Self-Driving Cars

This is a natural problem you get when you sell an early-access product, judgments are made based on the product at the time of judgment and aren’t going to be updated whenever the product updates itself.

There's something to be said for doing the same thing but making sure that you keep using a different name.

And never fall into the insane trap that for some reason the USB and HDMI specs have fallen into, where not only do they have different versions, they have at one point decided to rename existing versions. USB 3.0 is also something like USB 3.1 Gen 1 and USB 3.2 Gen 1 (with zero changes), and HDMI 2.0 became... HDMI 2.1? And the new 2.1 features that were expected were made optional. Or they made the 2.1 spec into the 2.0 spec but just declared the new stuff optional. Because that's not confusing at all.

It's possible, and again I caution I haven't watched the documentary, that they are making the mistake of thinking that Teslas are like other vehicles: when you buy one, the software is exactly what it will always be for the lifetime of the car unless there's a recall, you take it to the dealer, yadda yadda. The idea of a car downloading an update overnight and immediately becoming better at things is so alien to them that they can't understand it.

Or they are perfectly aware and are sensationalizing things. You decide.

Anyway, I don't normally watch FSD videos because, frankly, they're mostly just people talking while a car drives, but I was curious enough about the 10.12.1 release to wade in, and here is a relatively short one (eight minutes) that highlights just how far it's come and shows some of the remaining weaknesses.

First, the visualizations are tight — way better than they used to be. The visual field is steady as a rock when it used to jump around quite a bit, the road lines are sharp, cars animate smoothly, and so on. It is much more confident in remembering information across time. At no point do I get the sense that it's not seeing something that it should.

You can still see the occasional flicker as it's misclassifying something at the very edge of its visual field or something that is occluded, but that doesn't seem to affect its behavior. It almost feels human in this way: is that thing that I can just see the corner of behind the car another car, a mailbox, a shopping cart, or what?

In terms of route planning, there was one absolutely brilliant moment when the car successfully drove around a stopped Amazon delivery van on a two-lane road. It yielded to oncoming cars, and then just when I thought it might get cold feet because another car was approaching, it very confidently asserted itself and took the planned path. That's an incredibly complex behavior that many humans get wrong.

Where pedestrians are involved, I can clearly tell that the car is thinking about what they're going to do. In its "mind", it's trying to predict where each person is going to walk, bike, etc. The visualization highlights pedestrians that the car is "worried" about: "this woman might try to cross in front of me". At no point did I see it fail to identify a pedestrian.

And here we come to the one real weakness in the drive I watched: it is still very cautious around people. Now, you may consider this a strength, and that's fair, but it can be frustrating to other drivers when the FSD car waits for an unreasonably long time for someone that it thinks might try to cross the road, or doesn't enter an intersection because it's waiting for people to clear the crosswalk.

I also saw it make an odd lane decision: moving to the left lane before merging right again to make a turn. This may just be a planning thing: it wants to go into the fast lane but doesn't consider at that point that it will need to be in the other one in less than 30 seconds.

What is most obvious to me is that FSD Beta is not just driving by following a set of pre-programmed instructions; it is reacting to its environment in an increasingly humanlike way.

Edited by Fighteer on May 21st 2022 at 9:20:10 AM

Yeah. That scene where it stopped to buy a cup of coffee and go to the bathroom was really creepy.

@Fredhot: Here a short primer on the role of AI in autonomous vehicles. It covers types of sensors, data processing, path planning and execution, even collecting data for insurance purposes.

Edited by DeMarquis on May 21st 2022 at 9:47:11 AM

You jest, but when they're driving themselves to automated Supercharger stations, the robot revolution will be ready to begin.

~raziel365: Replying to this post in the Sustainable Energy thread, to keep things on-topic:

First, I think you're dramatically overestimating the number of antagonistic situations of that nature. We don't live in a Mad Max society where highway bandits are hijacking cars left and right. Those cars will have humans in them, who will be able to respond to such incidents by calling for help or using a panic button in the vehicle, which I'm going to assume will be a standard feature.

