Self-Driving Cars

I’m sure that lots of regular taxis already have CCTV and that this is thus a solved problem. Monitoring of active customers/employees by the company selling a thing/employing a person is very different from monitoring of a product user post-purchase by a manufacturer.

I guess a key part may be where the data goes. Data just used in-car for the car to operate safely is easily defended, but data be used back at Tesla HQ for business purposes is different.

Edited by Silasw on Mar 17th 2022 at 9:00:08 AM

Sure, CCTV in taxis exists now, but that's monitored by humans, and there are also humans driving. With completely autonomous robo-taxis, there will need to be an AI component to these monitoring systems, otherwise you'll have to hire people to do nothing but watch video 24/7. Ironically, you get more privacy with the AI solution because while you can erase electronic data, you can't erase someone's memory.

Edited by Fighteer on Mar 17th 2022 at 5:50:43 AM

What does that have to do with anything? The point there is that taxis get recorded because if something happens, the video is necessary. And thus, we're already used to cameras being in place in taxis and busses for that reason. Ergo, the self-driving has literally bugger all to do with it. <_>

Why would you need AI monitoring? Why not do what is done with most CCTV where it’s unmonitored but reviewed by a human in the event of an incident?

I'm still a bit confused by the relevance of stuff related to hired transport to not needing driver monitoring in private vehicles if there's no driver.

I remain unsure why we keep arguing about something that seems fairly simple to me. To recap.

1. Companies that deploy semi-autonomous driver assistance systems with SAE Level 2 ratings also use various methods to ensure that drivers are paying attention, since they must be actively supervising those systems at all times.
2. Regulators and consumer watchdogs such as Consumer Reports heavily criticize ADAS that do not adequately monitor their drivers.
3. Tesla has recently begun employing its cabin-facing camera to perform driver monitoring. This uses AI to ensure the driver is looking at the road and has their hands on the wheel. There are other ADAS that use similar techniques.
4. An Illinois driver is trying to mount a class-action lawsuit against Tesla for violating the state's Biometric Information Privacy Act, claiming that the driver monitoring uses "biometric data".
5. To the best of my knowledge, no data from the monitoring system is transmitted to Tesla, nor is it stored in long-term memory, making this lawsuit absurd on its face.

The lawsuit calls to mind the operations of fully autonomous vehicles in the future, which will need to actively monitor their interiors to ensure that passengers are safe, to confirm who and what are inside, to detect vandalism or cleanliness problems, and so on. This monitoring will be done mostly by AI, giving rise to similar privacy concerns.

I'm not trying to do a "Tesla is being picked on" thing; I am curious about people's thoughts on those privacy issues.

Edited by Fighteer on Mar 17th 2022 at 7:30:14 AM

The bold bit is where you loose me, why would a fully autonomous vehicle have to actively monitor the interior? Why does the car need to know who is inside it beyond determining if they have the authority to tell the car to go somewhere?

Monitoring to me seems like it needs to be only the very basics of seatbelt’s being on and doors being closed. I wouldn’t need my self-driving car to tell me if it’s dirty, I can check that myself. Likewise a robo-taxi company can just have someone at the office remote into the camera after each ride and do a visual inspection for mess/lost belongings.

Vandalism is similar, you only need active monitoring (and thus an AI) if you’re running a prevention system rather than a “evidence for the police” system.

Edited by Silasw on Mar 17th 2022 at 11:37:01 AM

Okay, so maybe need is the wrong word, but it's definitely going to happen. If you have AI that can drive a car 100 times better than a human, it's relatively trivial to use that AI to look inside the car and make sure people aren't messing with it or having heart attacks or whatever. If you can do that, why wouldn't you?

As for privacy, as a consumer hailing a robo-taxi, it's not your car and you should have no inherent expectation of privacy.

Edited by Fighteer on Mar 17th 2022 at 7:46:59 AM

Because running that AI is gonna use up processing power from my car and use some of the power supply for my car. It feels like a luxury and/or something being overengineered. Also, because I’m cheap and won’t want to pay extra for a self-driving car that can tell if the ice-cream it’s bringing me has started to defrost.

