Self-Driving Cars

Without meaning to sound insulting by stating the obvious, Tesla is well aware of this and will program local driving laws and customs into each regional version of FSD. It is already employing engineering drivers in Europe, China, and other areas to perform road testing in order to gather data to inform these versions, and it has access to all of the driving data of Tesla vehicles operated anywhere in the world (with some limitations for local data privacy laws).

That does raise a point actually, Will FSD be a region-locked program? Or will you have to do a big reinstall of the driving data when moving a vehicle from one region to another?

That’s not a huge concern in North America, but Europe is liable to have a non of variance in how the FSD software will need to operate, even if Europe can be a single dataset you’re going to get people taking their cars across continental boundaries.

Yes, that is a good question. It depends mainly on the capacity of the FSD computer. If it has enough storage to handle regional sets of driving laws, then it won't be a problem. If it doesn't, then some portions of the software stack would have to be periodically downloaded if you take your car from nation to nation.

It would make no sense to require this each time you cross a physical border, so there might end up being different sets for North and South America, Europe, Asia, etc. I really don't know what solution Tesla will come up with.

Note that the same would apply to any autonomous vehicle. If you drive your Chevy UltraHyperMegaCruise car from the US to Canada, it would also have to understand the different road laws, signage, and so forth.

Edited by Fighteer on Feb 15th 2022 at 5:48:05 AM

Me, I'd worry more about the cameras. Not every country in Europe is going to like Teslas filming and saving recordings of arbitrary surroundings that may have people in them.

They already do. There are tens of thousands of Tesla vehicles operating in Europe, and they are at a minimum capable of using their cameras for Basic Autopilot. As far as I know, there have not been any serious data privacy issues raised by this, although I haven't researched the matter thoroughly.

I've seen quite a few video clips from Teslas in Europe showing up on various YouTube channels, so that's not a problem. Further, there are lots of dashcams in Europe and I am not aware of anyone getting sued or jailed over them.

Edit: When used privately, a vehicle's built-in cameras should be no different in terms of legality than aftermarket dashcams. When retrieved and used by Tesla for FSD training purposes, it would be governed by Tesla's own privacy policy and, while the EU may have some interest in that, Tesla is a US company so I'm not sure how the regulations would apply.

Edited by Fighteer on Feb 15th 2022 at 6:01:22 AM

Tesla can pretty easily argue for a legitimate use for the baseline data, it might just be required to clear out identifiable personal data from the footage it holds. So faces and potentially house numbers. Same way Google street view in parts of Europe has some stuff blurred.

The holding of footage by Tesla users is their legal lookout, Tesla is jsut providing a camera.

I guess this isn't really a self-driving thing, although it is related to autonomy features. Tesla recently made the news (again) when it became known that the NHTSA opened a preliminary investigation into "phantom braking" incidents related to use of its Autopilot software.

Phantom braking occurs when a semi-autonomous vehicle incorrectly detects that an obstacle is in front of it that requires engagement of Automatic Emergency Braking. AEB is a standard feature in many new vehicles, usually relying on a combination of machine vision and radar. It can be disconcerting when a car abruptly slows down without any obvious reason and may lead to crashes.

Given the amount of fuss this raised in the media, it is fitting that we have just learned (Ars Technica) that the NHTSA is also investigating the Honda Accord and CR-V for phantom braking.

I put this out here not to play silly games but to caution people against excessive reliance on individual news items. Phantom braking is not unique to Tesla or any other company.

Edited by Fighteer on Feb 24th 2022 at 10:28:04 AM

I figure I'll give this thread a poke. Drive Tesla Canada reports that a Cruise driverless autonomous vehicle was involved in a two-car collision in California on February 11, 2022.

Cruise has been allowed to operate driverless vehicles in limited areas at speeds of under 30 mph since last September. The accident in question was not the fault of the AV, but of a BMW driver (because of course it was) who made an illegal left turn at an intersection. Damage was minor and the driver allegedly fled the scene.

Driverless vehicle operators are required by California law to report any collision that involves property damage or injury, no matter how minor.

Editorializing: We are beginning to enter an era in which autonomous vehicles will be driving more safely than humans, making it more and more clear that it is people who are the problem when cars are involved. It'll be especially hard for humans to avoid fault when these cars are continually recording their surroundings.

