Tesla’s new ‘light-years ahead’ navigation and maps engine is ‘almost done’ | Hacker News Reader

Tesla’s new ‘light-years ahead’ navigation and maps engine is ‘almost done’

electrek.co

37 pointsevo_922 comments

There was another Tesla software update announced as "almost done": a major update to Autopilot 2, with an all new neural net and what not.

This all new navigation system might include functionality to help, not only the human driver, but that new Autopilot as well.

At some point Autopilot and Navigation will have to merge. Right now, both systems operate completely separate.

But if those upcoming updates will allow Autopilot to follow the navigation system changing lanes thru highway intersections, than that would indeed be 'light-years ahead'.

"Almost Done" . -- Elon Musk suffers from the typical engineer's optimistic feature task/project estimation.

Some people have wrongly attributed his overpromising, and delivering late, as Elon being a salesman, while the more obvious explanation is that he is just an optimistic engineer that tends to miss deadlines because 'estimating is hard'.

We all have been there

My expectation is that when it comes it, it will kick ass, but it will be late, always a bit late, like most of our projects...

Looks more like "minor change to map GUI". "The main difference seems that it enables a significantly more important zoom and smoother engine."

While the light-years-ahead would imply compared to current available navigation systems such as Google Maps, Apple Maps, etc, the actual tweet in the article notes he's saying it's light years ahead of their own current system. So doesn't necessarily mean it will be as good as Google Maps nor any promise it will be better and certainly not light years ahead of Google Maps or Apple Maps presumably.

who gives a damn about maps, just use your phone. Sick of their false tweet timelines. FFS it took 14 months to "learn rain" and enable wipers...

Autonomous vehicles need HD maps to be able to localize themselves accurately on a centimeter scale, GPS and current maps do not allow that.

The higher the resolution of your map is the better and faster the localization is, better localization means better and more accurate decision making for your car.

To illustrate it simply we’ve all been at a point where the GPS app we use put us on the curb or on a different lane sometimes even in incoming traffic.

This is because your phone localized you erroneously. While we ignore it or make a joke about it if this happens to an autonomous vehicle someone can get hurt or worse.

HD maps are the new gold rush in autonomous vehicles and companies are rushing to find ways to cheaply produce, update and more importantly distribute them to vehicles.

About two months ago I opened google maps on my phone and departed for a neighbouring country. I took it in offline mode (am still weary of EU-wide roaming) and had a route going on and everything. Half-way on my route, I missed a turn and thought no big deal, I'll turn around and continue from there. Google maps had different idea. It turned off route completely and congratulated me on arriving to my destination, while I was still half-way through the route. There I was, god knows where on my route, with google maps mocking me. I finished my trip by following signs, ye olde way.

On my way back, I turned on wifi and dialed in route for going back, while specifically telling it to avoid tolls (for several reasons - scenic and because it was not really a long trip to my country so I wanted to avoid tolls which make you pay for 7-day vignette at minimum). Anyways, route was on and I departed. 5 minutes into the trip, there I was in-front of the tollbooth. Thanks, Google Maps. Never again for anything outside of city driving (traffic info is great). I can't imagine handing out my life to this crap guiding the autonomous vehicle.

Here Maps is great for offline, unlike Google Maps you can search for POIs and will generate routes offline.

Current maps might not, but afaik Galileo has (will have?) centimeter accuracy by default and sub-centimeter accuracy when enhanced.

That isn’t a correct way of thinking about this problem and in more than one way.

The first reason is that you want your car to be able to localize it self with its onboard sensors alone, this is always faster and more resilient than using GPS.

GPS would put you in the ballpark and your localization goes from there.

Second GPS and any other simmilar positioning system is only as good as your maps because they give you coordinates if your map isn’t accurate or the resolution of your map is lower than the resolution of your positioning system then you again lose localization accuracy.

Galileo and GPS offer about the same resolution however high precision is costly (both in time and compute power) for both systems and you still need HD maps to be able to localize yourself on a centimeter scale.

Snap to lane and quality learned vectors is good to compensate for bad GPS signals.

For lane snapping to work those lanes need to be in the map to begin with and this is where HD maps come into play since they are on a centimeter scale.

There are a lot of way to compensate or correct for many things however your autonomous vehicle does not and should not rely on GPS or any other positioning system for decisions it's just one source of information which it uses for localization.

You have your onboard base map, you have the map that the car constructs using it's sensors, you have GPS data, your car might also be downloading portion of updated HD maps say 10 by 10 KM grids as you go, data that it might get from other vehicles all of this is used for localization and decision making.

There is nothing wrong with GPS and modern GPS chips like the ones from broadcom are capable of centimeter accuracy (at highway speed it's debatable but still) however you still need a high definition map to take advantage of that because if each pixel on your map is 2 by 2 meters then localization yourself to 10cm won't be possible regardless of which positioning system you use.

Nobody needs centimetre accurate navigation.

I'm pretty sure maps are important because having good maps and routing data is a prerequisite for better autopilot and eventually full self driving. I'm still hoping the $3K I paid for FSD will pay off...

They use Google maps today already, so just another move to cut costs. They can make it pretty but data can't be better than Google's whose waymo.com has real FSD trials happening now.

Waymo appears to be ahead, but Tesla has little incentive at this moment to publicize anything partially complete.

Waymo is certainly leading the pack on published capabilities, but Tesla has another advantage: tens of thousands of vehicles shipped with a sensor package they claim can support self driving. I’ve heard rumors they may need to upgrade the compute on those vehicles to support full autonomy, but no other company on Earth is better equipped to activate a fleet of tens of thousands of self driving cars (if the sensor package is indeed sufficient).

The kind of do and don't use Google maps. The center display shows Google maps, however the instrument cluster map and nav system uses an on-board traditional GPS system (like the tom-tom or garmin systems) due to them not being allowed to use Google for routing. The differences in the maps are sometimes jarring. Eg, the Nav will be telling you to turn onto "unnamed road" when the center console clearly displays the name.

Technically Tesla has far more vehicles on the road than Google does, and if they came up with a way to use their cars to capture data they'd have better maps than Google for major cities in the course of a single weekend.

The problem is bandwidth for upload, but if SpaceX launches internet satellites the Tesla cars can get free access to...

I don't think b/w is a problem. They are capturing different large chunks of video data already for improving self driving, and uploading via owner's wifi.

Some owners with rooted cars have reported large video uploads, and have noted that Tesla seems to have a "trigger" system that uploads camera data based on certain conditions. Eg, "it was raining and driver reclaimed control from AP in a curve"..

I don't know about you, but I'd be pretty cheesed off if my car blew my bandwidth cap because it was secretly uploading terabytes of mapping information over the course of a month.

If it was doing it via some other method, like via satellite, I wouldn't be concerned. Little car, do your thing!

I suspect they're doing that already. I don't think the bandwidth is an issue, unless they plan on reporting full frames of visual data (ie pixel maps) rather than vectorized or telemetry data.

They can always use a store & forward approach if they want to limit their interface usage to WiFi.

For real time data (like traffic conditions), even their 3G network should be sufficient (remember that they are already streaming music over it).

Lightnyears ahead of current verwion but jo vetter than google maps

Reply on news.ycombinator.com