getting smoother every update for me.

 

[Snuffysasa]

Has anyone noticed a substantial improvement in Autopilot over last 18-24 months? To me there seems to be a bit better stability, but that's about it.

It's way smoother for cut-ins and when you have driver who uses the accelerator as an on-off switch in front of you. Those were my main complaints so I'm pretty happy. Lane changing is also much more refined with fewer aborts.

Ohh yea, and NoA is way more intelligent about when/how/where to make lane changes and cut in, pretty amazing actually how much better it has got. Traffic light control has already improved immensely from when it launched a few months ago. Smart summon has also improved quite a bit from initial launch about year ago. And obviously, a variety of new features have been added, so that's improvements.

 

[Mardak]

Has anyone noticed a substantial improvement in Autopilot over last 18-24 months?

If you're going 2 years back, Navigate on Autopilot was introduced October 2018 Introducing Navigate on Autopilot And 1.5 years back Tesla introduced "no confirmation" NoA April 2019 Introducing a More Seamless Navigate on Autopilot

Excluding those, looking at the Tesla blog, there is indeed no substantial improvement to highway or city street driving. There's no blog article about stopping at stop signs or traffic lights although there was a mention of Smart Summon in the v10 article.

So perhaps not blog-worthy, personally I've noticed more confident lane changes and predictions for vehicles and pedestrians. Although for the latter, sometimes the predictions are too cautious, e.g., warnings when another vehicle turns in preparation to cross the road after you pass or when pedestrians doesn't actually intend to cross the road.

 

[Cheburashka]

It's better for sure but it's not great.

The biggest issue is phantom breaking which has gotten worse. My wife now refuses to use AP.

 

[powertoold]

The biggest issue is phantom breaking which has gotten worse. My wife now refuses to use AP.

I haven't experienced any phantom braking while following a car. On the highway, I simply get behind a car, set the speed 10-15mph above the lead car, and let AP do its thing, while I only pay attention to road debris.

 

[Cheburashka]

I haven't experienced any phantom braking while following a car. On the highway, I simply get behind a car, set the speed 10-15mph above the lead car, and let AP do its thing, while I only pay attention to road debris.

We don't drive that much in heavy traffic especially as of late.

It hasn't been good for us at all, even with the latest updates. We live by a fairly curvy highway and the car slams the brakes randomly quite often. I'm not expecting too much at this point since Tesla can't even get the rear camera to work in the dark on MCU2.

 

[powertoold]

We don't drive that much in heavy traffic especially as of late.

It hasn't been good for us at all, even with the latest updates. We live by a fairly curvy highway and the car slams the brakes randomly quite often. I'm not expecting too much at this point since Tesla can't even get the rear camera to work in the dark on MCU2.

I don't know your particular situation, but I wasn't referring to heavy traffic. Whenever I use AP on a highway, I use it with a lead car (or when there's no car behind me).

 

[diplomat33]

2020.40.1 continues to incrementally remove driver intervention for traffic light control. Now, the car can continue through a green light in some cases without a lead car:

"As you approach a green traffic light, in some situations where there is a straight path through the intersection, the Traffic Light and Stop Sign Control feature may no longer require explicit driver confirmation. If this occurs, the stop line in the driving visualization will turn green to indicate that the car will continue through an intersection. Confirmation is still required if your car is already stopped at the light when it turns green.”

Tesla pushes new software update to enable cars to automatically drive through green traffic lights - Electrek

 

[powertoold]

This is huge. Every little update like this is a big deal. Essentially, Tesla is telling us that the car can confidently figure out which lane it's in, see the lights, and know which light applies to it.

Confidently figuring out which lane it's in (right or left turn only or straight, etc.) is crazy challenging IMO (even as humans sometimes).

 

[Bladerskb]

Man, you guys are hilarious. I can't find a single video or piece of evidence that these BMW traffic control systems have been deployed to consumers. All the videos or information I see is marketing and/or marketing demos.

I'd help a lot if you guys can provide videos of actual consumers showing off these features.

Once once once again again again, UNDER DEVELOPMENT:

BMW's next-gen adaptive cruise control stops at traffic lights - Roadshow

Get off your cool aid

New BMW Urban Cruise Control recognizes traffic lights

Introduction of the new Urban Cruise Control system which is an evolution of the previous technology. The new system recognizes traffic lights around town when is turned on and will be available in numerous BMW models starting this fall. The new function basically improves on the current Active Cruise Control system with Stop & Go function by recognizing traffic lights.

The Urban Cruise Control function can also be integrated into the car at a later date by means of a Remote Software Upgrade. With the exception of the BMW 8 Series models, this option is available for vehicles of all the above-mentioned series produced from July 2020 onwards and which are equipped with the Driving Assistant Professional.
​

This is huge. Every little update like this is a big deal. Essentially, Tesla is telling us that the car can confidently figure out which lane it's in, see the lights, and know which light applies to it.

Confidently figuring out which lane it's in (right or left turn only or straight, etc.) is crazy challenging IMO (even as humans sometimes).

It really isn't. Neural network for relevancy of traffic light was in EyeQ4 since Q4 2017.

 

[powertoold]

Neural network for relevancy of traffic light was in EyeQ4 since Q4 2017.

And the wheel was invented in -4017.