Edit: I incorrectly said "Chevrolet" when I originally posted. Cruise is a startup not directly related to Chevrolet, although it is owned by General Motors.

Edited by Fighteer on Mar 8th 2022 at 9:28:46 AM

Teslarati: Tesla FSD Beta 10.11 release notes tease critical improvements

While Tesla FSD Beta users champ at the bit for version 11, which promises to unify the city driving and highway driving stacks, Tesla continues to focus all of its energy on the city portion. Elon Musk has spoken about this on Twitter, saying that the number of interventions on FSD Beta has gotten so low that they want to nail it down before moving ahead.

Anyway, Tesla has just released version 10.11, which contains a massive list of improvements, largely around optimizing for current weaknesses in the neural network's prediction abilities. Someone with more knowledge of machine vision than I am can probably glean significant value from them, which I'll post here and stick in a folder.

One of the biggest takeaways for me is that Tesla is trying to move away from reliance on low-resolution maps to predict how intersections will behave and towards using direct input from the cameras. Many of these changes are designed to improve the vehicle's ability to model how intersections are supposed to work, such as the correct paths for both the ego car and other cars to take. What we should see as a result is the vehicle making less of an effort to yield to cars that shouldn't cross its path and picking the wrong path less often.

Another really big note is that the cars are getting much better at eliminating false-positive detections of VRUs, or "vulnerable road users" - pedestrians, bicyclists, animals, and so on. What I get from the explanation is that a lot of phantom braking occurs when a vehicle sees a skid mark or tar seam and interprets it as something that it needs to avoid or stop for.

Even humans may have difficulty with tasks like these. For example, I was driving today and there was what looked like a seat cushion in the lane. I couldn't tell what it was until it would have been too late to avoid it, so I moved over in the lane in case it was something that might damage my car. It is very important that self-driving vehicles be able to perform the same sort of discrimination: is it an oil stain on the road or a pothole, a paper bag or a brick?

Another important thing is "adversarial VRU" situations, meaning intersections and streets with lots of pedestrians, bicycles, motorcycles, and such all moving around very rapidly and without regard for cars. We've all seen videos of what it's like to drive in countries like India, where people swarm around moving cars without a care in the world. Eventually, FSD will need to be able to handle these.

The programmers also seem to be working very hard to improve how the cars detect and react to parked vehicles and open car doors, which is very important when driving on city streets with parallel parking.

Optimistically, what I'm seeing here is that Tesla is confident enough in its pure vision software stack that it is able to continue refining its capabilities without running into local maxima. Local maxima occur when you reach the limits of your hardware or software in its current model, such that you're spending an enormous amount of energy trying to get it to perform better without much benefit. The only way to get out of a local maximum is to change your model.

Tesla has hit such maxima many times in the evolution of FSD, which is why it's taken so much longer than Elon Musk originally predicted. Dropping radar got rid of one local maximum by eliminating the cases when different systems would provide contradictory data to the computer.

I recent saw an interview on Tesla Daily with a former Tesla board member who doesn't think that the company will be successful with pure vision: that lidar will be a necessary part of any future self-driving solution. The jury is still out on that, but what I don't see is any other autonomy company being so transparent with its development.

Edited by Fighteer on Mar 13th 2022 at 9:54:00 AM

That's an inaccurate description of what a local maximum is. More broadly speaking, it's just that whatever function you're trying to maximise has hit a point where any movement away from this point would decrease its value—however, it's not the global maximum, and there are in fact better solutions. With regards to ML, it can happen just because the training parameters don't allow for enough variation to get over this bump once entered and find a new maximum somewhere in the solution space. It does not mean that the hardware or software has hit a limit (though if you're defining the solution space in terms of the end goal, it might entail that).

The usual simple diagram for demonstrating local maxima is a polynomial. Probably with four terms? That way you can get two maxima and one higher than the other. Sort of a lumpy M shape.

Interesting that they seem to have been modelling lanes as, basically, images? Or at least some sort of cluster of points that defines a lane, and now they've moved onto, uh... lines? Well, makes more sense from a human perspective.

