Hacker News Reader

Comment by deepnotderp | Hacker News Reader

The bet on maps is interesting. I'm of the personal opinion that in the long run only machine learning can solve autonomous driving, but both MobilEye and now this company both have invested heavily into maps.

MobilEye makes sense, they have vested financial reasons that dictate what they do, so they can't turn their back on maps which they've been using for such a long time.

On the other hand, starting a company in maps for self driving cars, especially by what appears to be industry veterans, is far more interesting.

View on HN

Cruise (now part of GM) also relies heavily on maps, based on a talk I saw their lead CV guy make. It does make a good deal of sense - maps (at least street-maps) are mostly static and they make the problem way easier, so why not use them. In fact, most people these days are implicitly using maps - how often do people drive to new places without GPS navigation? Of course, vision and likely ML are still a necessary part of the equation.

By maps I mean high resolution 3D maps with lidar and stuff as this startup seems to be doing.

I don't doubt that 2d maps would be needed like how humans drive with a GPS.

> "In fact, most people these days are implicitly using maps - how often do people drive to new places without GPS navigation?"

I'd argue also they're using maps in an even more fundamental way - most driving occurs in places where the driver has already been, and relying on a wetware-stored map is a critical part of safe and effective driving.

Local knowledge - traffic patterns, weird traffic signal setups, difficult lane changes, etc, is pretty critical to driving well.

If I have been to the destination multiple of times, I usually don't open map, I leave my phone off, or I let Pokemon Go running. But sometimes I do use map to seek for alternate faster route if I am in a hurry even if I am familiar with some alternate routes just to be sure.

Since I don't use Waze (horrible experience), Google Map won't tell me there's a camera ahead, or police is ahead.

Right - I may have phrased my point badly. The idea is that all effective driving relies on pre-mapping, either via digital means (your phone) or human means (your brain).

The expectation that self-driving can be perfected without any kind of pre-mapping and the system can behave at high reliably in-situ seems unreasonable. Even humans can't do that.

Humans and computers have different strengths though. Just because humans can't do something good doesn't mean computers can't, and vice-versa.

A map in your head is closer to machine learning than to using GPS.

The betting is not only interesting, but actually an excellent bet, in fact, very strategic. Learning to drive is only part of the formulas. As I stated this probably 10th time on HN, we need real data about road condition in order to make self-driving really successful. DeepMap is only part that solution, but we still need real data in real time. We need a real network communication in the transportation space. Take flight control, instead of having a person coordinating, we can make computer programs to do that coordination; now apply that to driving. This is not only a data challenge (because there is very little authentic data to collect from to begin with), but also a huge challenge in network communication.

The problem with your idea is it's well outside a MVP for self driving cars, and we don't actually have self driving cars yet.

What's MVP? Of course we don't have self drivinv car otherwise we wouldn't be having this discussion.

Minimum Viable Product. He means that until self driving cars are not even as good as human drivers, it is pointless to discuss features that make them even better than human drivers. I.e. "one step at the time".

Well it probably is, but it would be too late if we don't start investing into the data collection. Convincing government to build infrastructure and services what I am suggesting is going to be a long investment. I know some states are working on toward that goal, and IBM's SmartCity project is one such example experimenting how to use technology to allow services to subscribe to data.

We have a MVP. Google has a MVP, Telsa is a MVP already. I feel we focus too much on building sensors.

With good sensors you can quickly get to good maps. But, getting to level 5 automation without good sensors and useful AI is impossible.

Again, I never said we don't need good sensor. I have been saying we need to start thinking outside of just sensor. The forumla can't just rely on sensors alone. I don't dispute we need to be able to recongize surroundings. Because we human drive with eyes open.

We humans also do fine with a our eyes, our brain, and a crappy GPS or Map. If you want to say that's not enough for computers and you need really good maps to get level 5 automation then I disagree. If you think you can be first to market with Level 5 automation while splitting your focus onto irrelevant details, again I disagree.

Your view on self-driving car is different than mine. I am not sure who came up with the five levels of self-driving car automation, but let me reiterate the last time before I give up trying to convince anyone on this thread.

First, for a computer to be able to navigate safely on the road the computer needs to be able to recongize and understand its surroundings. Checked.

Next, for computers to be able to be good at driving, we need to train the computer both in the real world and in simulations.

Moreover, to ensure we can optimize route, plan ahead we need data about road condition far ahead of us (and sometimes behind us). To be able to make turns on an intersection safely and swiftly, unlike humans who are terrible at making turns and confusing when to turn, computers can coordinate with one another. This requires good network communication. Bluetooth is weak and unstable. In my earlier example, if an ambulance says it is driving heading my direction, my car can coordinate how to empty up a lane for the ambluance to pass. This level of automation is what one sees in a sci-fi movie, which is really what you want besides being able to avoid obstacles like Bat mobile.

My point is, we should explore how to make traffic data accessible. There is so much one can do with such data. It takes a good amount of time to get the government onboard and help them build such infrastructure. To finish off my point, meter data is not available, right now apps use crowd source to determine whether a parking spot is available or not. Instead of driving around which causes more congestion and sometimes accident, governement should make meter data available, and we can still crowd source by using driver's camera to recongize whether a spot is opened and send back to datastore.

This is essentially the goal of Internet of Things, at rhe city scale. As I mentioned, IBM's SmartCities supposed to do stuff above but I really am not sure if IBM will ever pull off given many of their projects ended up on someoen's desk covering in dust.

Trafic analysis based on cellphone tracking is already part of free GPS apps like: https://itunes.apple.com/us/app/waze-gps-navigation-maps-soc... Even one user per 100 drivers is plenty to locate traffic issues and we are well past that. Now sure, the situation is improvable but the reality is knowing about bad traffic is more or less solved.

For higher precision tragic flow that's going to fail while their is still a high percentage of human drivers. So sure when there are 50+% AI drivers it's a useful optimization. However, for the first decade or so it's not going to happen making it a waste of time for now.

As to finding a parking space, again mostly a non issue for self driving cars, because they don't need to park near people they are better off going to a garage which already tracks open spaces than using street level parking when it's hard to find. Can you slightly optimize this sure, but again it's not a hard problem relative to driving without a person in the car.

Every car could observe its environment, refer to the map, and send back any conflicting information to DeepMap, so that the map database would be updated in real-time, and pushed to all cars in the area.

Not quite. We need data from toll, from government on planned events, getting data about where emergnecy vehicles are going. Obseving environmebt is only as reliable as the recongition. For example if I knew ambluancs is coming from behind I would pull over. But often this is last minute reaction and often we can barely make way for ambluancs. However, if dispatcher publishes signal and ambluance publishes its GPS then cars where the ambluance is heading can get that data in advance and calculate how far and when to pull over. Relying on doppler effect to calculate siren then get twenty cars to slow down is not a great design.

Fleet management will become a true need, so maybe that will turn out to be their main value prop.

Reply on news.ycombinator.com