Self-Driving Cars

It needs to be more than “the driver doesn’t have to pay active attention”, it needs to be “the vehicle does not need a driver”, as long as the user is legally considered the driver/operator it’s not a full self-driving car. A fully self-driving car means you don’t need a driving licence to be the only person inside one, it means you can be drunk in one, it means I can tell my car to meet me at home and not be in breach of anything when it drives there itself.

The fact that it will have to be tested in court is going to be the limiter, people aren’t generally going to pay for a self-driving feature they can’t use safely (due to legal risks), so that legal status has to be settled before we’ll see any scale of adoption.

Now, Musk is probably exactly the level of memelord troll that he might get it tested in court by (after receiving authorisation from federal regulators) having his Tesla do laps around a police station in every state while he sits in the back tweeting about Space "X".

He might do that. After all, this is the same Elon Musk who rented a truck to drive his Falcon 1 rocket to the front steps of Congress to show that he was able to compete with the big contractors. He also drives with the bleeding-edge alpha version of Tesla FSD in his personal car.

We need to make a distinction between levels of autonomy, though, because it's important to this discussion.

Tesla is currently rated at Level 2 autonomy, which is still classified as a driver-assistance system. In Level 2, the car controls steering and acceleration/deceleration with the human driver supervising and taking over in situations that the car can't handle.

At Level 3, the car handles all aspects of driving but may require the driver to intervene. This means there must be a driver in the seat and they must be paying attention, although they don't need hands on the wheel at all times (unlike Level 2). The human remains ultimately responsible.

At Level 4, the car is fully autonomous but may ask for intervention from time to time. If the human does not respond, the car can safely pull over. The main difference here is that the car is responsible for its own actions rather than the human driver. If it messes up, it's considered the car's fault.

At Level 5, the human is out of the loop completely. The car is totally independent. At this level, driving controls may be removed.

Full robo-taxi service requires Level 5 autonomy, since any human in the car is a passenger, not a driver. However, limited forms of it would be possible under Level 4, such as the car being monitored remotely. Some of the current lidar-based solutions technically qualify as Level 4 but only within precisely 3D-mapped regions.

These gradations allows self-driving systems to gain regulatory approval for increased autonomy over time rather than suddenly jumping from no assistance to "let 'er rip".

Tesla is very close to Level 3 with its current FSD Beta. I would be surprised if it isn't released to all FSD subscribers by the end of the year. At that point it's a matter of data. The NHTSA isn't swayed by emotional arguments; it wants to see hard statistics showing that the vehicle is as safe as or safer than a human driver. This means hundreds of millions of miles of road testing by real people.

Edited by Fighteer on Feb 26th 2021 at 10:41:00 AM

Yeah I think for self-driving to have the revolutionary impact you expect, we have to wait for it to hit level 5, because that’s what people think when you say that a car is self-driving.

This is what irritated me the most about that Cadillac ad in the Super Bowl. Sure, it may be rated for hands-off highway driving in limited areas, but you definitely can't get in at home and tell it to drive you to the grocery store. A Tesla (with FSD Beta) absolutely can do that, but because it's only 99 percent reliable rather than 99.9 or 99.99, it can't be advertised in the same way as the Cadillac.

Edited by Fighteer on Feb 26th 2021 at 10:44:28 AM

1/100 situations (or whatever comparative ratio it winds up as) needing intervention for a car is definitely too low to advertise it that way, in my opinion. All advertising has to be based on the assumption that people are going to be stupid and not pay attention.

Well, Tesla doesn't advertise, so the point is moot. The software is clearly labeled in accordance with Level 2 restrictions even if it's capable of Level 4 or 5 in internal builds.

Please don't take my "99 percent" figure too literally either; I have no idea what the actual statistical accuracy rate of FSD Beta is. Elon has talked about "chasing the nines" by way of saying that as safety increases past certain thresholds it becomes exponentially harder to keep increasing it.

ETA: He has also, however, talked about Tesla's latest 4D machine vision technology in revolutionary terms, saying that it can solve way more problems than just driving. Elon memes and hypes and occasionally overpromises but he doesn't lie about technical subjects.

Edited by Fighteer on Feb 26th 2021 at 11:43:45 AM

Yes, you've quoted that part a billion times. All ML techniques run into problems with improving accuracy eventually because of the difficulty of weighting uncommon cases and common ones whilst trying to keep generalisation broad enough; breaking it into smaller solvable problems is obviously one advantage but you still need to recognise the situation first. Though people are prone to some really dumb oversights, which doesn't help.

Great, and as long as it's a proprietary trade secret I hate him for it.

Edited by RainehDaze on Feb 26th 2021 at 4:45:16 PM

Huh, I didn't know about this "Level" involving restrictions in self-driving cars until right now.

