Hacker News Reader

Comment by Animats | Hacker News Reader

More cameras. Better sonars (very short range). Better radar processing, but apparently the same old single radar at bumper height. Still no windshield-height radar. No radar scanning in elevation. No LIDAR.

That's a better, but still weak sensor suite. It's probably enough for freeway driving under good conditions. It's far below Google's sensor suite. Or Volvo's.

Now they just have to write software smart enough to not plow into stationary vehicles on the shoulder. There are videos of three separate Tesla crashes where the Tesla plowed into a vehicle partially blocking a lane.

There have been several announcements of low-cost solid-state LIDAR units for automotive. Quantergy announced last year, but didn't ship.[1] Innoviz announced this year to ship in 2018.[2] Advanced Scientific Concepts can't get their costs down.[3] (They have a great unit that costs $100K; the Dragon spacecraft uses it during docking.) Those are all-solid-state devices. There are also some companies trying to use MEMS mirrors, like TV projectors. Eventually somebody will get 3D LIDAR technology working at a low price point, but it hasn't happened yet.

[1] http://www.quanergy.com/products/ [2] http://spectrum.ieee.org/cars-that-think/transportation/sens... [3] http://www.advancedscientificconcepts.com/applications/autom...

View on HN

It's staggering that Tesla don't use LIDAR at all.

Sure LIDAR is more expensive but the data quality is very stable and good. Cameras won't work that well or at all in night or when the sun rays blend in morning / evening or in bad weather conditions like heavy rain or snow. All other self-driving cars I know of have at least a small LIDAR beside several other types of sensors.

Lidar doesn't make sense for Tesla with the failure rate alone. They're mechanical devices, and they often have defects when produced at scale. They need to be replaced very often. This makes sense when you're doing vehicles as a service, but not as much for Tesla's use case.

What about the solid state LIDAR mentioned by the GP?

Needs to be mounted high up and then the car doesn't look cool anymore.

Once the cost goes down, multiple LIDAR units can be mounted behind the windshield. Cars designed for automatic driving should reserve the top inch or so of the windshield for sensors, with the optical sensor suite mounted at the front edge of the roof and covered with padding matched to the headliner. (Yes, auto glass today is often infrared-opaque to reduce heat input, but the IR-opaque layer could be omitted for the top inch.) Putting cameras behind the rear view mirror is already done, but more space may be needed.

The rotating Velodyne thing Google uses is an interim research tool. In production, expect to see better design.

With enough LIDAR units, full-circle coverage can be achieved. This is already being done with cameras. The rear and side units need less power and range than the front-facing units, so they can have a smaller collecting aperture, making them easier to place. For the front-facing units, the emitter can be spread out, allowing higher power. The rules on eye safety are concerned with the maximum power coming through a 1/4" hole (human eye pupil). If the laser emitters are spread out over the width of the windshield, the maximum power per unit area is reduced.

The CMU/Cadillac demo a few years ago managed to conceal all the sensors. It can be done. It just costs too much until one of the three companies claiming to be building low-cost automotive LIDARs Real Soon Now manages to deliver.

> Once the cost goes down...

> ...expect to see better design.

> ...costs too much until one of the three companies claiming to be building low-cost automotive LIDARs Real Soon Now manages to deliver.

All I see is future promises. Tesla wants to do this now, that too at mass market costs, independent of whether those promises materialize or not.

It also depends on how much you trust their R&D team. If they thought this was a week sensor suite and something else was just around the corner, they won't risk fitting their cars with inferior technology which would be obsolete in a couple of years, making all the collected data worthless. Yes I am aware they did this to existing cars lacking the full autonomy hardware, but I find that OK, since, for them, they never promised full autonomy to begin with.

> It also depends on how much you trust their R&D team

R&D doesn't make top-level budget decisions.

If the product fails, that's on Musk and Musk alone.

He takes credit for its successes. He can take responsibility for its failures too, as we would expect from any respectable CEO.

No, Tesla doesn't want to do automatic driving now. They want to hype it now. The software isn't ready yet. In their announcement, they say that the new hardware will, right now, support fewer automatic driving features than the current hardware.

Low-cost LIDAR is a problem that can be solved with money and a customer ready to buy a lot of units. So far, no car company has said "we'll take a million units a year at $100 if they meet these specs." It's lack of demand, not lack of technology. It's a risk for the LIDAR maker. There are automotive radars available better than those currently on cars, but they don't have a volume buyer yet.

