Ask HN: Why aren’t autonomous cars using signals? | Hacker News Reader

Ask HN: Why aren’t autonomous cars using signals?

19 pointspipeline_peak16 comments

It feels infuriating to think AI and machine learning are becoming the future way of the road. That instead of sound logic, cars will learn important behaviors exhaustively in the name of safety.

How about instead of 1000 cars on the road guessing and sharing “neural information” we have all manufacturers agree to make future cars equipped with a standard communication signaling system. One that describes turning, speed, dimensions, all necessities for safety. Then older cars can be the AI fallback.

View on HN

The FCC allocated a band for V2X communications long ago

https://www.fcc.gov/wireless/bureau-divisions/mobility-divis...

Early on the automotive industry was betting on a version of WiFi modified to reduce connection overhead but this took forever to get deployed. Telecoms would badly like to introduce a system where they mediate V2X communication over the cellular network, which would be good for security, but between the appearance of rent seeking and the fact that telecoms have no interest in upgrading their network to avoid dead spots makes that kind of system dead in the water too.

If you can't get people to stop buying $70,000 gas cars I don't know how you're going to get them to stop buying cars that don't include whatever car-to-car communication system you can come up with. People don't do what's good for the commons; instead they do what they think is good for themselves.

So for a long time going forward, autonomous cars will still have to deal with other cars that are basically rolling bricks. Since they can't count on any smarts or any of the communication that you speak of, they have to rely on themselves.

So eventually, yes. But meanwhile we do need the fallback of autonomous discernment.

Disclosure: I worked at GM/Cruise, opinions are my own.

Right now, most cars are human-driven, and this fact will remain true for the foreseeable future. When there is a larger share driven by just computers, a standard will likely emerge, but to build that now would just be a distraction from making the self-driving cars safer in the presence of human drivers.

That's exactly what's being done in Europe, sponsored by EU and multiple national governments. I was a contractor on a related software project.

https://www.audi.com/en/innovation/autonomous-driving/car-to...

Companies in the US don't like open protocols or standards, and will push their own proprietary systems in the hope that the market adopts it so it can be licensed to everyone else that wants to use it.

Ideally, cars that need to drive on roads with human operators should be depending on their surrounding and conditions, not some secondary signals about those surroundings and conditions. As long as roads go through places that people live and work in, humans will be on the roads as drivers, pedestrians, cyclists, etc.

The fallback has to work reliably, certainly for the foreseeable future, when most cars won’t talk or listen to other cars, but probably also if all cars share the information, to handle the case where some car stops broadcasting due to a bug or hardware breakage, and if your fallback works reliably, you don’t need anything else.

Because of that, there, for now, is little benefit w.r.t. safety in this (there may be benefits in efficiency, for example if enough tell the traffic system what they’re up to, so that the traffic system can optimize traffic lights)

Also, I hope (but I know I’m wrong in that) nobody is going to say cyclists and pedestrians should do this, too. You shouldn’t require people to gear up with electronics before they can go on the street (even if mobile phones were enough to do that), so the fallback will have to be there forever.

There also is work being done on this. For example, the EU has projects in this space (https://www.connectedautomateddriving.eu/)

OP you have a good point that I've thought about often. However, instead of digital communication, what cars lack is optical information about themselves projected on the rear and front corners with speed, acceleration, and other metrics. This is easy to retrofit on 'dumb' cars with cheap LEDs, it will provide value to human drivers (as long as the format is standardized, intuitive and easy to read), and it's also valuable for camera-based AI driving, which all forms of driverless cars use.

To me, this is something I wish every car would have, so I could tell at a glance if someone is accelerating/braking and their current speed.

Perhaps this could work for highway driving as a first step. Mandate all car be “on-network” to use certain highways or lanes.

Car-to-car communication would allow cars to run in dense packs, inches apart, increasing efficiency through the use of drafting. Except for controlled pedestrian crossings, they would never need to use traffic lights or stop for other vehicles: they would flow past one another smooth as butter.

Until the bug that causes them to crash in each other.

That likelihood is no more than any other form of autonomous driving.

Because sometimes people don't drive cars. Unless you want to ban all pedestrians and cyclists from roadways, cars will always need to interact with humans.

A (very) small step was done by Scania's automated truck platooning, a driver in a truck leads a small convoy of driverless trucks, they communicate with each other using a variant of 802.11 WiFI

https://www.scania.com/group/en/home/newsroom/news/2018/auto...

this all exists. dsrc for v2v plus the basic safety message, etc

If it's V2V you have the potential for malicious actors to send signals not actually describing what the car is doing. Visual/radar/lidar and AI interpretation thereof are needed as a fallback to prevent crashes. If it's not V2V you need mediation, authentication, and extremely high availability, which has antitrust xor lack of non-telco infrastructure. Cars need visual/radar/lidar and AI interpretation thereof as a fallback to prevent crashes just because a cell tower goes down, and will be completely inoperative until they go up in the first place. Human drivers change their minds all the time, and represent the vast majority of drivers right now. Humans don't even do a good job of following their GPS directions, let alone using turn signals. So visual/radar... you know the rest of this sentence.

Basically, the ground truth is always more important than whatever signals are being sent. If a malicious actor can't fool a human, it's mustn't fool a non-human driver. If missing or unavailable infrastructure can't make a human unable to drive, it must not make the non-human driver unable to drive. Decreased efficiency, okay. Slower trips, fine. But total denial of service (either by failure to operate or crashing) is unacceptable. All of it eventually relies on the AI fallback. So we should be (and are) building the AI fallback first. Every car will have the AI fallback, including V2X-equipped cars, and all V2X signals will be treated as potentially unreliable forever.

I don't doubt we can have some efficiency gains from V2X, but we'll likely never have bumper-to-bumper at 60 mph, interleaving though intersections, etc. Cars just aren't mechanically reliable enough. Say someone drops a nail or caltrop or cylinder of compressed gas out of car #0's window. How many milliseconds does car #2 have to react to the pending tire blowout of car #1 immediately behind? Even with an AI fallback, its sensors might miss the nail. Is a sudden pressure drop (presumably signaled) going to be transmitted and processed fast enough by the entire convoy of cars to avoid a large multicar pileup? Are ALL the cars' brakes in perfect working order? And that's not just a linear problem - all the cars in all the other lanes will have to brake too to allow car #1 to swerve unpredictably as the tire explodes as the car learns the new driving model it has to contend with (and it will need the fallback AI to figure out how its inputs are translating into real outputs). Not that this is an easy situation for human drivers to deal with now! But now we have 1-4 seconds (de facto-de jure) of following distance. V2X comms could make 1 second following distances safer. But it'll need an intelligent (artificial or human) fallback anyway, and having a human one means that the efficiency gains will disappear. So we need the AI fallback.

It's hard to ignore the fact that signals can be dangerous, especially when they are a distraction while you're driving.

However, autonomous cars are not making it easy for us to navigate the roads safely.

Let's take an example. Say you're driving along, and all of a sudden the traffic lights turn red. You don't want to stop, because you're in a hurry—but you also don't want to be caught in a situation where you're ignoring red lights and being pulled over by law enforcement. The problem is that both options will end badly for you: you'll end up getting in trouble with the cops or having a near-accident with other drivers.

So what do you do? Well, if you were using an autonomous car, using your signal would probably be your best option—but this could actually lead to more accidents.

Remember that self-driving cars are programmed not to react to things that aren't important to them—and signals, which are a vital part of driving, can be dangerous distractions in addition to being an obstruction on the road that could cause accidents with other drivers or pedestrians. So even though we know signals can be dangerous, we still need them.

Reply on news.ycombinator.com