As a consumer hailing a regular taxi you already don’t, so this feels like an already solved problem. Especially as the lawsuit you mentioned is about a vehicle owner rather than a rider in a robo-taxi.

Edited by Silasw on Mar 17th 2022 at 2:14:31 PM

"Why WOULDN'T you have a privately owned vehicle constantly monitoring and studying its owner at all times": because that is weird, overengineered, and frankly creepy. Presumably, you could have laptops have their webcams on at all point checking their owners aren't suddenly dead, but we don't go around saying it's a great idea for every device we own to actively study us. That cars can go places is not much of a change to this notion.

Our devices are actively studying us all the time, in case you didn't notice, and cars are increasingly software platforms on wheels, so it's a logical extension. Anyway, the in-car driver monitoring only occurs when a Level 2 ADAS is engaged and in use. It's not like the car is WATCHING YOU every second.

And most people with any sense correctly realise that always-online always-active cameras are an incredibly stupid thing. Plus, you're swapping timeframes again. As Silas said (and I agree), it doesn't really matter in the long term about driver monitoring concerns, because the idea is to not have a driver. The only time in which you actively need such a thing is when the driver is actively required to be monitoring the situation even if the car isn't.

I am clearly talking about two different things, and while they are not perfectly similar, both involve active monitoring of the occupants of a vehicle, the data from which could potentially be a privacy risk if not properly controlled.

The question I am asking is: what do we think of that privacy risk in the current scenario, in which a Level 2 ADAS actively monitors a vehicle's driver, potentially using "biometric analysis"? Is the Illinois lawsuit justified or frivolous?

Edited by Fighteer on Mar 18th 2022 at 2:05:47 PM

Oh, unless the data is being sent somewhere the lawsuit is a crock of shit.

If the car is just gathering the data to use internally then the owner never looses possession of their data, it’s just sitting inside a devise they own.

Unless Tesla has been doing Facebook style data harvesting it’s the equivalent of a lawsuit against the fact that a car will beep at you if you drive without a seatbelt.

As far as ensuring that drivers are still paying attention to the road, I'm pretty sure that the feed from the internal camera would only be used by the car's computer - much like the steering wheel sensors on every car with a similar system (as well as the driver-facing camera used for a similar purpose with GM's Supercruise system).

The only possibly questionable thing I can think of would be the "Safety Score" system used to decide who gets to participate in the full self-driving beta. Even that may not run afoul of the law in question if the car is calculating the score by itself and not exchanging data with Tesla HQ. Though if they ever start using these scores for their new car insurance, that may also be cause for concern.

Tesla is already using the Safety Score for its self-offered insurance in all states where it is available except California, where the use of vehicle telematics data to set insurance premiums is prohibited by law.

I am not sure whether the complete data set is transmitted to Tesla or only aggregated data, but I do know that the company is capable of collecting that telematics data from any vehicle in its fleet regardless of whether the owner is subscribed to Tesla Insurance or FSD Beta. This can be used for a variety of purposes:

• To optimize general vehicle performance.
• To establish fault in a crash or injury. For example, when there was a rash of accusations of sudden unintended acceleration (SUA) in Teslas, the company was able to prove conclusively that alleged incidents were due to pedal misapplication. Same with reports of brake failures, which was a brief thing in China: Tesla was able to prove that the drivers failed to apply the brake or did so inadequately.
• To develop Full Self Driving, as manual inputs may be compared to what the software would have done in a given situation while it is operating in "shadow mode".

In these cases, the data are associated with the VIN, not with the driver specifically, although the use of Driver Profiles could allow the company to know who was ostensibly driving at the time the data were collected. In theory, the in-car camera could also be used to identify the driver — in fact, a frequent request from customers is for Tesla to use facial recognition to automatically select the driver profile, much like your smartphone can use facial recognition to unlock itself. Like the phone, this recognition would be done entirely by the vehicle and would not transmit biometric data to Tesla.

