Actually, it does: I don't know what you mean by "faster lane to pass". Doesn't sound like a nav feature.
If it doesn't, you'll eventually get an OTA The difference here is it's saying "take lane X", versus "you are in lane Y but need to be in X, so change lanes." Autopilot and nav right now are not connected, and Autopilot lacks awareness of which lane you're in — it just knows that you're in a lane and (most of the time) whether there's a lane to your side(s).
Nav tells you which lane you need to be in to continue your journey. If you are not already in that lane, AP will (apparently) tell you which lane to be in to continue your journey. Are there people out there that hold a current driving license that would find that extra functionality genuinely useful? "Oh, wow, AP has not yet notified me that I have to change lanes, and as I am nearly at the junction, I must therefore assume that the lane shown by the nav screen one mile earlier was in fact the lane I am currently in!" Powerful stuff.
I'd imagine it's a pretty simple feature in practice: 1) Nav says you need to leave the motorway at the next exit (2 miles away) 2) 1 mile away from the exit, Autopilot makes a beep and says "lane change required for exit, proceed?" 3) You confirm it somehow, with a tap of the stalk or the indicator or something 4) AP users the side repeaters and ultrasonics and some object ID tracking etc to try and dodge cars whilst changing over however many lanes it needs to. 5) Once you're in the correct lane, it'll signal when you need to take the exit, and then also take the exit 6) It'll slow down and bring you to a gradual stop at the end of the exit And, I imagine, this will be a US only feature for the foreseeable future (as the current slow down for the 'off ramp' feature is) Useful? Well, I guess so - but not with the current nag rate...
It's the first step to automatic lane changing, merging, and the car taking the exit by itself. It could be especially useful in complex and poorly signed interchanges in areas the driver is not familiar with, depending on how early it prompts and how smart the lane choosing knowledge is.
We already have the first step to auto lane changing, live in all AP cars today. Works really well in my experience. Next step is for the car to #justdoit. Auto-indicate for 3-5 seconds then make the lane change if safe to do so. Not convinced that we need "AP lane advice" as an intermediate step. If the nav is already telling you to get in a specific lane, there is minimal utility in AP giving you the same information in nag form. The feature set for EAP is that the lane change is done fully automatically, without driver involvement. "AP lane advice" is not in any of the AP/EAP/FSD feature sets. Personally, I would prefer it if V9 arrived without "AP lane advice", and the real EAP auto lane change arrived in a later point release.
This may not be possible if Tesla is going to use drivers’ input to train or validate the lane changing software. But I wouldn’t worry too much about this for now. We will find out what’s actually in the release. Electrek’s information might be wrong or it might be a broken telephone problem.
In Elon's appearance on the Joe Rogan podcast this week he described the on-ramp to off-ramp functionality and said it was about to be released: "We're about to release the software that will enable you to just turn it on and it will drive from highway onramp to highway exit, do lane changes, overtake other cars, go from one interchange to the next. If you get on say the 405 and get off 300 miles later and go through several highway interchanges and overtake other cars and hook into the nav system." Of course as others in this thread have said, that may be a 9.x release, with the initial 9.0 release requiring human intervention. I suppose it all comes down to the definition of 'about to release'. Knowing Elon, this could mean anything.
Oh and maybe I'm just being overly optimistic, but it seems reasonable to me that an 'air gap' for automated lane changes would be part of the early dev builds, but will be removed for the general release....
Not a nav feature, an autopilot feature. That's what we're talking about, isn't it? From Tesla's EAP page: "In addition to ensuring you reach your intended exit, Autopilot will watch for opportunities to move to a faster lane when you're caught behind slower traffic" Autopilot Your photo shows the car indicating an offramp when it's the only possible choice to continue following your nav. What about on the open highway when there's no upcoming nav maneuver? What about ramps/junctions/etc with multiple possible lane choices (for example, at many major junctions by me, there are two or three valid lanes for a given maneuver). No navigation system I've ever seen will make recommendations such as "the car in front of you is driving 10 below the speed limit and there's a clear passing lane, so you should use it", much less offer to execute the maneuver for you. Automated lane changes with confirmation required is absolutely an incremental step towards fully automated lane changes. There's a fundamental difference between a driver-initiated action, even if it's completed automatically, and a technology-initiated action, even if it requires confirmation. This is, by definition, increased autonomy. Within the realm of automation and AI, the party that initiates the action is extremely significant to determining level of autonomy (in a general sense, not necessarily by the SAE levels of autonomy). So much so that the recommended-action-with-confirmation paradigm is an extremely common incremental step between technologies that assist a human worker and full automation, especially in problem domains where the cost of making a mistake is high. If nothing else, this is a fundamental shift in how the car operates, even if the practical effect is significantly less than that of full autonomy. You might not find this incremental increase in autonomy useful or exciting, but please disabuse yourself of the notion that something you don't find useful or exciting is universally useless or unexciting.