You provided no evidence for the hype claim but even then you are contradicting yourself:

Statement A: Tesla is just hyping autonomous driving since no technology can deliver fully autonomous driving right now.

Statement ~A: Low-cost LIDAR is available to any high-volume customer, of which Tesla is one, given Model 3 demand. And LIDAR can deliver fully autonomous driving.

The point simply being, I have a hard time believing that even with good-looking, cheap LIDAR being a possibility, as you claim, Tesla chose to go with an inferior sensor suite, for no apparent reason.

What they're doing now is a minor retrofit to their existing crashes-into-big-solid-objects "autopilot". The Model 3 is at least two years away. By then, the technology should be better.

Volvo already has a concealed LIDAR in their self driving car.[1] Volvo will put 100 self-driving cars in the hands of ordinary drivers in Sweden in 2017. This is self-driving by the pros. Makes Tesla look amateurish.[2]

And Volvo is way ahead on marketing self-driving.[3]

[1] https://www.youtube.com/watch?v=2q00jIBhkq4 [2] http://www.volvocars.com/intl/about/our-innovation-brands/in... [3] https://www.youtube.com/watch?v=uDB6fFflTVA

As far as I understood the Volvo LIDAR is windshield mounted?

http://www.sae.org/dlymagazineimages/web/516/11702_15456.jpg

Making the necessary gubbins look cool is a job for the company's designers.

Spinny LIDAR on top is never gonna look cool unless your aesthetic is "Ghostbusters", but it's also already looking technologically obsolete.

   >> then the car doesn't look cool anymore.

This doesn't look cool either:

    >> There are videos of three separate Tesla crashes where the Tesla plowed into a vehicle partially blocking a lane.

Besides being almost as expensive as the car they mount on top and stick out like a sore thumb. This is the reason. It doesn't look cool. Amazed that people don't realize that.

In that sense, aerodynamics might play a role as well.

there are systems with LIDARs mounted next to the headlights.

LIDAR's high cost[1] is prohibiting it from being commercialized. This article below was dated 2015 so price might have actually gone down a bit - but still a few orders of magnitude higher than cheap ultrasonic sensors and cameras.

This is actually common problem not only for Tesla. We won't see LIDAR in production cars (you will certainly see them in prototype cars as usual) until its cost drastically reduced

[1] http://articles.sae.org/13899/

It's more than just high cost, LIDAR has issues with rain making it questionable in real world conditions. AKA, if you need some fallback, just use that fallback.

With range gating, you can do a lot with LIDAR in rain and fog.[1][2] It's possible to ignore all reflections out to a given distance, and you can adjust that distance. There's been a lot of work on this for the military. Automotive LIDAR units don't have this yet, but it could easily be added. If you have both range gating and "first and last", where you get the time of the first and last reflection, you can easily distinguish fog and rain from hard objects. The distribution of first and last times on fog and rain should look like Gaussian noise, while a hard target has near-equal first and last.

This is something we'll probably see in second-generation self-driving cars. With range-gated LIDAR and millimeter radar, self-driving cars will do far better in fog and heavy rain than humans do.

[1] http://www.sensorsinc.com/applications/military/laser-range-... [2] http://www.obzerv.com/en/range-gating-technology/core-expert...

> That's a better, but still weak sensor suite.

All you have is two overlapping cameras with poor resolution and they let you behind the wheel, right?

This is a software and processing hardware problem only. Human beings drive OK for the most part and we don't have lidar/radar/sonar/etc.

With 8 cameras providing 360 vision and an advanced NVIDIA hardware suite, this problem is entirely solvable and this is a very reasonable solution to it.

In theory, sure. But you're effectively saying computers have to process vision as well as humans, which some experts think is comparable to solving AI. I'm not saying Tesla won't make a lot of progress, but considering how many decades of research has gone into this very problem (I worked on it at Carnegie Mellon as a graduate student, and they've been doing it since the 80's) it will likely take longer than most people like to admit.

I actually don't know a single expert who thinks we're anywhere close to having a stage 4 autonomous vehicle. Most of the people I respect are pegging it at decades instead of years.

I think you must mean that you don't know an academic expert who thinks we're close. The teams at Tesla must think it's achievable in the next few years.

Or they want you to think that.

