Comment by Animats | Hacker News Reader

Comment by Animats | Hacker News Reader

Parent Full thread

Tesla's system doesn't have enough sensors. Musk forced his engineers to try to do this almost entirely with vision processing, and that was a terrible decision. Vision processing isn't that good yet. Everybody else uses LIDAR.

I've been saying for years that the right approach was to take the technology from Advanced Scientific Concepts' flash LIDAR and get the cost down. I first saw that demonstrated in 2004 on an optical bench in Santa Monica. It became an expensive product, mostly sold to DoD. It's expensive because the units require exotic InGaAs custom silicon and aren't made in quantity. Space-X uses one of their LIDAR units to dock the Dragon spacecraft with the space station.

Last year, Continental, the big century-old German auto parts maker, bought the technology from Advanced Scientific Concepts and started getting the cost down.[1] Volume production in 2020. Interim LIDAR products are already shipping in volume. Continental is quietly making all the parts needed for self-driving. LIDAR. Radar. Computers. Actuators. Cameras. Software for sensor integration into an "environment model". They design and make all the parts needed, and provide some of the system integration.

Apple and Google were trying to avoid becoming mere low-margin Tier I auto parts suppliers. Continental, though, is quite successful as a Tier I auto parts supplier. Revenue of €40 billion in 2016. Earnings about €2.8 billion. Dividend of €850 million. They can make money on low-margin parts.

Continental may end up quietly ruling automatic driving.

[1] https://www.continental-automotive.com/en-gl/Passenger-Cars/...

Interesting research on Continental. FWIW, looks like Tesla started using Continental's radar to replace Bosch just this week: https://teslamotorsclub.com/tmc/posts/2266769/

I suspect that if/when LIDAR is cheap enough, Tesla will use it.

In the meantime they outfit every single car with the best hardware that is realistic from a cost standpoint today, instead of waiting til 2020.

In the meantime they outfit every single car with the best hardware that is realistic from a cost standpoint today, instead of waiting til 2020.

And so far, it just sits there and does nothing useful, since the self-driving software that can do the job safely with those sensors doesn't exist.

It depends on what you're optimizing for. Others using LIDAR are optimizing for speed to market, while potentially sacrificing ability to solve the problem as fully. Musk's argument is that we know for certain that the entire road system can be navigated by visual cues, because that's how humans do it. We do not know for certain that this is possible with LIDAR.

  Musk's argument is that we know for certain that the 
  entire road system can be navigated by visual cues

We know for certain that human brains can be assembled out of regular atoms, but raising funds for a company that manufactures brains would be getting rather ahead of our current level of technology. The same might be true of computer vision and autonomous vehicles.

Please don't quote with code blocks. It messes with formatting on both mobile and desktop.

  It messes with formatting on both mobile

HN has never been mobile friendly, (element spacing, upvote button size) so I'm not terribly worried about that. If YC wanted HN to work on phones, they would fix their CSS.

  and desktop

Oh?

Block quoting can screw up desktop rendering on HN if the quoted line is too long. Say you put a 1000 character line in a literal quote:

  1.......1000

The whole page would render at the length of that line. So this text that I'm writing now, it'd normally wrap once or twice depending on your window's width. Instead, it'll only wrap after approximately 1000 characters. This is why when I block quote, I put in a line break every 60-70 characters (similar to what you did a few posts up).

I know this behavior happens with IE, I believe I've seen it in other desktop browsers but can't verify at the moment.

On desktop long block quotes go in a scrolling box. One that's pretty wide, so it doesn't usually cause a hassle to have to scroll it. But not so wide it breaks anything.

Additionally to what Jtsummers said, quoting on desktop can drastically reduce the line length compared to the rest of the lines in the comment, which is both ugly and annoying, especially since it makes it more difficult to read since your eyes are whipping back and forth.

There are a multitude of mobile apps for consumption of HN, most of which solve the code-block-line-overflow problem by adding a horizontal scroll, which only makes the problem worse by forcing users to scroll side to side to read the full quote.

Just use `>` at the start of the line to signify that it's a quote, please.

> There are a multitude of mobile apps for consumption of HN, most of which solve the code-block-line-overflow problem by adding a horizontal scroll,

HN does that natively - just drag the code block around with your finger. Still sucks to use though, so > quoting is the way to go.

So you believe the technology gap for assembling a human brain out of regular atoms is similar to that for navigating a road using cameras?

JFC. I don't get why otherwise rational people become so stupidly aggressive when faced with analogies.

Obviously not, the former is far easier.

Humans also learn how to drive using several orders of magnitude less training data than self-driving car efforts are using. Do you think that means we should strip down the training data sets to human-equivalent levels?

People and machines behave and understand the world differently. Just because it works for people doesn't mean it'll work for computers.

How on earth do you reach that conclusion? Every person has at least 16 years of high bandwidth sensor data experience before getting behind a wheel, plus millions of years of evolutionary training built into the structure of their brain.

A vast amount of data has produced every driver.

Data specific to driving is still quite limited for people, even when counting all that passenger time in childhood. The average driver in the US drives maybe 10-15k miles a year, so optimistically a person may have 200-250k passenger miles before driving? Consider that Waymo's cars drove 3 million real world miles in 2016 alone, and 25 billion virtual miles in a simulation, and all those miles are potentially used in a model shared by all the cars.

