Hacker News Reader

I agree with you in some aspects but in the project description it is not mentioned the "self-driving" part, and I quote: 'Currently it performs the functions of Adaptive Cruise Control (ACC) and Lane Keeping Assist System (LKAS)'. This I can believe is performed, at some degree.

Do you think the media can be overinflating the whole project?

The reality is that Adapative Cruise Control and Lane assist is good enough for 90% of driving.

Let computers handle the boring parts of driving, and send the hard stuff over to the human when it needs to.

That is a lot of gains that we could be running right NOW on the streets.

What exactly will humans be doing while the computer handles the boring parts of driving? If the answer is 'supervising', that sounds even more boring than actually driving.

Right - they will not be supervising or paying attention or ready to take over in an emergency. Software that is designed with a dependency on drivers doing any of the above is an accident on its way to happen. The usefulness of certain systems that do not implicitly make use of human attentiveness cannot be assumed by those that do.

They should be supervising, but I imagine they will be staring at their phones. That said, there are 4-5 hour halls down I-65 that I wouldn't mind downgrading to "supervising" :)

I'd be happy if this type of system would scan really aggressively to warn/avoid the deer barely visible off the road before it jumps out in front of me...

What car that comes fitted with the required radars/cameras doesn't have those features already though?

That's entirely true, but it's also problematic because although you have "computers handling the boring part" you still need constant human supervision.

If the human driver mentally switches off or stops paying attention, things get dangerous as has been demonstrated in numerous Tesla autopilot crashes.

That isn't to say that these features don't improve safety. Widespread introduction of radar-based automatic emergency braking (AEB) will significantly reduce crashes and save many lives. But until we're closer to 100% autonomy, features that reduce driver workload need to be implemented carefully.

This isn't the entire codebase, it's what they've stripped out and open sourced (and a critical chunk isn't even OSS, it's a binary blob in the github repo).

Furthermore, ACC and LKAS are standard features provided by Honda/Acura with the required feature package needed to run this code! Neo will provide more functionality, but as of now, NeoOS hasn't been open sourced at all and is just a binary Android image.

It's not just the media though, he has a history of drama and overplaying things and is repeating the pattern here. I don't want to downplay his achievements, he's done a lot, but the hype is far greater than even that high bar of accomplishments.

Now I get the whole picture, thanks for the clarification.

Hotz described his product as 'fancy cruise control'. He actually doesn't play it up, and there are any number of fly-by-night autonomous driving startups working on rudimentary autonomous driving systems on next to no budget, but of course it's Hotz getting all the attention.

I'm seeing a lot of internet commenters reacting emotionally to Comma.ai, but the only thing they're getting upset about is their own inflated expectations. The actual product is too niche, too esoteric to warrant a strong reaction. It's a toy for nerds to tinker with, nothing more.

He claims to be on par with Tesla:

> It’s about on par with Tesla’s Autopilot 7, the one that came out when it launched.

That's not playing it up to you?

He's using a car that already has ACC and LKAS functionality to provide exactly that, yet scorns Tesla for using Mobileeye.

I'm not sure which part of your claim is in opposition to my own, but you seem to be disagreeing with something I said.

> He actually doesn't play it up

That part... ;)

I agree with your point. By the way, I would like to point the "elephant in the room": what if after so many hype, heuristic/ad-hoc AI becomes able to beat expensive "deep learning" AI in most cases? In my opinion, that could bring next AI "winter". E.g. imagine a 5-10$ RISCV (or ARM/PIC32) 300MHz microcontroller with 4MB of embedded RAM/ROM beating 400-2000$ ASIC/FPGA stuff with teraflops and gigabytes of RAM in most AI problems with "good enough" results.

edit: grammar corrections

I think I know what you're trying to say, but it's really difficult to follow your writing. Just some friendly feedback.

I'm not 100% clear what you mean by heuristic/ad-hoc. The first AI winter happened because researchers (e.g. Lenat) thought they could build true AI (consciousness) out of millions of rules.

If rules, guidelines, and heuristics can beat neural nets at driving (which they might, I agree with you on that) it won't start a new AI winter because it won't erase the other things that neural nets are good at (e.g. classification).

Sure, "real" AI is important. I meant that using massive resources for solving simple problems with expensive AI will drive to deception, as soon as someone shows equivalent results by using cheaper AI (e.g. minimax, pathfinding, linear optimization, expert systems, etc.). In my opinion, it is like hunting ducks with cannons when using a simple gun could be enough for it.

In my opinion, real advances apart, a large portion of the "new AI bussines" is just a way of getting money from investors telling following magic words: "deep learning", "big data", "jump the wagon before it gets too late", etc.

> "it piggybacks on the hard work of engineers at major manufacturers"

What hard work? It's a COTS part.

Not to mention, front-facing radar, even combined with image recognition is woefully inadequate for true self-driving. (As opposed to, for example, Tesla's so called "Auto Pilot".

What do you need for "true self-driving" in your mind?

Cameras + radar is more than humans have at least, so I think that should be enough.

Solid state LIDARs will soon be cheap enough and a good complement of course but I find it hard to believe autonomous driving would be impossible without it.

> Cameras + radar is more than humans have at least

execpting perhaps, that kilogram lump of gray matter floating around in our skulls.

You don't need LIDAR for self driving cars. You also don't need megawatts of power to play top-level Go: humans do it with a few tens of watts of energy. Yet Google needed ~megawatts of energy to run AlphaGo on their massive server farm.

Imagine two companies competing to win at Go, and one company had the attitude that megawatts of energy was not necessary, and another company threw whatever resources they could. The second company just played top-level Go this year. The first company is at least another 5-10 years away from it.

I'd also like to point out that humans do SLAM already (LIDAR solves SLAM). From a previous comment:

Humans implicitly perform SLAM (simulataneous localization and mapping). What do I mean? Look around your room. Close your eyes. Visualize the room. As a human, you've built a rough 3D model of the room. And if you keep your eyes open and walk through the room, that map is pretty fine-grained/detailed too and humans can keep track of where they are in the map.

LIDAR solves SLAM? LIDAR is a sensor you can use as part of a SLAM system.

Yes. The problem of SLAM from LIDAR, GPS / IMU data (which are cheap sensors), and pre-recorded maps was solved years ago.

You can do SLAM with only cameras ...

No you can't. Have you tried it?

Firstly, the state of the art (which is not deep learning based) still has trouble when the camera doesn't turn/swivel a lot (e.g. it has trouble when the camera is just moving straight).

And that's not even getting to the most critical assumption. All of visual SLAM right now assumes the environment is static (!!!), and doesn't work when things in the environment are allowed to move.

In my opinion, a system that combines LIDAR, GPS, and a pre-recorded map is not really doing SLAM. This scenario does not suffer from the "chicken and egg" problem which is a defining characteristic of SLAM. GPS and the pre-recorded map both make the problem much, much easier than SLAM in a GPS-denied, unmapped area.

Maybe we are just arguing over terminology.

We are. SLAM, LAM, same thing as far as building a 3d model of your environment is concerned.