V11.4.2 is the second version of FSD Beta after the supposedly revolutionary improvements going from V10 to V11.
There have been intense discussions on the web lately about huge advances in artificial intelligence (AI). Apparently, programs like ChatGPT have reached a stage where they can answer conversational questions and write essays and poems, make medical diagnoses, write legal briefs, and create images and movies that do a credible imitation of what a competent trained person, camera, or video camera can do. This progress has come from using huge new supercomputers with massively parallel NVIDIA brand graphic processors repurposed to train on massive comprehensive datasets.
What are the implications of these advances in artificial intelligence for automatic driving programs? Tesla is using one of the 10 largest computers in the world to process data from my car and up to 400,000 other cars using FSD Beta to “train” future versions of the software. The Tesla approach uses cameras that substitute for human eyes and a Tesla custom designed “supercomputer” using neural nets that try to imitate the ad hoc brainy decisions that a human makes. Tesla is not using the approach that some other companies have used of completely mapping the driving environment so that all the decisions the car makes are specified ahead of time.
What’s my take on how this is working out so far?
We recently were able to download the second V11 version of Tesla’s Full Self Driving Beta (V11.4.2). We are spending the summer in rural Northern Wisconsin, so my ability to evaluate it in heavy traffic and complex situations is limited. However, there are still things that we can evaluate here.
In the low-traffic situation here, I can often put an address into the navigation and have it drive me the whole route without intervention. A big advantage to FSD Beta over regular Autosteer here in Northern Wisconsin is that the car will drive automatically and accurately on roads with no lines (no yellow lines in the center and no white lines on the edges of the road). It will also accurately navigate sharper turns than Autosteer.
One of the issues about previous versions of FSD Beta was too-timid (slow) behavior at stop signs. I’ve been making quantitative observations of V11.4.2 behavior at stop signs. V11.4.2 now always stops my car at the correct place at stop signs. This wasn’t always true before. The next question is how timid it is at stop signs. When leaving my house here, I have two stop signs where a smaller road tees into a bigger road and later tees into a state highway. I’ve been timing the period from complete stop to brisk acceleration. If the visibility of the bigger road is good, I’m seeing a pause of ~3 seconds until brisk acceleration. Unfortunately, I didn’t measure this pause on the previous V11 version, so I don’t know if that is an improvement. 3 seconds is quick enough that I don’t think a car behind me would be significantly inconvenienced. If the visibility of the bigger road from the stop sign is not good, the car creeps forward for a better view and it takes a total of ~5 seconds to brisk acceleration. If there is a car behind me or a car in the distance approaching on the highway, I will add a little accelerator encouragement.
For all 11 versions of FSD Beta that I have used, phantom braking has been an issue. However, I find it to be an annoyance rather than a show stopper. Phantom regenerative deceleration might be a better term in most cases. When using FSD Beta, my right foot is always resting lightly on the accelerator. If I have a phantom braking event and there is a car behind me, I apply some pressure to the accelerator to keep from slowing down significantly. With no car behind me, I let the car slow down and speed back up. I try to observe the “non-intervened” behavior of FSD Beta whenever possible.
I have observed new behavior with the last two versions of FSD Beta. I call it phantom swerving! I see this quite frequently on the two-lane roads where I usually drive here in Northern Wisconsin. The car acts like there is a person or object in the middle of my lane. It swerves abruptly into the opposing traffic lane and back. Once it even swerved onto the shoulder of the road. I intervened because I thought it might drive off the road. The problem: There has been nothing that I could see that justified this swerving.
For 6 months, my wife and I used 10 versions of FSD Beta V10. We recently got the so called “revolutionary” upgrade from V10 to V11. This upgrade was supposed to incorporate big improvements that Tesla has been working on for a long time. We got our upgrade to V11.3.6 a little over a month ago. In a previous article, I reported on my first impressions. This article is my report on what I have observed with the latest version, V11.4.2.
FSD review summary
Yes, there have been significant improvements. For example, when working correctly, FSD Beta V11 moves you over and gets you into the correct turn lane earlier. The lane changes are also much smoother. You also get a wider blue pathway center line that lets you know that FSD Beta is active (as shown in the picture at the top of this article). You still have the option of FSD-only on limited-access highways, which gives you a narrower blue center line, or Autosteer-only, which gives you two very narrow blue lines on each side of the lane.
However, still too frequently, FSD Beta puts you in the wrong lane. The worst is when it puts you into a turn lane when the navigation calls for going straight. This, to me, is the biggest failure of the current system. Note: Here in Northern Wisconsin, I don’t get a chance to observe this often. However, the town of Rhinelander (population 8,308) is 25 miles to our south. Today, while still in Rhinelander as we returned to Three Lakes, FSD Beta wrenched my steering wheel and put me into a turn lane that was inconsistent with the navigation route. So, it is clear that FSD Beta V11.4.2 has not solved this problem.
Why haven’t these problems been fixed? Tesla has huge teams of software specialists working on their total AI approach. Yet, I’ve seen no improvement in phantom braking and bad lane choices in the 11 versions of the software I have used. My take: The AI approach Tesla is using doesn’t address these problems.
I would love to hear about your experiences with FSD Beta in the comments section. Have any of you observed phantom swerving?
Have a tip for CleanTechnica, want to advertise, or want to suggest a guest for our CleanTech Talk podcast? Contact us here.
Former Tesla Battery Expert Leading Lyten Into New Lithium-Sulfur Battery Era — Podcast:
I don’t like paywalls. You don’t like paywalls. Who likes paywalls? Here at CleanTechnica, we implemented a limited paywall for a while, but it always felt wrong — and it was always tough to decide what we should put behind there. In theory, your most exclusive and best content goes behind a paywall. But then fewer people read it! We just don’t like paywalls, and so we’ve decided to ditch ours.
Unfortunately, the media business is still a tough, cut-throat business with tiny margins. It’s a never-ending Olympic challenge to stay above water or even perhaps — gasp — grow. So …