It's a bit serendipitous that we were just speaking about self-driving cars needing to dodge aggressive human drivers, because guess what just popped up on my Twitter feed?

Teslarati: Tesla Autopilot foils BMW SUV’s apparent pit maneuver attempt on Model 3. Since the article is now paywalled (really, Teslarati?), here's the original tweet with video.

In the video, a Tesla Model 3 on Autopilot is driving in the fast lane on a highway when a BMW driver overtakes it on the right at high speed. The BMW swerves into the Tesla's lane in what is unmistakably an attempt to run the car off the road by striking its front quarter panel.

The Tesla sees the maneuver and perfectly dodges it without any intervention from the human driver. No damage is suffered, although there may have been some soiled garments afterwards.

The BMW driver is lucky. Teslas are extremely stable due to their high weight and low center of gravity. The BMW would almost certainly have flipped and rolled, seriously injuring or killing the driver.

Edited by Fighteer on Feb 27th 2021 at 2:01:36 PM

That's going to present an issue for police. Either: A) they have no way to stop a car if someone is using it for illegal purposes, or B) all the problems with building a back door into every vehicle on the road.

I agree. It would be ridiculous to assume that Tesla and other self-driving developers haven't thought of it, but I have no idea what their solution space looks like. For a machine vision system, it's a labeling problem like everything else. You need a label "police car" that encompasses all the different kinds that might be found out there, plus "police car with lights on" and "police car with lights on telling me to pull over".

Put in those terms, it should be solvable, although not without some trial and error.

What I would do, actually, is have the self-driving car communicate back to its fleet that it is being pulled over. A human operator would jump in and evaluate the situation for safety, such as determining if it's a real cop or someone pulling a trick. That operator could contact the police on behalf of the car.

Correspondingly, police departments will need training on how to handle self-driving cars. You can't interact with them in the same way as human drivers.

I’ve done some googling and it looks like some self-driving car developers we doing just that, they’re teaching the system to recognise police officers/police cars and to use the length of time an active police cars sits behind them to determine if they’re being pulled over.

Now obviously the system would still be susceptible to someone faking being a police officer, but so are regular drivers, so that’s not a new problem. Hell, a self-driving car could probably more easily to the safety check of phoning the local police station and asking for conformation that the pullover in real.

Teslarati: Honda launches $102k Level 3 autonomous Legend sedan after 807k miles of highway testing

Honda has announced the launch of the world's first certified Level 3 autonomous driving technology in Japan, but it comes with a few caveats.

First, only 100 units will be available of this Legend sedan, and they're priced at a hefty 11 million yen ($102,000). Second, the Level 3 functionality is limited to congested traffic conditions and speeds of 50 kph (31 mph) or less. Third, Honda says it's done 807,000 miles of road testing with this feature. The self-driving equipment reportedly costs "several million yen" and the system as a whole is not intended for mass market use.

The Legend is already equipped with some Level 2 features equivalent to Tesla's base Autopilot or GM Supercruise, such as lane changing and speed following. The driver is not required to maintain eyes on the road during Level 3 operation, but is responsible for monitoring the system and resuming attention when it exits congestion mode and resumes normal speeds.

I'm sure this will come across as defensive, but anyone who talks this system up while ignoring that it only has 800K miles of real-world data vs. Tesla's billions needs to rethink some things.

Edited by Fighteer on Mar 5th 2021 at 10:53:54 AM

How many miles of real-word data does Tesla have in the specific conditions this is trained for? That seems like the obvious potential split, that and how the data was gathered, Tesla has been gathering lots of data via crowdsourcing, which will provide less valuable data than trained testers testing under specific conditions that need testing.

Not all data is equal, so we should be comparing more than just the size of the data gathered.

Honda also has how much past experience in robotics and visual processing again?

That's not even getting into how data requirements differ between different approaches.

Well, the important part is that it works and is safe. I'll give that a win rather than argue over statistical interpretation.

Elon Musk tweeted that Tesla is accepting requests to join the FSD Beta program, which is being expanded with the upcoming 8.2 and 8.3 releases. His enthusiasm over the software is infectious but he keeps stressing the need for caution lest it go into full use too quickly.

Edit: As an update to the above, apparently the next software release will give existing FSD customers the ability to opt-in to the beta program [1].

Edited by Fighteer on Mar 6th 2021 at 8:10:41 AM

CleanTechnica: Lidar May Be Harmful To People & Cameras

First off, it's important to note that the issues raised in this article are not well-studied at a formal level, but should be considered nonetheless. That out of the way, there are concerns that lidar systems can damage both human eyes and camera sensors.