Autonomous vehicles without occupants will be in communication with a fleet control system and would similarly be able to report incidents, such as crashes, aggressive or antagonistic behavior from other vehicles, unexpected stops, and so on.

Second, I think you are underestimating the capabilities of a fully autonomous vehicle. Teslas with FSD Beta are already capable of evading dangerous situations, maneuvering around stopped cars, and so on. They will only get better at it over time. (They can also recognize and distinguish police vehicles from other vehicles, so should be able to respond correctly to pull-overs.)

The approach that Tesla is taking is to create a generalized solution for machine vision. If this solution space does not include antagonistic driving from other vehicles, then they can train it for that and, hey presto, the cars can evade those sorts of situations! At the very least, they can recognize them and signal for help, either from the occupant(s) or from a fleet controller, as stated above.

If I can think of these things, so can the developers of autonomy. You aren't going to blindside them with a scenario they haven't envisioned. It may just not be something they've had time to work on yet. For example, people have shown that it's possible to mess with road signs to make a Tesla drive at the wrong speed, but all this is doing is training the software to understand when that is happening and not to fall for it again.

Addendum: Autonomous vehicles are capable (by necessity) of recording their surroundings with a high degree of fidelity and resolution. Anyone attempting to mess with one would be on full Candid Camera display for the authorities. The YouTube channel Wham Baam Teslacam posts weekly videos that include situations in which a perpetrator was ID'ed or arrested thanks to these recordings. They also clearly establish fault in most situations, acting as built-in dashcams with a 360-degree field of view.

In fact, I predict that, aside from those with very specific motivations, most people with malicious intent are going to learn to stop bothering autonomous vehicles because they will get caught and they will be held liable thanks to incontrovertible video evidence.

Edited by Fighteer on Jun 2nd 2022 at 8:04:10 AM

A multi-vehicle robber assault feels like something a driver of a regular car couldn’t do squat about either.

As has been mention before, a self-driving car should be equal to or better than a human, not perfect.

It’s the same with the fake signs thing, that’d work fine on most human drivers, so why is it suddenly a worry only now that a robot will be being tricked rather than a human being tricked?

I think a more likely issue is GPS problems. I assume the cars are designed to not like...drive into a lake when the GPS thinks there's a road there, but some GPS maps unintentionally ported in the fake roads used for copyright checking and they can go for some real strange and circuitous routes when they think the direct route is non-viable for some reason.

Also, what happens in places where GPS is very unreliable. When I lived in Switzerland, the GPS had a hell of a time trying to figure out some of the mountain roads, especially the really hairpin turn heavy ones.

I assume tunnels are easy (just keep...following the road until a signal is re-established), but what happens if the car is receiving conflicting signals that tell it that it just made a snap u-turn somehow and has now reversed direction...and the visual evidence also shows an extremely sharp turn?

(This is not a "why we shouldn't do this" thing, this is a "these are actual problems I've run into, how are they being handled" thing)

Edited by Zendervai on Jun 3rd 2022 at 8:56:08 AM

The answer to this one is fairly straightforward: self-driving cars do not use GPS in the way you are suggesting, and this applies to both of the broad classes: lidar and machine vision.

In a lidar-based system, the car is loaded with high-definition 3D maps and uses its lidar sensors to understand its position. GPS is used for redundancy but is not necessary for driving. If the 3D maps do not correspond to what it sees, it may ask its driver or controller for assistance, but won't drive itself into a lake.

In a machine-vision system, the car uses its cameras to understand its environment and recognize where the roads are. GPS is used for route planning but cannot contradict what the car sees. If GPS says a route exists but the car sees that there is no driving surface available, it will either attempt an alternate route or, as with lidar, ask for help. For example, Tesla FSD Beta is already able to re-route if it encounters a road closure or obstruction along its planned route, and if you don't give it instructions, it will follow whatever road it's currently on.