Or it's the marketing team and not the engineering team :-)

Well I also used to work with the team that's now at Uber (I worked at the National Robotics Engineering Center which is where they got almost all of their initial 40 engineers), and some of the guys I used to work with went to Google as well as Tesla. Tesla has a retention problem - their top engineers regularly quit every few years. Google lost nearly their entire team working on self-driving cars because Sergey insisted on making a fully self-driving vehicle instead of rolling out incremental levels of autonomy.

If you're only reading article after article hyping the technology instead of talking to the people actually building it, of course you're going to believe the journalists instead of the engineers.

I don't know anybody actually writing the code that thinks we're close. Of course the PR teams and managers are going to hype everything up - that's their job.

Excuse my rather hand-waving explanations below but I'm not an expert in this field.

The guys you spoke to, was this before or after they went to Google/Uber? As I understand the approach at Tesla would be (a) collecting an enormous amount of real-world driving data that possibly others are not or have not done yet, (b) do as little "coding" as possible, but rather take a deep-learning approach. I.e. the "algorithm" gets better the more data you throw at it, it doesn't depend on human intelligence, but rather on how much data you collect. Similar to how Google Translate got so good (it's not good because of any linguistic model or because a team of linguists "coded" it, it's just good because of the sheer amount of data it was trained on). (c) they have enough money to throw at any hardware requirements for a platform that could train on such an amount of data.

Are the conditions above not different perhaps from what you experienced whilst working at NREC and perhaps different from the way you guys approached autonomous-driving? I'm trying to think from Elon Musk first-principles. If it was technically possible, what are all the lego blocks you need to go about building this?

That certainly seems like the idea for Tesla - collect all the data from the cars in real-world condition and use it for training. I think it's brilliant. Time will tell to see how well they can leverage that data. Having a better sensor suite would certainly make it easier though. It's a lot easier to interpret LIDAR (it basically makes a 3D map for you) than it is to interpret a camera (you have to figure out how to turn 2D into 3D).

My skepticism is in the leap towards a fully autonomous vehicle, safe for driving on real roads in harsh conditions. Very few people have attempted anything in bad weather yet, although Ford has actually done some work on it.

If we are talking about incremental improvements - sure that will happen constantly. What I'm saying is I don't believe you're going to get into a car without a driver anytime soon.

Tesla has been recording driving data, sensor data, and GPS data for years. They already have a system in place for machine learning driving data. This combined with their sensors is IMO a great basis for fully autonomous cars.

Tesla is planning to have their factory produce 500,000 cars in 2018. Even if they only meet half of that they'll still be collecting a lot of data.

I would guess most major cities and freeways will have full autonomous support by 2019.

I'll give you 10 to 1 odds on that if you want to bet.

Driving in snow is simple. Just hit the brakes hard whenever you are uncertain.

I think the comparison to google translate is interesting. How much confidence do you have that a translation by google translate is correct? How would you compare the complexity of translating documents between various languages to the complexity of driving a car in arbitrary environments and situations. And the most interesting think: Would you trust your life on google translate? At least in the sense that it does not corrupt the meaning of a translated text (small errors ok)?

The way I see it panning out is that the first generation of autonomous vehicles is going to have a tough time. As more and more cars become fully autonomous there will develop some sort of protocol where every autonomous vehicle will be relaying exactly what it's "about to do" to anything nearby. I.e. your car will know if the car in front / behind / on the other side of the freeway is about to do anything and thus eliminate what is currently seen as "unpredictable environmental factors" to a very large extent. Internet of Things will also increase the amount of additional data being fed to the car's computer, e.g. it will know exactly when a traffic light is about to turn red because they will receive this data directly from it.

Once you reach a certain critical percentage of fully autonomous vehicles regulation will probably change to such an extent that car insurers will void your insurance if you drive it manually, or give a significant discount to your insurance premium if you don't ever put it in "manual" mode, etc.

I think the problem is really difficult right now because humans are still allowed to drive, but as that ratio goes down it becomes less of a serious problem.

Google is at stage 3 and getting close to stage 4 for slow (< 25MPH) vehicles. See Urmson's talk at SXSW, where he shows Google cars handling many unusual situations. (A woman in a powered wheelchair chasing a turkey with a broom is one example.) If you're willing to drive slow, more situations can be handled. It's not necessary to have as much sensor range, situations are less ambiguous, braking hard will resolve most problems, and predicting what's going to happen becomes less important.

At higher speeds, prediction becomes more of an issue. Will that car on the side street drive into the intersection, or not? At slow speeds, you can wait and see what they do. At higher speeds, you have to predict behavior. That's hard, but a machine learning problem.