Edited by RainehDaze on Mar 13th 2022 at 2:12:32 PM

Sorry, you are right. I was describing local maxima incorrectly even though I had the right idea in my head. However, it is also possible to reach maxima because of your hardware, such as the aforementioned sensor fusion issues with radar. Taking the radar out of the loop allowed for better performance because the cars were no longer getting conflicting data from different sensors.

Example: The radar sees a highway overpass but can't tell if it's obstructing the road or not. The cameras see an open road ahead. If the computer sides with the camera, it proceeds. If it sides with the radar, the car attempts to stop for the "obstacle". The only way to be sure which is correct is to make the car's vision system good enough to understand the situation by itself, at which point it no longer needs the radar.

What I understand about FSD development is that they've been progressively switching from raster space to vector space in more and more areas. As you correctly observe, this means that the software is no longer looking at a pixel and asking, "Is this pixel part of a lane," but rather plotting where the lane boundaries are as vectors in 3D space to which certain parts of the visual field belong and others don't. I could of course be off on the precise details.

Changing from rasters to vectors is one of those shifts you are talking about in the solution space. And yes, the goal is a fully autonomous vehicle, so anything less than that would be a local maximum from the point of view of the designers. They are working under the assumption that there is a global maximum that will achieve the goal.

Edited by Fighteer on Mar 13th 2022 at 11:23:55 AM

Sorry for the extra post, but the NHTSA has just paved the way for driverless autonomous vehicles by suspending its requirement that all vehicles have driver controls (Teslarati). This would be the equivalent of an SAE Level 5 system: able to operate in 100 percent of situations without supervision or intervention.

Note: NHTSA is not saying that any vehicles currently meet that standard, but that it is now prepared to allow them once they do.

Still need to transform the initial rasterised image into vectors, though, but it probably makes for easier route-plotting etc. when what you're working with describes a path.

I disagree with the idea of describing something as a local minimum if it's a limitation of approach or hardware, because that's a more precise term than needed. "We lack a physical component!" doesn't mean you're at a local minimum in a mathematical solution space.

Realizing that you are at a limit with your hardware can be a difficult thing, though, because engineers often fall into the trap of becoming married to their physical designs. Even more importantly, the people who pay the bills don't want to keep hearing that they need to spend a lot of money redesigning that hardware.

This brings us to the current debate going on in the autonomy world. On the one hand, we have the common wisdom that a sensor fusion of lidar and cameras, plus maybe radar, will be required for full autonomy. On the other hand, we have the comparably heterodox idea that cameras alone can fully solve the problem.

If we plot this as a mathematical solution space, we have two maxima that represent the best possible autonomy solution in each of those hardware environments. It's possible that each could be a full solution or that one of them could fall short.

Edited by Fighteer on Mar 14th 2022 at 8:37:56 AM

But we can't plot this as a mathematical solution. The solution spaces for two entirely different sets of inputs are non-comparable. Furthermore, the problem might lie outside of anything to do with fitness functions and gradient ascent (well, more likely gradient descent and we're looking for a global minimum, due to minimising loss) in the first place. Local maxima and local minima are a fairly defined thing.

A claim that everything, including hardware and the design of the approach and random things like road law are points in some bizarre nebulous "autonomous driving" mathematical solution space isn't making the use of the term any better, because it's not a mathematical problem with a mathematical limitation we're talking about. Really, just seems like one of those attempts to sound more scientific. But winds up not being because it's just overcomplicating and inaccurate. <_>

Edited by RainehDaze on Mar 14th 2022 at 2:08:23 PM

I won't try to argue the distinction any more, but what do we think of the lidar vs. machine vision debate?

That the advantages or disadvantages of either would seem to be best argued when we're not in a world mostly based around human drivers and signage.

That's an interesting point of view. The way I see it is that autonomous and human-driven vehicles will have to coexist for a minimum of three decades and developing entirely new roadways for autonomous vehicles will be prohibitively expensive.

This means that they will have to understand how to navigate in a world built for humans. Having done so, they might as well keep doing it. Sure, we'll probably adapt roads over time to be friendlier to autonomous vehicles, but it won't happen quickly and there will still have to be allowances for manual driving in certain areas.

Specific adaptations that I predict include the allocation of traffic lanes specifically for AVs, much like HOV and EV lanes today, improved standardization for signage and lane markings, and dedicated systems, like traffic tunnels, that are meant to be navigated exclusively by AVs. Other things that we've talked about already would need to occur in a world where robo-taxis become a dominant form of transportation, such as transforming parking lots into pick-up/drop-off lanes.

