50 comments

[ 7.4 ms ] story [ 2148 ms ] thread
Time to “Cut the Gordian Knot” for self-driving cars?

https://en.wikipedia.org/wiki/Gordian_Knot

Let’s figure out how to make the problem easier to solve.

The solution is lidar but Elon refuses to use them for cost reasons and still thinks everything can be done with cameras because "that's how our vision works".

Most weird autopilot goofs and accidents stem from the car not knowing the exact distance to the object its seeing with its cameras.

Our vision only works because of "brains" and "optic nerves".
A neural net properly trained with binocular vision should be able to estimate the size and distance of the sign and the distance of the sign from the actual intersection. LIDAR is still probably way easier to get right, though.
We don’t have AGI yet so those are not enough for a car.
Does a fly have an AGI? Or a bee?

They seem to be able to do pretty complex navigational tasks with just visual and olfactory system with a minimal power budget.

True, so Tesla is not clearing even a much lower bar.
Sure it’s a general intelligence, why wouldn’t it be?
> Most weird autopilot goofs and accidents stem from the car not knowing the exact distance to the object its seeing with its cameras.

What about crappy roads with worn markings, missing / damaged signs or out of order traffic lights? How well are these handled?

I wonder if there's some patent issue preventing them from using LiDAR and by extension preventing Musk from admitting its superiority.
I don't think lidar is the only solution, passive\smart infrastructure changes and car to car communications could overcome many issues.
I think once you get to "overhaul civil infrastructure" vs "install lidar on cars" you've lost the thread a little bit here. Those suggestions are at least an order of magnitude harder to accomplish than simply installing a better sensor suite on the car.
Lidar isn’t a silver bullet. Google has been at it for over a decade and their taxi service still gets confused by traffic cones.

The cheapest Tesla is the base Model 3 at $49,000. How much more expensive would it be with lidar? It’s already inaccessible to the masses.

I'm not sure Tesla was ever aiming to have cars "Accessible to the masses" as their initial target. They seem, at least to me, to always be aiming at the more luxury end of the market.

Regarding lidar itself not being the silver bullet, that's true. The real silver bullet will be when we have an actual AI and not just stacked ML models. The difference between parsing visual and other sensor data and using that to navigate the world and understanding that visual and sensor data is a very sizeable one and I suspect until we get a reasonably decent task-focused AI capable of understanding we're going to see lots of these kinds of shortcomings.

Well I called out the Model 3 specifically as it was supposed to have been the car for the masses. I consider it way too expensive for the masses, but it’s likely one of the most affordable pure EVs you can buy, even at the high price.

I suspect actual self driving to follow the Pareto principle. The last 20% will take 80% of the time. And I’m not even convinced we’ve hit the 80% yet, anyway.

So basically, their self driving software assumes that stop signs are all the same size and is using some simple scaling algorithm to judge where to stop?

And this is what passes for AI? The singularity is upon us folks.

Just give me a simple car with analog guages and a Bluetooth radio so I can talk hands free and listen to the 100+ albums I have stored on my phone. I'll provide any "intelligence" required, thank you very much.

>And this is what passes for AI?

Autopilot is kind of a scam. Marketing and economic pixie dust from Elon over a basic camera lane assist system to make sure Teslas attract customers and hold their prestige and value.

But no, FSD is perpetually 6 months around the corner for him.

I do not know what buyers actually expect, but the phrase "FSD Beta" should sound very stupid to anyone with half a brain. It means that it almost drives like a human or better and it needs a bit of polishing. It is not only false, it is irresponsible too in this context.
Autopilot is kind of a scam.

AI is kind of a scam --- and autopilot is just one example.

FSD is perpetually 6 months around the corner for him.

And will probably remain so for decades to come.

“AI” is not a scam, but it’s a very wide term that people abuse.

The self driving software packages that work, the Cruze and Waymos and Mercedes of the world, are in limited careful rollout. These companies and teams want to make sure it’s safe and robust before entering mass market adoption.

Tesla is one of the weakest and most immature offerings of the major players, and it has been given the widest distribution. It is utterly frustrating.

"AI” is not a scam, but it’s a very wide term that people abuse.

The abuse starts with using the phrase "artificial intelligence".

If it's a scam, it's a really old one. It's been called AI since the state-of-the-art was a few "if" statements chained together.
It's an oldie but goodie --- but still a scam.

Selling what we call a computer as "intelligent" is like selling a DVD player as "creative" for it's ability to produce video. They are both just play back devices --- reproducing the work of others. Useful for sure but no more "intelligent" or "creative" than a hammer.

What would you call the field of Artificial Intelligence then? I can't think of anything.
Pattern Recognition? Automated Learning? Adaptive Software?
A new one that I'm kinda fond of --- Probabilistic Programming.

Any answer that "AI" provides has no assurance of correctness. It's up to you to judge if the probability is acceptable.

Do you really want to trust a self driving vehicle that misinterprets it's environment 20% of the time?

That's not a property of all AI techniques though, so calling the whole field by that seems very misleading.
Intelligence is not a property of all "AI" either.
None of those really cover much of the field.

AI also includes things like expert systems, which don't autmatically learn, adapt to much or recognize any patterns.

These type of systems are just a database with a multi-factor weighted search of some sort. Any "intelligence" involved comes directly from the real experts who set the weighting parameters and the user who interprets the results
Even if true, that doesn't make any of those other names fit any better.

The AI name works for me because it is goal oriented. Artificial Intelligence is implementing artificial systems that perform tasks that would otherwise require an intelligence (eg a human) to perform.

It doesn't really specify how you do that, it's just about what you're having it do.

Autopilot is just a marketing name for assisted cruise control. I think a majority of new cars have assisted cruise control at this point. My 2019 CRV has it.

Full self driving is the scam. And it goes up in price to $15,000 next week. It’s almost comical to imagine even one single person buying that.

This is true because they use cameras and camera lack depth information despite it being used with “state of the art AI”. This would not be an issue if they have any real depth based measurement system like a LiDAR.
How hard is it to use multiple cameras to gauge depth perception, as per human binocular vision?

(Serious question, I know nil about machine vision)

It's well understood to be impossible / not there yet, the argument then shifts to "but ML is awesome so it could happen?" to which the correct response is "anything is possible but we're nowhere close"
In some ways, it can be better than lidar. The colour signal can be perfectly matched with the depth signal to provide further context. This is particularly useful for tracking moving objects. Vision-based systems also have a far greater resolution, especially when additional cameras are added.

But having both would be best. The drawbacks of camera are additional compute being necessary for stitching and the risk of occlusions occurring due to something like adverse weather conditions. It also only offers a calculated estimation rather than a true distance measurement.

It’s extremely difficult, the biggest issue is that any vision based depth system is so damn noisy.
I'm not sure about USA, but here stop signs placement doesn't indicate the location to stop. Instead it is the line on the road if present. Or in such location where the visibility is sufficient. Fun to know that the system would fail here...
In many places in the US, and indeed other countries, the "stop line" can be worn almost completely away, or a road just repaved may not yet have one.
So I think I have a pretty good idea for successful hitchhiking or hold-ups…
You can already make fake road signs and trick human drivers.
Humans are unreliable. Thanks to Elon’s incompetence, my scheme will stop all the FSD Teslas all the time. And what better target, than those who can afford to piss away five figures on a half-assed scam!
Some campgrounds and trailer parks use small stop signs. Does this mean it will stop too late?
If only Tesla could have outsourced FSD to Google.
Large off-ramp stop signs are clearly an edge case nobody should expect to be handled well. /s