There are regulations on the maximum power of lidar emissions precisely because of these risks. Early sensors used 905 nm light that penetrates to the retina and can cause serious eye damage or even blindness. More modern sensors use 1550 nm light that the cornea is largely opaque to, but cheaply made ones can still "leak" 905 nm light. These sensors can also operate at much higher power, giving increased range and resolution. It is not thoroughly demonstrated, however, that they are safe around human eyes.

Further, lasers of all types can damage CMOS sensors used in modern cameras.

Right now these risks are limited because of the relatively low quantity of lidar-equipped cars on roads, but if this technique becomes widespread in self-driving vehicles, we could be forced to deal with them on a wider scale.

Edited by Fighteer on Mar 12th 2021 at 1:31:45 PM

That's well into infrared, right?

Both 905 nm and 1550 nm are in the infrared range. Specifically, near infrared.

Visible spectrum bottoms out at about 700 nm, so yeah.

The article feels very just-asking-questions, though. "We aren't saying that this is an actual health risk, but no one's proven that it isn't one. Anyway, thanks for the click."

I agree. It's very speculative, which is why I added the disclaimer.

Ars Technica: Federal investigators blast Tesla, call for stricter safety standards

A little background is in order. In the United States, there are two government agencies that share overall responsibility for automobile and road safety. The National Highway Traffic Safety Administration (NHTSA) broadly regulates vehicles and sets standards that manufacturers comply with. The National Transportation Safety Board (NTSB) performs investigations into safety incidents after the fact and issues recommendations.

Over the past several years, the NHTSA has been very relaxed (some would say very lax) about regulating automated driver assistance systems (ADAS) and claims related to self-driving technology in general (as opposed to Europe, which has been extremely strict). The NTSB has investigated several serious crashes involving the use of these technologies and issued stern recommendations that have largely gone ignored.

Since 2016, six crashes involving ADAS and autonomous systems have been investigated by the NTSB, four of which were fatal. One was a pedestrian fatality involving an autonomous Uber vehicle. Several incidents involving Teslas were caused by drivers not supervising the Autopilot system properly and/or using it in environments for which it was not intended. As an example, Tesla's baseline Autopilot is only supposed to be used on limited-access highways, but drivers can activate it on almost any road. As another example, Tesla's software measures steering wheel torque to determine driver attentiveness, but many report being able to easily fool it.

The NTSB has recommended more aggressive regulation of self-driving vehicles, including software restrictions on their use in unapproved scenarios and better monitoring for driver attentiveness. However, the NHTSA largely ignored such recommendations under the Trump administration. This new recommendation comes as the Biden administration is selecting its top-level staffers for the Department of Transportation and the NHTSA, so it appears to be a play to get these concerns recognized by a more attentive White House.

Edited by Fighteer on Mar 17th 2021 at 11:35:46 AM

CleanTechnica: IIHS Dishonestly Using Speed Limits To Bash ADAS

Here's an interesting one. The Insurance Institute for Highway Safety recently published a study claiming that drivers using advanced driver assistance systems (like Tesla Autopilot) are 10 percent more likely to have fatal crashes because they speed more while ADAS is on.

CleanTechnica is obviously pro-Tesla and pro-self-driving, but the conclusions are grossly dubious. The actual data found that drivers using ADAS go an average of 1 mph faster than drivers without, and that these differences are most significant on roads with lower speed limits. It is widely believed that many such roads have limits set well below what is appropriate and drivers speed on them anyway.

Further, most speed-related crashes occur when someone is driving much faster or much slower than ambient, speed limits notwithstanding. A driver assistance system slavishly obeying limits when other cars do not is far more of a hazard than a car going over the limit but keeping up with traffic.

The IIHS represents an interest group that stands to lose big from self-driving: auto insurers.

Edited by Fighteer on Mar 17th 2021 at 5:13:23 AM

CleanTechnica is on a tear: Hey, NTSB, Where’s The Evidence Tesla’s FSD Beta Is Dangerous?

It's worth reminding everyone that this is an opinion piece, although frankly it looks like the NTSB is writing opinion pieces on the potential hazards of Tesla's FSD Beta without any evidence of actual harm, a level of attention that nobody but Tesla receives.

• The software was released to an extremely limited pool of testers who have been actively monitored to ensure they're using it properly and dropped if they are not.
• Tesla has not advertised FSD Beta as anything other than a Level 2 solution.
• There have been no reported accidents or safety incidents related to vehicles using FSD Beta.

If anything, the extreme caution with which Tesla is releasing these features should be a mark in its favor.

CleanTechnica also offers the opinion that the NTSB seems to be requiring perfect safety of autonomous solutions when (a) millions of people die every year thanks to human drivers, (b) perfect safety is not just unattainable but undesirable at a fundamental level, (c) informed consumers may choose what level of safety to accept as long as they aren't causing undue risk to others.