This is with the caveat that the weather is nice and you're in one of Google's specially mapped areas. As I understand it - and I never worked at Google and I keep waiting for someone who does to correct me on it - those maps are very difficult to scale and maintain.

The limited area/weather isn't a huge problem.commercially . You just need to start offering it in a single city when possible, prove the tech and business works , and than you could get tons of money to scale to other cities .

I don't see how that solves the weather problem, unless you're limiting yourself to cities with nice weather. Even then - all cities have bad weather occasionally.

I'm not arguing against progress. I think it's great people are working on this and obviously you have to start somewhere. I'm just pointing out the fully autonomous vehicle is not a few years away.

What do you make of this?

https://www.tesla.com/nl_NL/videos/full-self-driving-hardwar...

Part of the problem I think is that the image is not fully used; I mean, generally these systems consist of a black-box routine that extracts interest points and then passes it onto to a SLAM routine, which inturn keeps an estimate of the car state and the interest point positions. There is no "physical" model of the world being inferred from images, and I imagine this makes things rather tricky (and also why a LIDAR is so much more useful).

AFAIK deep-learning hasn't really brought much change to this manner of doing things - the mapping part esp.

Researchers already passed human level image classification performance in 2015.

https://arxiv.org/pdf/1502.01852.pdf

That's a really cool paper. While I wouldn't say they're doing better than a human (I really think they're twisting semantics with that one), it is impressive and I hope to see a lot more progress in this area. Thank you for sharing!

Hey, the human eye is a damn fine camera and has the equivalent of a room's worth of NVidia cards in processing power behind it. It does have very poor depth perception compared to radar or lidar, a quite limited field of view, and an inability to see through fog but lets not sell a person short.

> processing power behind it

And the best that all that processing power can come up with is: tailgating someone. No thanks.

With modern driver assistance packages, that's not true anymore. For example Mazda's safety technology apparently gives drivers access to radars and radar coverage that Tesla simply don't have; see e.g. http://www.mazda.com/en/innovation/technology/safety/active_... There's also been talk of mandating active safety systems on new cars because regulators are worried about the safety of drivers relying on just their own two eyes.

> It's probably enough for freeway driving under good conditions.

It might rule out driving in fog or snow, but if the weather is OK then the sensor suite is probably enough for driving on any paved road.

I think you are underestimating how quickly Tesla's AI software can improve. The AI will always be running and checking to see if the driver does anything unexpected. For example if you brake or swerve suddenly when the AI did not expect you to do so, all the sensor data and state of the AI will be sent back to Tesla headquarters. Tesla will amass millions of these 'significant events' and be able to back-test updates to their AI logic against every single event. It will be test-driven development on steroids, allowing rapid software improvement while maintaining high quality.

"I think you're underestimating how quickly Tesla'a AI software can improve"

Do you have any basis for that statement, other than that you like Elon Musk, and read an article about "big data" once?

I have actual experience in this field. The parent is right: the sensors matter more than hypothetical gains from "big data". Because right now, sensors suck. There's a reason every serious player in this game is plopping pairs of $10,000 lidar units on their test cars (and still failing to work in the rain). You can't distinguish signal from noise by collecting more noise, and until you have a decent sensor suite, "big data" is just garbage-in, garbage-out.

Reading this discussion, you'd almost forget that the reason Tesla is doing this is because its existing sensor suite was proven to be a total joke in actual use...but hey, here's to the emperor's new outfit! One hundred percent less naked than the old one!

I feel sad that actual sensible responses on HN are now downvoted mercilessly because they are not aboard the hype train/zeitgeist

What irq11 has said is valid and important. You cannot just handwave actual engineering problems with some buzzwords

Perhaps irq11 was downvoted because of his/her belittling tone rather than the substance of his/her argument.

It's valid and important to say, "I have experience in this field and sensor quality will always be more important than a suite of regression tests, no matter how large".

It's not valid and important to say, "You're just saying that because you like Elon Musk, and read an article about 'big data' once"

Well, yes. That's why I didn't say that. I asked if the parent had any evidence for the otherwise bald assertion of a dubious fact (other than being a fanperson).

It would be a perfectly reasonable response to my comment to say "why yes, I do, and here's my evidence (along with a great reason that I didn't provide it in my original comment!)"

It stands to reason that if the parent had a better argument than "test-driven development on steroids", then they'd have provided it. But asserting that someone's well-informed comment about sensor quality "underestimates" the power of big-data-AI-TDD-learning-rapid-innovation-foo, well, that sounds like magical thinking.