In both cases, vehicles do not need an active external signal (cellular, GPS, WiFi) to operate on a moment-to-moment level: staying on a road and not running into things. However, a lidar-based car does need to have a high-speed data connection to download updated maps. I don't know how frequently this occurs, but I would expect it could handle a temporary blackout from going into a tunnel. A machine-vision car remembers its planned route and resyncs once it recovers a GPS signal.

Edited by Fighteer on Jun 3rd 2022 at 9:18:48 AM

NPR: California has given the green light to GM-owned Cruise to start charging passengers for use of robotic taxi service without any human drivers in the vehicle. For now, the taxis will only be allowed to run in less congested parts of San Francisco, between 10 PM and 6 AM, and only if there's no heavy rain or fog.

The Guardian: Study by Emissions Analytics indicates that car tires are a major source of micro- and nanoparticle pollution, including carcinogens.

Things like this are why I'm cynical about cars. Even if you take out the internal combustion engine, even if you take out driver error, there are still a lot of problems with cars, and they shouldn't form the cornerstone of our transportation system going forward.

Okay, but what does that have to do with self-driving technology, which can also be applied to trucks, buses, and trains? Telling people in developed countries to give up their cars is just a fool's errand, and using it as a reason not to improve cars is almost a textbook case of the perfect solution fallacy.

Edited by Fighteer on Jun 3rd 2022 at 8:26:32 AM

This is literally the "Self-Driving Cars" thread.

Also, nice strawman. I said cars shouldn't be the cornerstone of transportation. I'm not demanding people "give up" their cars.

Okay, but why the self-driving cars thread and not, say, the automobiles thread? Or one of the more general environmental threads?

It's not like self-driving cars have a wheel monopoly.

Earlier we were talking about reliance on GPS for self-driving vehicles. Well, it seems like the gods of timing are looking down on us, because Elon Musk just tweeted a helping of information about Tesla FSD Beta's upcoming features.

• [1] "10.13 smooths out intersection control, especially long lefts, and starts to handle roads with no map data at all. Last point is a big deal. Within a few months, FSD should be able to drive to a GPS point with zero map data."
• [2] "[C]ar will navigate to a pin location, even if in a complex surface parking lot or hotel entrance. When in covered or underground parking lots, car will have to navigate using only inertial measurement, wheel movement & vision, as GPS signal is no longer available."

So I was slightly wrong and slightly right. I was wrong in that FSD Beta currently does rely on maintaining a GPS signal. I was right in that it is soon going to be able to fully navigate without one as long as it gets an initial route plan from GPS.

Also, Musk tweeted today that Tesla is "close to the point where offering something for EU regulators to review makes sense". The EU has a much tougher regulatory environment for autonomy than the United States, so getting approval there is a big deal.

Note that Mercedes' Drive Pilot is currently qualified for Level 3 use on the Autobahn, but only under very limited conditions. (The tweet I linked has a typo and calls it level 4.) The driver can be hands-free as long as weather conditions are good, visibility is good, there is a lead car, and speeds do not exceed 60 km/h... on the Autobahn. So really it is more like a traffic jam navigator than a fully capable self-driving solution.

Edited by Fighteer on Jun 4th 2022 at 6:58:09 AM

I asked a couple people I know from when I lived in Switzerland and I found out why I hadn't heard anything about issues in places like the mountains in Switzerland. People living or driving in those areas on a regular basis are heavily discouraged from getting electric vehicles. Like, those are the areas where you need to keep tire chains around in case of heavy snowfall. You need both the really powerful engines and the weight of the engine to make it through safely. You also need vehicles that aren't impacted by the weather in terms of range and that you can carry around extra gas or whatever to keep the engine running if you get stuck.

I also get the impression that those areas might ban the use of self-driving for cars, long-term, because snow driving with chains controls weird compared to normal driving or even normal snow driving. But in the long-term, those problems will likely either be solved or just be a sort of small aberration in regions where electric cars really don't have great use cases.