However, the exact manner in which the adaptations occur will depend on the dominant technology. If we go with lidar, then roads will have to become "smart", capable of handling the fast wireless connections needed to deliver high-resolution maps to the vehicles. If we go with machine vision, then this won't be an issue, but we'll still need to spend extra time ensuring that roadways and intersections present unobstructed lines of sight for the cameras.

Edited by Fighteer on Mar 14th 2022 at 2:09:51 PM

Mmm, but the entire question is "which of these is better at getting the necessary information to deal with the excessive unpredictability of the current roads". Eventually, the vehicles should be pretty behaved (and better able to signal their intent than flashing lights), and it becomes more of a question of if needing to build and maintain an entire visual recognition stack is entirely necessary.

Though this is also counterbalanced by the whole bit about not trying to make the world as obsessively designed around cars as the US has become, which will always introduce uncertainty.

I bet self-driving trucks wouldn't start fascist convoys that fuck up everything for everyone because they're insecure about their dick sizes. At least until the robot rebellion starts.

Nor would self-driving cars cut you off and brake-check you because you weren't driving fast enough in the passing lane. They would not fall asleep and drift across the median. They wouldn't rear-end a car in front of them because they were distracted by looking at their phones.

The list goes on and on. I would happily "sacrifice" driving myself if I could have sufficient confidence in the vehicle.

Who would have thought the robot rebellion was kicked off by not giving lorries proportionally large dongs?

Teslarati: Tesla’s camera-based driver monitoring system triggers legal complaint in Illinois

An Illinois plaintiff has filed a class-action lawsuit against Tesla for violating the state's Biometric Information Privacy Act with its driver-monitoring software. This software uses the cabin camera and AI to determine if a driver using Autopilot is paying attention to the road.

So now we have a double-standard. We expect semi-autonomous vehicles to ensure that drivers are paying attention while using their features, but we don't want our privacy violated by having cameras looking at us. Except this person's complaint is bogus since, as far as I know, the car doesn't store or transmit the data. I also expect that the defense will include the fact that their usage of Autopilot is voluntary and they agree to the terms when they use it.

I guess this is the sort of thing we're going to see more of as self-driving is more widely adopted.

Not really, as once cars are actually self-driving there will be no need for driver monitoring as there will be no driver.

That's at best a small part of the picture. While it is true that an SAE Level 5 vehicle does not need to monitor its driver (because there is none), such a vehicle employed as a robo-taxi will most definitely need to be able to recognize who and what is inside it. The company operating the taxi will have an interest in preventing vandalism, determining responsibility for damage, returning lost items, etc.

Now, I don't know that this monitoring would reach the level of biometric analysis, but facial identification might be used. Heck, the car might even be able to identify when someone is having a medical emergency and take them to a hospital instead of their destination.

As an example of in-car monitoring that we already know about, Tesla (plus possibly others) is either planning to deploy or currently deploying high-resolution radar systems in their cars to detect occupants and to distinguish them from luggage or other objects. Partly, this is so it can monitor seat belt use, but also for obvious liability concerns, such as the aforementioned robo-taxi driving around town for hours with a forgotten child in the back seat.

Even in the short term, since there's no guarantee that Tesla or any other semi-autonomous solution provider will be upgrading to SAE Level 4 or higher in the immediate future, these vehicles will continue to get scrutiny from lawmakers, regulators, and safety analysts over their ability to ensure driver attentiveness. Any attempt to do so will in turn create privacy implications.

Editing to add: I saw another report — from Twitter, so grain of salt — that BMW is switching its self-driving technology platform from Intel and MobilEye to Qualcomm and Arriver, with the first equipped vehicles in 2025. The goal of this collaboration is to develop an in-house Level 3 solution. Searching for reputable sources, I found this Forbes article from March 11.

What makes this notable is that it involves Green Hills Software and Dan O'Dowd, the man who ran a full-page NYT ad trying to get Tesla FSD banned from the roads. Green Hills writes code for MobilEye. Might want to disclose that conflict of interest, Danny boy.

Edited by Fighteer on Mar 17th 2022 at 10:54:53 AM