Object detection is an issue, but it doesn't only apply to "moving" objects. The current state of the art is having trouble differentiating between a "solid" stationary object like a tire or chunk of metal that must be navigated around, and something like an empty plastic bag that can be safely ignored. Regarding infrastructure, AVs also seem to have a lot of trouble with faded lane lines on roads.

If you're relying only on cameras, then yes, that's an issue. Level 4 vehicles can locate themselves without the need for lines on the road. Everything has been carefully mapped out ahead of time.

>Everything has been carefully mapped out ahead of time.

That sounds sustainable and scalable to earth-size, for sure.

Well, it's been mostly done once with Google maps and street view. Presumably once it was done once and in use cities would have a vested interest in keeping it updated themselves. Judging by how many large potholes stick around for years near me I wouldn't keep my hopes up about it being well maintained though

Earth size? When was the last time you drove in Madagascar? Most driven miles are the same commute in the same city 10 times a week.

I wonder if you know what some of the roads in Madagascar are like, or just picked a country.

https://www.dangerousroads.org/africa/madagascar/3415-route-...

(A colleague is there now.)

People in Madagascar would answer differently.

People in Madagascar depend on 4x4 monsters. Yet manufacturers still sell RWD coupes. It's worthwhile without them as a target market.

> Everything has been carefully mapped out ahead of time

This doesn’t scale. This is also how Zoox, Tesla, and Cruise do their demo videos to scam more money out of investors: they collect ultra-HD maps in a very narrow area or a very specific route. Then they drive the route/area about a thousand times, recording each drive. Then they upload the drive with the fewest mistakes to YouTube. Just like me taking a thousand half-court shots with a basketball, hitting one, and then claiming I can do it every try. There, I just gave you the formula to raise $100mm from FOMOing VCs.

There’s also no such thing as a Level 4 Autonomous vehicle. It doesn’t exist.

Many of the original Google self-driving car project members came over directly from Google Street view, which has since scaled all over the world. Autonomous vehicles have big scaling issues, but it's not the maps holding them back.

Street view is not an ultra-HD map for autonomous driving. Cruise isn’t using street view, or anything so primitive and easy to collect in comparison, to navigate SF. And it doesn’t scale in and of itself without being backed by a search ads monopoly that prints money at will.

Telsa doesn't have such ultra-HD maps at all. They build some maps for some simulations and training but the car does not locate itself in such maps when driving.

> Then they drive the route/area about a thousand times, recording each drive. Then they upload the drive with the fewest mistakes to YouTube

This is certainty not what Tesla does, from Tesla vehicles you will find literally 1000s of videos uploaded by Testers on every possible routes.

And Tesla has also never relied on VC funding for any of its self-driving tech.

> Telsa doesn't have such ultra-HD maps at all. They build some maps for some simulations and training but the car does not locate itself in such maps when driving.

They do have such maps, but only for the routes of their demo videos. That’s the entire accusation — it’s a Fugazi. A man behind a curtain. A mechanical Turk. Fake.

This video was posted 3 years ago: https://m.youtube.com/watch?v=tlThdr3O5Qo

I have never seen a Tesla perform as well as the one in that video in ANY user video. I drove a Model 3 with FSD for 4 months and it never came close to performing that well. Not even close. They either recorded a special map for that scam video, drove the route a thousand times and took the best recording, or both.

> This is certainty not what Tesla does, from Tesla vehicles you will find literally 1000s of videos uploaded by Testers on every possible routes

In the real videos, the cars fail to navigate simple scenarios constantly. You can’t even watch the videos without cringing at the constant mistakes and dangerous maneuvers. For Pete’s sake, just search “Tesla phantom braking” on YouTube and behold.

https://m.youtube.com/watch?v=Zu18KYAhSzo

https://m.youtube.com/watch?v=9iGWDdnoONE&t=50

> And Tesla has also never relied on VC funding for any of its self-driving tech.

No, just unsophisticated customers paying $12k for vaporware and massive government subsidies. So sorry — they’re not scamming VCs. Just real customers and the taxpayers.