My friend didn't mention the regulatory thing, but that'd almost certainly be the best explanation. No one's tested self-driving in that area.

I do have to admit, I'm...concerned about self-driving cars in heavy snowfall or in conditions of freezing rain. It kinda makes everything erratic and difficult to account for.

Edited by Zendervai on Jun 4th 2022 at 7:06:13 AM

I understand the concerns, but to the software it's all just data. There's no obvious reason why it can't be trained to recognize hazardous conditions like snow and ice, and in a scenario where traction is poor, the car will be able to react much faster and more accurately than all but the best-trained human drivers.

The point is that it's not about absolutes, like there's a hard line between where a self-driving car can go and where it cannot. Rather, the space of driving conditions that autonomy is capable of handling will gradually get filled out, from easy to hard, until it's complete.

If the car has a private owner, then putting on snow tires or snow chains would be equally as important as it is now, and if the car is in fleet use, its operator would be responsible for that.

Edited by Fighteer on Jun 4th 2022 at 7:12:47 AM

Here's a Twitter video of Tesla FSD Beta v10.12.2 navigating an unmapped roundabout in a construction zone. The car hesitates for quite a while, apparently unsure of what to do, and then it makes a very human-like maneuver by following a pickup truck that is navigating the same space.

If one were to anthropomorphize this, one can picture the thought process. "I'm not sure what to do here. It's not on the map, and it's a very complex driving space. I don't want to make any mistakes, since my priority is not crashing. Oh, look, there's someone who looks like they know what they're doing; let's follow them."

Notably, the car does run up on the curb a little bit, but it's a sloped curb like you'd find on a racetrack so there's no risk of damage. It also feels like it's going a little fast, but that may just be the perspective of the video.

This is one of the scenarios that we were just talking about: how does an autonomous car handle a situation that it doesn't have any map data for? I'm not saying that the Tesla did a perfect job, but it's not any worse than I'd expect from an inexperienced human driver.

Later in the video it slows down for crosswalks even though nobody is in them. This clearly shows that the software recognizes the crosswalks, and it may be programmed to display an excess of caution in these situations. That's probably better than ignoring them, but the behavior will definitely need to be tweaked over time.

Edited by Fighteer on Jun 5th 2022 at 1:35:52 PM

Re the video: I believe that's the speed in the upper left hand corner of the control display. If it's MPH, then it never went much above 30.

It feels faster than I would drive in that situation, but as I said it's really hard to get a feel for whether it was safe from the camera's perspective.

Speed was about 15mph at most through it. Given there was no exit to the right, so it was mostly straight through, that's definitely fine. Even if there was, I don't see why it wouldn't, unless you want the car to overcompensate for the idea someone might not accept that those already on the roundabout have priority.

Yeah, the main reason I was concerned about the speed was that I was thinking the car might be following the speed limit on the GPS-mapped route it was trying to use, which might make it go too fast in the roundabout, but if it went 15 mph throughout, that's probably fine.

My main takeaway is that it did an acceptable, though not perfect job of navigating through a situation that it had no preexisting information about. A lidar-based vehicle would probably not have succeeded.

Edited by Fighteer on Jun 5th 2022 at 7:54:42 AM

Minor item: Multiple Tesla owners are reporting on Twitter that they have started receiving the FSD Beta 10.12.2 download with a Safety Score as low as 93, indicating that the company is making good on Elon Musk's promise to expand the program substantially with this version of the software.

So it seems a little bit suspicious that Autopilot disengages automatically just before collision as if to be more about "Well we're not liable, the driver was in control at time of impact".

Surely a smarter system would warn and disengage earlier to give the driver time to take action.

I have seen it proposed that this is caused by people panicking and aborting any such assistance when they think it's about to crash, which would also lead to the conclusion that the driver was in control at the time. Which would explain the report not having this happen in all cases.

But it would also suggest that the driver wasn't in a position to stop it anyway.