Tesla's autonomy development: lumpy and fast or slow and smooth?

Elon Musk says that by the end of 2020 all Teslas produced since May 2019 will be capable of full self-driving with no one in the driver’s seat (and all Teslas produced since November 2016 will be eligible for a computer upgrade that will put them on par with newer cars). Many people have the intuition that if we were only about a year from full autonomy, the capabilities demonstrated in Teslas today would have to be close to full autonomy already. They therefore reject Elon’s timeline as ridiculous on its face. But this is not necessarily true.

Progress on autonomy could just as easily be lumpy as smooth. Driving capabilities could develop or improve as the culmination of a long process of research and development, capped off by a relatively quick phase of neural network training.

Elon’s timeline certainly may prove to be too optimistic, but I think the intuition that progress will be smooth and slow rather than lumpy and fast should be contemplated more deeply.

My article that elaborates on this idea:

https://seekingalpha.com/article/4295418-tesla-autonomy-possibility-fast-progress

I haven’t seen any examples of Tesla development being fast and lumpy in the past. Seems to be slow and lumpy.

As an example they have 600,000 cars to learn from. But their supposed cut-in detector works maybe 5% of the time before the vehicle is less than a foot already into your lane. That doesn’t instill me with confidence that if they can’t yet figure out how to leverage their data to reliably figure out that somebody with their blinker on moving toward your lane is about to come in…

1 Like

When asked, “Any word on when Navigate on Autopilot for street level (aka read traffic lights and signs) will be out?”, Elon responds:

At the Tesla shareholder meeting in June, Elon reiterated the timeline of Full Self-Driving being feature complete by the end of 2019. He also said:

When I drive the latest development version of the software, it can take me from my house to the office, but there are interventions at times. It doesn’t perfectly deal with every intersection or every turn.

So, as of June, it sounds like feature complete was, uh, feature complete, but not ready to release to the general public.

The ripeness of the software would be (at least in large part) measurable by the rate of disengagements.

Smart Summon was in Early Access testing from early June until it was released in late September, so that’s 4 months of Early Access. If Full Self-Driving is going to be in wide release by the end of the year, then shouldn’t we expect to see it in Early Access soon?

1 Like

“Smart Summon was in Early Access testing from early June until it was released in late September, so that’s 4 months of Early Access.” Do you have a source for “early June”?

Tangentially related to lumpy/smooth/whatever, I notice that (current week number) - (Tesla software week number) is now ~10. This is largest number I have seen.

1 Like

:wink:

What a tease

This article is from June 7:

I have an intuition — really just an intuition, a hunch, a gut feeling — that we’re going to see new FSD features for city streets in Q4 or Q1 and that they’re going to work about as well as Smart Summon or Navigate on Autopilot. If not better.

I think we’ll see these features at least in Early Access, if not in wide release.

In April, Tesla had FSD in a state good enough to shoot that demo video and take people on rides. After alledgedly just 3 months of development focused specifically on city driving (and 22 months since Karpathy joined Tesla).

By June, Elon and presumably others were driving the dev branch of FSD on a daily basis and Elon claims it was working decently.

It doesn’t feel like a stretch to think within the next 2-5 months we could see at least some FSD city driving features get pushed to customers in the Early Access program. And then to wide release a few months later.

FSD features aren’t new science; they don’t need to be superhuman. They just need to work most of the time.

In terms of commutes automated from beginning to end, if you’re a Tesla driver and you have 1-2 interventions on your morning commute and another 1-2 on your evening commute, I could imagine that being a pretty good experience, depending on how on edge you feel you need to be.

What’s most exciting to me is how this could be a stepping stone. Enabling FSD features will enable learning from driver interventions. Learning from interventions is a flywheel for imitation learning. Imitation learning is perhaps the key way Tesla’s approach is different from everyone else’s and, therefore, the main reason it could work better than everyone else’s.

Interventions could also potentially surface computer vision errors and behaviour prediction errors that lead to bad vehicle behaviour on city streets. So, it could accelerate progress on those parts of the stack as well.

FSD features don’t have to be that great at first to still be a big deal for Tesla’s overall progress on autonomy.

1 Like

Here is a similar article from April :slight_smile: :

1 Like

Oh my! So Smart Summon was in Early Access testing for at least 6 months. I can see how maybe FSD could require less time to test than Smart Summon because you would rack up miles more quickly with FSD, but still… If we’re going to see a wide release by the end of 2019, surely there has to be an Early Access release pretty soon.

I wonder if “:wink:” means FSD will come out (in Early Access?) on November 1:

This is the tweet the “:wink:” was in response to.

Of course, Elon could plan for November 1 and end up pushing the update much later, as we saw multiple times with Smart Summon.

I would have so much more faith in it being able to drive around city streets, if it could drive straight down the highway with an incoming merge lane and not swerve into the incoming lane or not drive out the lane on relatively sharp interstate turns. Or maybe even just pull up behind a stopped vehicle less than one full vehicle length back. Or maybe start moving less than 3-4 seconds after the car in front of you started moving. I would even accept as proof of progress not changing lanes randomly in the middle of an empty interstate with no exits for 15 miles to “follow the navigation”…

Compared to navigating a turn lane… these are trivial tasks.

Elon just said “it’s going to be tight”, but he thinks feature complete Full Self-Driving is going to be out in Early Access by the end of the year.

I am concerned with quotes like this:

it will fill in the gap from low-speed autonomy with Summon. You’ve got high-speed autonomy on the highway, and intermediate speed autonomy, which really just means traffic lights and stop signs.

This might be the most worrying quote I’ve heard from Elon Musk on FSD. “It’s pretty much just highway driving but with stoplights.”

And even then, again, they can’t even follow a lane properly when another lane merges into it. There is so much more to highway driving that’s safe and enjoyable than “Stay centered in the lane, and don’t rear end someone.” Waymo hasn’t spent 10 years trying to solve stop signs and stop lights.