Now if you want to just look "total experience in the world as a whole" the numbers look a bit different, but if anything that just accentuates the differences here. We don't currently have a way to teach computers to construct mental models of how the world works the same way humans think about it, which would be necessary to use training data that was about the world as a whole.

What's more important, we can teach people much more effectively than machines by communicating concepts. Imagine the driving performance of a hypothetical person who is not capable of speech, so the only way to train them is having them drive in a simulator, and feed them candy when they get it right, while administering electric shocks for every driving error.

Humans are pretty bad at driving though, so I'm not sure that basing self driving technology on us is a good idea.

Humans are bad at driving because they can get distracted and can only look in one direction at a time. A computer system based on visual processing shouldn't have either of those issues.

I agree fundamentally though that it's a weak argument to say a certain approach should be taken just because it's how humans do it.

The argument is not that it's because how humans do it. The argument is that the entire system was designed to be consumed as visual cues.

As far as we know, humans are better drivers than every other species in the universe.

> Musk's argument is that we know for certain that the entire road system can be navigated by visual cues, because that's how humans do it.

i'm hesitant to argue against Musk, yet the actual goal is to navigate better than humans, and it seems reasonable to suspect that having wider range sensors would be a good (and may be necessary) foundation to achieve it.

just watching the current system navigate at night on poor lightly stripped streets is frightening enough to keep me away. search youtube for autopilot fails and there are some videos from just this year.

You seem to be assuming that the limiting factor is cost and parts. To me it appears the limiting factor is software. AIUI no one has the software to do level 4 autonomous driving, at any price. So why would a parts supplier end up "ruling" automatic driving?

I don't think thats what he said. Parts and hardware aren't a limiting factor, but they are an enablement for better software to be created (note Tesla in his example).

Without good hardware then you're stuck jumping through hoops to do all the processing software-side which leaves less time to react quickly and accurately.

I've seen you plug flash LIDAR (especially ASC's unit) several times here, but is anyone actually using it on SDVs? I've seen things that could have been flash LIDAR on test cars, but never as the primary sensor.

I worked on a team that evaluated the ASC unit a few years ago, but they found it unusable due to bloom issues. Has that changed?

I don't know how Continental is doing about staring into the sun. Here's a recent overview of flash LIDAR detector options.[1] That's more about the detectors used for looking down from aircraft, where ranges may be thousands of meters. Automobiles don't need that kind of range.

Flash LIDAR has a tradeoff between range and field of view. You can concentrate your energy or fan it out. There's been some interest in systems where you can narrow the output beam when you need to. Or you can combine wide-beam short-range units with narrow-beam long range ones.

There are other flash LIDAR companies, but many of them are vaporware. Quanergy, which bills itself as "the leading provider of solid state LIDAR sensors" announced a unit last year, but doesn't seem to have shipped. Velodyne wants to come out with a solid state product. TetraView is trying for a low-end semiconductor solution that uses common sensors. Luminar and LeddarTech want to use MEMS mirrors.

A long-range narrow-beam MEMS-steerable flash LIDAR might be useful. Look out 300m in the direction you're going when at high speed, and use other wide-angle sensors for a side view and in cluttered city environments.

Somebody is going to get this working at an acceptable price point reasonably soon.

[1] http://spie.org/newsroom/6466-comparing-flash-lidar-detector...

Why are flash LIDAR the way forward? Is it just packaging, and the ability to blend the sensor with the lines of the car? Cost? It certainly isn't image quality AFAIK.

(as an aside: I don't buy the moving parts argument -- a modern car has thousands, many of which have much tougher jobs than "spin at 900 RPM while on".)

I just recently heard about a company building automotive headlights with, let's say "selective lighting", using micro-mirrors. I was then surprised to find out micro-mirror devices have been available for quite a long time. [1] I wonder how come they haven't been used for a solid state LIDAR yet, they seem to be _very_ well suited for this. Are there any limitations because of the power of the LIDAR?

[1] https://en.wikipedia.org/wiki/Digital_micromirror_device

>Tesla's system doesn't have enough sensors. Musk forced his engineers to try to do this almost entirely with vision processing, and that was a terrible decision. Vision processing isn't that good yet. Everybody else uses LIDAR.

I think I agree with this, but is LIDAR expected to work in the rain?

I think this is the wrong question to be asking.

Instead ask: "In good weather, can LIDAR-less systems match the safety of LIDAR?" The answer is "not yet"

Until LIDAR-less systems can match the safety of LIDAR they will simply be banned from the roads, or limited to certain situations such as rainy weather.

Regulators, politicians and society will not allow Tesla to operate a system which in good weather has a much higher accident rate than is necessary - and they will not accept "cost reasons" as a valid excuse for not installing a LIDAR.

Laws are frequently named after a single child who died.

"In a gold rush, sell shovels"

You forget the (smaller, repeatedly broken up) American auto counterpart of Continental: Delphi. Google them and you'll find they're doing pretty well of late.

This is an example of Tesla's RADAR sensor at work.

https://m.youtube.com/watch?v=BE2lQK_0CDw

And Tesla's engineers aren't the first to bellyache about being asked to make the impossible a reality.

Reply on news.ycombinator.com