Hacker News Reader

As someone who's old enough to remember road trips armed with nothing more than a Rand McNally road atlas for navigation (really not that long ago - and I do continue to take one along nowadays as a backup), I'll never trust an autonomous vehicle that needs to be fed a constant stream of up-to-the-minute mapping data in order to make decisions.

I am with you here. Autonomous vehicle may need GPS for navigation, but it should be able to drive down the roads without it.

Definitely. GPS can't be relied on: there are lots of places with naturally bad reception and the military sometimes tests GPS jammers, like during the recent exercise.

It's interesting to note that the first in-car navigation system used no GPS: it relied solely on inertial navigation and map-fitting to determine its location. For navigation, an autonomous car needs something like that as backup for GPS. All the actual driving needs to be done with on-board sensors, though.

Dead reckoning is actually fairly accurate if done right.

I'd be surprised if the lidar units aren't being used for a similar algorithm.

Doing dead reckoning "right" over an extended distance requires extremely expensive laser ring gyroscopes. It's going to be a long time before those (or something with similar accuracy) will be affordable for consumer use.

Would a self-driving car need perfect dead reckoning? I was able to update my assumptions about my distance and location based on street signs before I had GPS...

Couldn't the device correct using local land marks? A minimal offline map, maybe even just topographic, that shouldn't change too often. "Expected hill was crested 5 meters sooner then expected, update estimated position accordingly".

If you have local land marks then you don't really need dead reckoning at all. Cheap dead reckoning sensors have so much error and drift right now as to be mostly useless for driving. The expensive sensors work but aren't economical for cheap passenger cars.

Here's the system I was talking about. It used dead reckoning and landmarks (turns) to periodically re-calibrate the fix. This was all done with 80s technology.

https://www.fastcompany.com/3047828/who-needs-gps-the-forgot...

Live demo: https://www.youtube.com/watch?v=CHCCjlSWbHE

> Cheap dead reckoning sensors have so much error and drift right now as to be mostly useless for driving.

I don't think anyone's saying that it would be used for the driving part (i.e. car control), just as an option for orienting the car in the wider world beyond its direct perception.

GPS is a nice-to-have, but wouldn't be necessary for a true AI, given an autonomous car can have more and better on-board sensors than your basic-issue human (we don't have eyes in the backs of our heads, or LIDAR).

Why would a True AI (tm) need a car in the first place? "Give me your clothes your boots and your motorcycle" and all that.

I don't see how GPS is a "nice to have". How does an entity know to get from A to B without some sort of navigation assistance?

Unless your "true-AI" stops to ask locals for directions, its going to need some sort of GPS. Can't say I'm looking forward to arguing with my car about how it should have taken a left on Main St, instead of a right.

> I don't see how GPS is a "nice to have". How does an entity know to get from A to B without some sort of navigation assistance?

GPS ≠ navigation assistance. You can infer where you are on a conceptual map from reading your surroundings, especially signage. Getting a latitude/longitude estimate from the GPS network and trying to match it with an overly precise and possibly out of date map isn't the only way to infer your current position. It's only useful when you don't already know where you are, and computers are good at precisely remembering the path they took from the last known observed reference.

Plus, self-driving cars need to be aware of their surroundings, and not solely rely on blindly following a hyper-precise map as a prescription of the road details ahead. Having the local observations become the canonical precision knowledge, and keeping the map as conceptual knowledge of how roads connect places, implicitly gains a lot of this sort of benefit.

I mean rear-view and side mirrors are the eyes in the backs of our heads, though a level 5 self-driving car would be as you say.

Brains pretty much need a constant stream of up-to-the-minute data as well: eyes and ears, both of which can fail.

That's not what he's saying. He's saying a stream of data from a server rather than directly from sensors.

Ah. Well, they don't need a stream do they? Why not periodic downloads of the map area 10 miles in every direction?

Safety is not a boolean value, you have to look at this from a risk perspective.

If you have a self-driving car that's pretty safe, but you can make it even safer with maps, and those maps are going to be accurate 99.9% of the time, the 0.1% of the time will be less safe, but overall you've significantly reduced your risk of an accident, and maybe you can even mitigate the risk by doing things like slowing down when your maps aren't accurate, the same way human drivers slow down when there are cones on the street.

The issue isn't so much that they "couldn't handle" some debris, but that "handling" something is not a yes/no proposition.

Having said that, I haven't worked for these groups, so I don't really know what these maps are used for, maybe these maps are just used for localization because GPS isn't accurate enough.

I understand all that; but in terms of liability, safety is a binary issue. There will be tolerances for safety that can't be lessened because there's a tire in the road that wasn't mapped yet. "It wasn't mapped yet" won't be acceptable, so the vehicles will need to maintain a lowest-common-denominator safety profile at all times based upon their sensory input - not upon networked debris information.

The debris type information should only be useful in terms of routing.

I don't think anyone outside the autonomous groups really knows how map dependent they are.

It's easy to imagine reasons to have detailed maps in the areas they are operating even if their systems are not wholly dependent on the maps. Stuff like being able to more quickly analyze how well the system is characterizing the environment or reducing power consumption or the map is necessary above a certain speed or …

GM made a detailed map of limited access highways in North America just to geofence the super cruise control in the new Cadillacs, so I'm also not sure the duplication of effort matters all that much.

> you wouldn't expect that they would rely on anything too precise for safety reasons alone

I don't think they do. It's just more data that gets mixed in with multiple streams of real-time sensor data.

> Seriously, this mapping thing smells like a boondoggle.

It makes sense to me that autonomous cars would simultaneously be map users and map creators because I assume that data is shared with other vehicles. It's going to drive the cost of the map making way down. I think eventually the company with the biggest fleet will have the best maps, perhaps even real-time. Waze does this now but it relies on user reporting.

You need the map because you want to know where to go, not in order to avoid debris.

Pick two points in your town: say your home and your office. Imagine a stranger had to get from one to the other, and they know how to drive. The map if for all the information that stranger does not have.

Why do you need anything with more precision that current-day Google maps for that?

That is a different question about how often they should be updated, what sort of updates, when does construction happen etc.

You don't need a map for that, just a sense of direction and distance. I have often navigated someplace by making seemingly random turns - I knew where I was going was southeast so any east or south road was fine. Once I got close street signs got me the rest of the way.

> You don't need a map for that, just a sense of direction and distance.

Without map you also need time. In some cases, a lot of it. Not all cities have perfect grid structure and there are one-way streets that can mess you up a lot. In the city where I live it would be really time consuming, because there's a huge pedestrian-only area in the center and many the streets around it are one way, or turning is prohibited at some crossroads so you cannot suddenly decide that you made a mistake and you want to make a U-turn and go back.

Without maps you need time or luck.

That works in cities that are perfect grids. Try navigating the scottish highlands like that. You might have gone miles down a road before you realise that it's not going the way you want to go.

If you are from the US you may be used to use cardinzl points. I was often given directions like "the south side of the building" or "go north east" when visiting the US.

This is not common at all in Europe. Sure, we have a general idea of the direction ("it's with of Paris") but we would not use that for anything else. Especially the building case was weird to me, I woukd have said "the side chose to the entrance" rather than "SE"

> ... just a sense of direction and distance

OK, how does a first time visitor with a rental car get a sense of direction and distance without a map or driving directions from a local?

How do you reliably solve the beginner's mind problem for a self-driving computer?

Without a baseline input, the car will have to rely on the passengers to do back-seat driving which defeats the point of driverless.