Edited by Fighteer on Mar 26th 2022 at 4:57:30 AM

Oh yeah, Shadow Mode. That's been running for what, 2-3 years now? Odd that privacy advocates have yet to complain about that, but Tesla is facing a highly questionable lawsuit over a driver-facing camera used for what is by all logic a legitimate safety feature (and again, not much different from what GM uses for Supercruise - yet no one's complaining about that either).

Tesla FSD Beta 10.11.1 was released last night and has begun rolling out to subscribers in Canada. The release notes seem to be the same as for 10.11, or at least I can't see the difference when I look at them, so Canada is probably the major addition. The rollout seems pretty small at the moment, limited mainly to drivers with a Safety Score of 100.

Some people immediately noticed that there is a geofenced area in downtown Toronto where FSD Beta will not operate. In response, Elon Musk said that the reason is that Tesla does not yet have accurate modeling data for Toronto streetcars.

Teslarati: Toyota adopts Tesla’s camera-only approach to self-driving development

In another "if you can't beat 'em, join 'em" moment, Toyota is investing in camera-only self-driving technology through its subsidiary Woven Planet, according to a Reuters report (paywalled).

Citing cost reductions of up to 90 percent on hardware and the simplicity of installation, the spinoff plans to process a large volume of "low-fidelity" data rather than a small volume of "high-fidelity" data, meaning lidar-based 3D mapping. This is precisely the approach Tesla is taking.

The technology won't be ready overnight. Toyota still plans to go with lidar and radar for its near-term robo-taxi and other autonomy projects, but Woven Planet's approach seems to be intended to replace it over time, and especially in vehicles intended for private use.

Here we go. The New York Times released a documentary today on FX and Hulu entitled "Elon Musk's Crash Course". This apparently explores the difficulty that Tesla has faced in making its self-driving concept a reality.

I have not watched it, but from what I've heard it focuses on fatal and serious crashes involving Autopilot that have already been extensively litigated in the court of public opinion as well as in actual courts. I don't know if it acknowledges that the Autopilot of 2019 is not the same as the Autopilot of 2020, 2021, 2022, etc.

This is another case where the narrative is more important than the reality: Tesla's own data purports to show significant safety increases when using Autopilot and even better results with the recent iterations of FSD Beta, but the perception in the media seems to be of death machines attempting to kill their drivers.

Anyway, Tesla has just pushed out FSD Beta version 10.12.1 to beta program customers with the promise of substantially expanding the release on version 10.12.2. Tesla currently claims around 100,000 users of FSD Beta and Elon Musk has indicated that it may expand to one million drivers by the end of this year. There is some question as to where that figure comes from since nobody thinks that the FSD program has that many takers.

It doesn't look as if anyone has taken the time to transcribe the release notes, but screenshots of them have been shared around Twitter. I'll link to this article that shows the notes. They are far too long for me to reasonably type out myself, and I don't speak ML, but I'll try to summarize.

• The primary emphasis is on improved reliability when making unprotected left turns across a wide road. In particular, the vehicle will be more certain in its decisions (it won't try to go and then suddenly change its mind), it will accelerate faster through intersections, and it will make smarter use of wide medians in which it is possible to wait for traffic to clear.
• Significant improvements in lane predictions and also an upgraded planner for maneuvering in restricted space.
• Improved traffic control behavior, particularly false slowdowns when approaching yellow lights. This seems similar to the left turn improvements: the vehicle will be more certain in its predictions and not suddenly appear to change its mind.
• Improved "understanding of visibility", which I believe means modeling occlusion and line of sight blockers in vector space for improved awareness of what parts of the driving space it is and is not able to see. Related to this is improved "creep" behavior, where the car edges out into an intersection to improve visibility.
• Improved recognition of motorcycles, so the car will "see" them more reliably.
• The car will more accurately recognize vehicles with open doors and will display them in the user interface. In previous versions it would drive uncomfortably close to them.
• Improved recognition of when a car is parked on the side of the road versus when it is stopped for traffic. This should resolve cases where the software would try to "pass" a lane of stopped traffic.
• Better lane changing behavior when a car in the destination lane slows down unexpectedly.
• Improved system frame rate, which is impressive when you consider how much FSD pushes the hardware.