This pretty much. You've got to be able to rapidly identify everything you see and predict what it's reasonably likely to do, which sounds like an AGI-scale problem.

Unless… the infrastructure change they propose is removing the “other things on the road”. Sounds brilliant. You could call that system a “rail” road as the cars would be on rails in a manner of speaking

Vancouver's Skytrain has been fully autonomous since, I think, the 1980s.

Teslas, despite the ludicrous promises of "full self-driving in 6 months, just you wait", completely choke if there is so much as a traffic cone or a road-work barrier for it go around.

Tesla doesn't come anywhere close to having the sensor suite or compute to be taken seriously as a developer of level 4 autonomy.

What sensor suite is that?

Radar, ultrasonic proximity sensors, and/or LIDAR, presumably.

Tesla has famously removed all radar/ultrasonic sensors from their newer cars in favor of a purely camera-based system.

But they removed them well after they showed they could accurately predict distance with vision.

LiDAR doesn’t make sense as a sensor to me because it only works in good weather. Its Like a car without windscreen wipers.

Appreciate the downvote/disagreement.

But here is Karpathy explaining how vision can be used to measure distance to objects accurately[0]

Here is the fact that LiDAR doesn't work in the rain[1]: "... In heavy rain, for example, the light pulses emitted from the lidar system are partially reflected off of rain droplets which adds noise to the data, called 'echoes'."

Which logically implies you need to revert to vision, as see [0] also for why Radar is unreliable.

[0]https://www.youtube.com/watch?v=g6bOwQdCJrc [1]https://en.wikipedia.org/wiki/Lidar

How do humans drive without these sensors?

With superior reasoning abilities. Extra sensors generating more data make reasoning simpler.

Except that Radar and LiDAR send false negatives in poor weather/conditions. Making reasoning harder.

Extra sensors also generate a lot of extra complexity in fusing all the sensors into a common consistent view of the world. And if your goal is just 'more data' then you might as well just add more cameras, that's also 'more data'. I guess what you wanted to say way 'diverse data'.

As with everything else there are benefits and costs to those approaches.

I recently drove a rented Tesla 3. On a highway in Norway it failed to detect speed limit signs in like 25% of cases. And if the speed limit sign was a temporary one due to road repairs it failed with those like 50% of times.

And this is with stationary objects designed to be seen and easily grasped by humans.

Signs are nice but the simplest way to know the speed is to store it in a database and look it up from position.

In fact roads should have barcoded position markers everywhere. That way you could navigate without gps and all road signs could become virtual. Just download a road marker update.

That's not really how speed limits work though. People take their cues from road design and conditions, not what the speed limit actually is. Virtually changing speed limits is dangerous because it causes you to think you are altering the behaviour of traffic without actually doing so.

https://youtu.be/bglWCuCMSWc?t=280 goes into detail on this

We could put GPS beacons on all pedestrians.

The sarcasm tags are missing.

Teslas do not read speed limit signs or basically any signs other than stop signs, including not reading “Do Not Enter” street signs. It is using the wrong speed limit because their maps do not list the correct speed limit.

In Norway you are not getting Tesla FSD.

This reminds me of something the Rocket League devs said. Something to the effect that bots wouldn't be effective at the game because it's too challenging.

This is -not- my line of work, so I have no idea if that's true. But if it is, I don't see how we could have perfectly safe self driving vehicles.

This is where I see the problem being with autonomous vehicles, too. They can only react to other vehicles, they can't predict them.

Humans are good at predicting.

You don't consciously know you've seen the guy a couple of cars in front checking his mirror and his shoulder but you're hanging back because you just know he's going to pull out any second. The guy that's wavering a bit in the middle lane is about to dash across to the far lane of the sliproad that's coming up, clipping the zebra stripes a bit, because he's concentrating on the sat nav not the road, but you just know - out of all the other drivers in your space at the moment - that red Ford is the one that's going to do something boneheaded.

Autonomous Vehicles won't be able to do that, probably not ever.