In any case, this is a meta-discussion about voting, which is definitely not okay, so I'll leave it at that.

Now is your chance to show what other you evidence you have and dispute his snarky fanboy comment.

Agree on that. Would however love to read about some confirmed info whether AI/NeuronalNets/BigData really plays a significant role for any of the current automatized drive developments. And how this fits together with safety regulations (ISO26262, ASIL) and the required deterministic behavior.

> existing sensor suite was proven to be a total joke in actual use

The existing sensor suite was is sufficient for the existing driving assist functions. The existing suite was never expected to be sufficient to provide the predicted, future levels of autodrive.

What do human sensors cover? Primarily our eyes. Have we had sensors that provide about the same dynamic range as our human vision?

I am NOT a part of this field, so please excuse my presumption.

I believe that most/all(?) current/previous 'automated control' systems used sensors that were lower resolution than our eyes, but that also provided some indication of range data.

Do our eyes provide distance data? I think that data is calculated with quite a bit of accuracy in our brains.

Again, I'm not part of this field, so this is just speculation.

In theory, given a fast enough computer and advanced enough algorithm, given nothing but visual input streams no more advanced than are available with existing optics, could a car autopilot not be just as good a driver as a human?

This would be a high quality comment if you removed the first and last sentences.

Those are good points. Here's some backup.

Here's Mobileeye's CTO talking in detail about their technology.[1] They use a a deep neural network that recognizes cars. Just cars. Look at this point in the video.[2] There's a stationary obstacle on the left, and it doesn't recognize that. Maybe this is why Tesla, using Mobileye hardware, plowed into three different stationary vehicles at the left of a highway. They were all trucks (one was a street sweeper), and seen from the back while stationary, may not have matched the trained model of "car".

Mobileeye's algorithms are good at recognizing moving cars from most angles, and that's valuable for feeding into the model of "what's the other player doing". But it's not enough for the "where are all the obstacles" and "where is the road surface" questions. Those questions have objective answers, which can be answered with a LIDAR, and maybe millimeter radar if it has the angular resolution.

[1] https://www.youtube.com/watch?v=GCMXXXmxG-I [2] https://youtu.be/GCMXXXmxG-I?t=1758

Could you share with us what the state of art in SLAM is ? Last I looked, the backend is essentially the same old Kalman-smoother, but the frontend is generally used with hand-tuned features (stuff from the 90s - 00s). Big Data, incl. all the rapid advance in object-recognition CNNs can't really seem to do much in this case; unless there are radical new techniques that I'm not aware of.

Here's a very useful blog that deals exactly with this topic, http://www.computervisionblog.com/2016/01/why-slam-matters-f...

I'd advise others on this thread, who tend to dismiss the difficulty of vision, to realize that techniques in ML and Deep-learning are very problem specific and not the holy grail of AGI they're made out to be.

Humans manage with only two cameras.

Actually we also have very sophisticated direction finding sound processing, and tactile sensors on the wheel and seat plus proprioceptors (a trade secret that no at tesla has been able to reproduce). This allows for e.g. driving over snow and being able to process the sound of the snow as it is driven over as a hint to the rest of the driving system.

You are giving humans way too much credit.

There's no reason why a microphone and an accelerometer would be hard to connect as input features. The car can make use of the same hints, and most probably make better use than even humans by learning to distinguish more informative patterns.

I agree it "could" happen - but nothing is shipping yet.

When we were designing our vehicle for the 2004 DARPA Grand Challenge, we considered installing a guitar pickup on the frame to sense vibration. Early plans for the Grand Challenge involved more off-road driving, and we wanted something to give us a sense of surface roughness. In practice, the off-road aspect wasn't important and we were overdesigned in that area. We had a 6-wheel drive vehicle and a windshield washer/compressed air cleaning system for the sensors, which were totally unnecessary.

Yeah, but our cameras have adaptive resolution, a large field of view, very good dynamic range + low-light performance, multi-axis control, and they are tightly integrated with a sophisticated processing system.

On top of what others have said, we manage very poorly.

Google's technology was already working years ago without machine learning. If you look at the object recognition/detection benchmarks, they are already human level, so cameras should be enough (with a good enough GPU). 5 years ago deep learning wasn't working. Algorithms are getting better faster than hardware getting cheaper.

if Tesla doesn't need LIDAR for level 4, I am skeptical of the "importance" of it made by the media

Reply on news.ycombinator.com