Curious question: how are self-driving cars supposed to work? How do they detect other cars, people in front of you, random pieces of garbage or debris that shows up in your way, etc?

How much of the car, to put it simply, is actually driving itself?

Do you want the long version, the short version, or the short, short version?

The short, short version

It depends on which technology you're talking about and which level of functionality you are talking about. There are two broad categories: the car assists the driver with basic functions but the driver remains in control; and the car is driving itself, understanding its environment, recognizing road markings, vehicles and pedestrians, traffic control, and planning its route with limited or no human intervention.

The short version

There are two broad approaches to self-driving and a classification system by the Society of Automobile Engineers (SAE) that describes what level of autonomy a vehicle is capable of. The two approaches are high-resolution mapping using lidar and machine vision. The SAE chart gives ratings from 0 to 5. At levels 0 to 2, the driver is completely responsible for the vehicle. At levels 3 to 5, the vehicle is increasingly responsible, with the driver acting in a supervisory capacity or even completely out of the loop.

Lidar vs. machine vision

A lidar-equipped vehicle operates by driving along pre-planned routes using high-resolution 3D maps that are created by specially equipped vehicles that drive through each geographical area. The lidar allows for an extremely accurate understanding of the vehicle's position on the map, and it uses a combination of radar, lidar, and optical sensors to detect obstacles and navigate along its defined route. It cannot navigate outside of mapped areas, however.

A machine-vision equipped vehicle uses only cameras, without lidar or radar, and depends on low-resolution maps for navigation, such as Google Maps or Apple Maps. It applies machine learning (artificial intelligence) to build a comprehensive understanding of its environment and makes driving decisions on the fly within that environment. It does not require high-resolution maps so it can in principle go anywhere that it is capable of understanding.

The SAE scale

• Level 0: No autonomy. The driver does all the work, but the car may provide warnings and emergency assistance (lane departure warning, automatic emergency braking).
• Level 1: Steering or speed control, but not both. This can include features like lane keeping and adaptive cruise control.
• Level 2: Steering and acceleration, which usually means both lane keeping and adaptive cruise control at the same time. Most consumer driver assist systems currently on the market are Level 2, including Tesla's Basic Autopilot, Subaru EyeSight, Ford BlueCruise, etc.
• Level 3: The car is driving itself under limited circumstances. The driver is required to take control when the car requests, but can take their eyes off the road otherwise. Some ADAS software claims to offer Level 3 capabilities in geofenced regions, such as specific sections of highway.
• Level 4: The car can drive itself under most circumstances with occasional assistance. Manual controls may be removed. This is what Tesla is working towards with its Full Self-Driving software.
• Level 5: The car can drive itself under all conditions and there no longer needs to be a human supervising. This is the ultimate goal of all self-driving systems.

The long version

Ha, thought you'd get me to write a research paper, did you? I can point you at more resources if you're really interested in the details.

Edited by Fighteer on May 23rd 2022 at 10:27:44 AM

Give me the long version, Fights. I didn't come here for "Dummy's Handbook for Self-Driving Cars", I came here to learn!

Edit: How dare you stiff me on a wall of text, Fights! You better have some good and varied sources as recompensation!

Also, my spell-check is saying that "recompensation" is spelled wrong. It's weird, I heard the word enough that I can't be making it up, it must be a real word.

Edited by fredhot16 on May 20th 2022 at 8:37:08 AM

The word you are looking for is "recompense".

It's 10:30 PM my time. I'm not writing much more. I did already link you to the SAE website, where you can learn more about the classification system. The Wikipedia articles on vehicular automation and self-driving cars should give you all the references you can handle.

Huh, it's 9:38 P.M here. You're an hour ahead.