Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

FSD rewrite will go out on Oct 20 to limited beta

This site may earn commission on affiliate links.
We will decide whether Tesla uses HD maps or not.

I'm surprised that there is even an argument against using HD maps when/if available for L2 driving, and geofence an L3+ system to only areas that have HD maps that are constantly validated for correctness.

I've encounter numerous driving situations in my life where visual perception of where I'm supposed to be either failed me or failed another person I had to avoid.

Even for human only driving having augmented reality with HD map provided lane marking along with arrows of where to go would improve safety a lot.

Now this doesn't mean vision based navigation isn't also essential. You really need both running parallel to cross check each other, and its very much how humans drive. We have the equivalent of an HD map in the back of our brains if we've traveled on the same road in recent memory. We generally know if something has changed, and we likely can predict how other drivers might mess up.

I would not get in an automated vehicle that did not have HD maps as redundancy to vision.

My desire would be for live HD maps where the maps were updated with visual information from the fleet of autonomous vehicles on the road. So what one car saw the other car would also see even if it wasn't the car that actually saw it. Now it would be only cars near me talking to my car through some kind of V2X standard.
 
Reliance on HD maps (to the degree someone like Waymo uses them) makes scaling a general self driving solution basically impossible.

See also how Waymo after 10 years of work offers consumer-facing robo taxis in ONE tiny geofenced area with perfect weather and no where else- and STILL gets into situations requiring remote intervention.

But then you get to argue about what "HD" really means for maps, and what "reliance" does too.
 
makes scaling a general self driving solution basically impossible.

My interpretation of reliance is really redundancy. It is 100% reliance because redundancy demands it, and not because vision can't take you most of the way.

Redundancy drives the interpretation of what is HD maps as well. Can the maps be used given fairly precise location estimates (based on Vision + GPS) when vision is information is lacking? If so I would interpret them as HD.

General self driving is currently impossible because not only are we asking a computer to do what a human does, but we're expecting it to do so at a safety threshold 10x better than a human. I use the 10x because the 2x threshold Tesla uses likely won't be accepted.

The way I see it is Waymo will continue to slowly roll out their L4 system to more areas where their operation area grows. Where they'll push for massive safety thresholds.

Tesla will hopefully act as a wrecking ball to lower expectations for safety to something more achievable between 2x and 10x.

Now that Biden is going to be president hopefully will have a better national level framework to establish standards to more quickly rollout self-driving. Along with a different perceptive at the nation level we have three things really pushing self-driving forwards: There is Waymo/Cruise pushing geofenced L4 systems, there is Honda among others pushing limited L3 systems, and Tesla with their "lets just make everyone into safety drivers" L2 FSD approach that might or might not happen in 2021.

Ultimately everything comes down to safety thresholds, and they're quite high for most companies as we live in a time where any death is viewed as unacceptable.
 
General self driving is currently impossible because not only are we asking a computer to do what a human does, but we're expecting it to do so at a safety threshold 10x better than a human. I use the 10x because the 2x threshold Tesla uses likely won't be accepted.​

The way I see it is Waymo will continue to slowly roll out their L4 system to more areas where their operation area grows. Where they'll push for massive safety thresholds.
Even the best driver in the world would probably find it impossible to achieve 10x safety. It's impossible to dodge all the other bad drivers out there.
Waymo has a collision rate (real and counterfactual simulation after disengagement) of 1 per 130k miles in their recently published paper. That raw rate doesn't sound better than a human to me but there were no serious injuries real or simulated. They're going to need a lot more data to conclusively prove greater than human safety.
 
Don't think an increase from 4.5GB to 5GB means anything besides half a year's worth of new map data. Increase of 500mb is nothing...
I think the part @boonedocks was referring to was this:
there's a new capability flag set in the newer version.

And this too from @verygreen : https://twitter.com/greentheonly/status/1326592718370459648
What’s the flag?

@greentheonly: It's a bitfield, so I don't really know what it denotes without some extra digging, just an extra bit set
 
Will HD maps be required for L5?

The autonomous driving industry says "yes". HD maps allow the car to reliably know the precise location of static objects like stop signs, crosswalks, curbs even when the cameras can't see the objects. Being able to know the location of an object that the cameras can't see is a pretty important ability for autonomous cars.

Elon insists that "no" because he believes in the power of neural networks. He believes these problems can be solved with neural networks where autonomous cars won't need to "cheat" with HD maps. Although, I think Elon is being a bit misleading since Tesla does use detailed maps that serve a similar function as HD maps, even if Tesla does not consider them to be "HD".

Also, Tesla fans don't think HD maps are scalable (debatable IMO). Hence, they consider HD maps only good for L4 maybe, but must be abandoned in order to do L5.
 
  • Like
Reactions: mikes_fsd
Tesla uses maps but doesn't use HD maps, period. There's no evidence for HD maps whatsoever. They haven't shown a single instance of HD maps in any presentation. They specifically deny using HD maps.

Tesla uses maps, as far as we know, to:
Log the rough position and orientation of traffic controls
Navigate FSD routes similar to humans (turn right in 500 feet)
 
  • Like
Reactions: mikes_fsd
If we really want to get into this HD maps mess again.
Green thinks that the Google Maps available today to everyone are the HD Maps equivalent that Tesla is using...
You can follow this reply from green both up and down to try to make sense of his logic on HD maps. FSD rewrite will go out on Oct 20 to limited beta

At the 6 minute mark of the CVPR presentation, in June, Karpathy has this slide:
upload_2020-11-12_9-1-39.png

and says, "HD maps are not a scalable solution" in the same presentation he says Tesla uses maps with geocoded info.
So, whatever distinction Karpathy is making between "HD Maps" and "Tesla maps" is all about scalability.

We seem to want to bring up this subject over and over again, but this is not really our speculation, the main dude of Tesla AI says this, not some schmuck on a forum.
 
Last edited:
Reliance on HD maps (to the degree someone like Waymo uses them) makes scaling a general self driving solution basically impossible.

See also how Waymo after 10 years of work offers consumer-facing robo taxis in ONE tiny geofenced area with perfect weather and no where else- and STILL gets into situations requiring remote intervention.

But then you get to argue about what "HD" really means for maps, and what "reliance" does too.

What I'd like to see, though I doubt we'll ever see it, or at least see it in HW3, is personal route HD maps. A human driver knows their regular commute fairly well. I'd like my car to build it's own HD or HD-light map of my daily commute. At the very least, retain the general geometry it captured while driving with more detail than typical GPS maps.