My photo is proof that nav knows which lane to be in to continue the journey. In this situation, having AP also telling me, the driver, to confirm a lane change is a pointless annoyance. I do this already, using auto-lane change, because I am paying attention to the road. Why do I need a nag to do the same thing? I understand the need for confirmation. It is probably safer and it definitely keeps the liability firmly on the driver. However, manual confirmation is not an EAP feature: "Your Tesla will match speed to traffic conditions, keep within a lane, automatically change lanes without requiring driver input". Extremely common? Trying to think of a technology that has moved past the recommended-action-with-confirmation paradigm, and struggling. Maybe you can identify some common examples that we would all know? Anyway, I prefer a notify-action paradigm. Car decides to change lanes. Signals for X seconds. Starts the process. User (who is still legally required to be holding the wheel) can cancel at any time. The system that was described will merely be a back seat driver, constantly nagging about what to do next, sometimes getting it wrong, and adding little-to-no value over the current system. I guess some such as yourself will find this exciting and useful, but personally I would rather wait a bit longer for a more confident approach.
There is some uncertainty still about whether the production software update will actually require driver confirmation, or whether that is just the dev version. So we may be splitting hairs over nothing. I would reiterate that with autonomy features we are in the deep learning a.k.a. “Software 2.0” paradigm. Meaning that large labelled datasets are necessary for progress. There may be new autonomy features where for several months or quarters Tesla will need drivers to label the data by taking an action like confirming or cancelling a lane change. Saying yes or no to the car’s decision to, say, overtake a slow truck is still more autonomous than making that decision yourself. But, more importantly, this sort of set-up would a temporary bridge to get us to the point where the car can safely make decisions with no driver input. It isn’t a matter of engineers spending more time coding. The neural network is doing the coding. The engineers are just designing the pipeline that collects, cleans, labels, and curates the data before feeding it to a neural network. That’s why we can’t just wait a bit longer for the engineers to figure it out. They need the data from everyday Tesla drivers.
I think you are taking Karpathy's talk a little too literally. There is no evidence that the Tesla Vision NNs are doing any vehicle control. I also think you have misunderstood the role of the Tesla Vision NN and the amount and type of data that is being collected. Suggest you track down posts from @jimmy_d who has decomposed the actual production NN in detail on here. Of course v9 might be another step change, but there again it may just be a fresh coat of paint over what we already have. With MAME. But you are right that this whole request/confirm approach is just a rumour at the moment.
We don’t know which modules of Tesla’s software use hand-written code vs. neural networks. We just know that Karpathy wants more elements of the system to use Software 2.0, and this would be in line with Elon’s views on AI as well. Maybe path planning and control for lane changing are both still hand-coded. Could be! But I would bet the prediction software that estimates the trajectory of other vehicles (a crucial aspect of safety changing lanes) is a neural network. I think everyone else is using deep learning for prediction. My impression of the folks who have hacked their cars or looked at information from hacked cars is that 1) they only have small fragments of information and 2) they are largely just making educated guesses based on those small fragments. A lot of the detective work is just looking at the names of files and speculating from there. It’s a big inferential leap to go to a file name to a hypothesis about how the system works. So I think there’s very little conclusive that can be gained from those threads. Interesting fodder for speculation, sure, but few hard facts. Tesla is super secretive about its technology, and the very limited stuff that hackers have been able to access is only what Tesla decided it didn’t want to keep secret. It could use encryption or take even simpler steps to prevent anyone from gleaning even as little as they have so far — like giving the files deliberately misleading, coded, or non-descript names. Tesla is of course aware of all the detective work happening on TMC, and has disabled OTA updates on hacked cars. So if there is anything Tesla would consider a sensitive trade secret or valuable IP, I would imagine that Tesla would render that inaccessible. Tesla doesn’t even like to disclose its battery pack costs. Tesla sued it owns former Autopilot lead to check his devices from any stolen info. This makes me skeptical how much we can really learn from hacked cars.
I guess you didn't read the posts I referenced earlier. jimmy_d has a background in AI and has been looking at the actual Tesla NN.
I have read the posts, yes. Based on what he’s written, I think jimmy is basing a fair amount of his educated guesses on the names of files. I don’t believe jimmy has access to everything that constitutes Enhanced Autopilot. I don’t see any post where he indicates that. In the Neural Networks thread, jimmy acknowledges there may be neural networks he doesn’t have access to, and he acknowledges that he’s sometimes speculating. You’re right: the vision neural networks are not doing path planning or control. They are doing vision. But that doesn’t mean that other neural networks aren’t doing path planning or control. One of the parts of path planning is prediction. It looks like Waymo is using deep learning for prediction. The same could be true for Tesla.
The release will not be as fast as you want, but I think they'll introduce an alpha or beta version to warm up interest.