Hacker News Reader

Comment by cameldrv | Hacker News Reader

Tesla "Autopilot" cannot see stationary objects. It has a radar, but it is 1-D and low resolution. It uses this to localize other moving cars. In theory, it could use this to see a highway divider or firetruck it's about to hit, but since it's 1-D that would also mean that it would have to slam on the brakes when it sees an overpass coming up, because it can't tell the difference between an overpass and a highway divider or a firetruck. It can assume that overpasses aren't driving at 60mph, so it will see other cars. The Tesla algorithm is "look for lane markings and try to stay between them, and don't hit any moving vehicles. If there is any stationary object, including another a vehicle in your lane, the "Autopilot" will plow right into it. If they're not explaining this to customers, it's negligent.

View on HN

> If they're not explaining this to customers, it's negligent.

Offering a footgun to consumers which puts other people in danger is a questionable choice.

How is this a footgun significantly more than the plethora of other vehicles with the same functionality?

The problem is calling it "Autopilot" in the first place compared to other brands. If you call it "Autopilot", people expect it to be an autopilot.

This is kind of a kicking-a-dead-horse discussion, because as others have pointed out, "autopilot will fly you into a mountain" (complete with warnings).

But does the average driver know that? It is their perception that's important here. The Utah driver expected the car would not fly her into the mountain, so to speak. Naming, marketing context of self driving alongside this feature, and driver instruction are all likely part of the puzzle.

If you think the name has nothing to do with it, would you also say that if it had the name "Deluxe Cruise Control" people would treat it the same? I'm not so sure about that. Engineers maybe.

I agree that the name issue is a red herring, but not for this reason. Autopilots are operated by people who are trained in their use, and, for the most part [1], understand their limitations. 'Understanding their limitations' is exactly the issue here.

[1] Following a number of WTF-type accidents, there is some concern that airplane automation has become too complex for pilots to reason about when it partially fails, but if that is actually the case, it raises the bar for all partial automation, including for cars.

Pilots are professionals who are trained and paid to use autopilot.

It's a dangerously misleading name in aeroplanes too.

You are confusing autopilot with unmanned vehicle.

That's what autopilot means to me, both look synonymous and their marketing also plays on that confusion. They could just call it "assisted driving" to be more honest.

autopilot has a technical meaning which this is totally compatible with, however it has a common meaning that diverges.

The common meaning has been derived from the technical and is affected by some common misunderstandings, such as:

autopilot flies planes all over the place and hardly ever runs into anything becomes autopilot is very safe and not autopilot doesn't run into anything because in the sky there is not very much to run into.

I guess Tesla should have named it something else, but this kind of mistake is quite easy to make and I would say is the norm if anything.

I heard about significant number of plane crashes with autopilot involved, however I'm 42 year old. Maybe meaning of the word was changed with new generation. How old are you?

There is nothing in the post you are replying to that suggests this is any more of a footgun than genuinely similar cases. Any potentially lethal and undisclosed or insufficiently emphasized corner case is a footgun, and not responding to stationary objects is not even a corner case for collision avoidance: it's front-and-center.

What do you mean by "1-D"? If it doesn't measure height, then it would still be 2-D.

EDIT: it's a 1-D array of distances, ie, an array of polar coordinates on the 2-D plane. So, it's 2-D data.

Definitely, these things are subject to differences of opinion in terminology. What I mean is that Tesla's radar produces something like 1-3 1-D arrays as output. Those arrays are probably distance, intensity, and doppler. A normal camera produces three 2-D array of R, G, and B. Something like a Kinect produces arrays of R, G, B, and depth/distance. When one of the parameters is distance, sometimes people talk about a sensor having an extra 0.5D. A Kinect might be 2.5D. In this world, the Tesla radar is 1.5D. Fundamentally though, it can't discriminate by height, and so it can't solve the overpass/overhead sign problem. In a very real sense, that is why these accidents have happened.

I got it later - I'm not used to the terminology, thanks for the explanation :)

So its 2D if it returns a 1D array of distances?

If it returns a 2D array then its 'seeing in 3D'.

So I guess that works out.

Ok, now I understand the confusion. The array is 1-D, but it is a representation of polar coordinates on the 2-D plane. So, the data is 2-D.

But the Autopilot has cameras, and therefore it has 3+1D data (2D frames, 1D inferred depth from stereo or training, 1D time from frame diffs). Even with a 30fps frame rate at 60mph, two frames from a video feed would see the truck move closer by a meter. The network can be trained to distinguish a truck in the lane from an overpass, so those two frames should be sufficient to signal danger. Instead, the car accelerated over 70 meters before the human, again using just her eyes, applied the brakes.

This is an edge case, and clearly something that requires more training to avoid. The challenge in edge-case learning is to encounter them often enough, or perhaps create them artificially, and then to ensure the learning can be transferred and generalized to all similar situations.

This is not even the first time they hit a giant red stationary fire truck.

Seems like it happens whenever a car in front moves out of the way. It goes full stupid-cruise-control mode and just accelerates to the stored speed. Of course there's a lot of noise in terms of things in the stationary reference frame, but still, big red firetrucks don't exactly look like lane markings.

At the very least, as a driver, I'm slightly cautious when I see a car move out of the lane in front. I feel like the solution lies in: (a) separating the signal from the noise in the set of things moving with the ground, and (b) getting the car to 'understand' why other cars change lanes. Sometimes it's so they can avoid obstacles in the road. It almost feels like cars need to have a theory of mind of other cars. I certainly drive with an intuition of driver intentions stuck in my head.

The thing is Tesla will never be an autonomous car or a reliable semi-autonomous car with its current hardware. Elon needs to admit it (he insist, this isn't the case) and he needs to reliably convince his customers this isn't the case. Tesla needs to see a couple of cars before, in order to drive safely (these are no edge cases, this happens in real life a lot), but it can't see ahead without LiDAR but Elon insists they don't need LiDAR.

The problem is you can't have a pre-existing bias when working with this kind of technology, you need to be able to adapt to whatever changes that need to be made.

Well it's not completely clear to me that cameras can't look a few cars ahead. If we can, cameras can too. Maybe they're not smart enough yet, but I don't see the fundamental limitation. Apart from that, I agree that cheap + effective LIDAR would help significantly to get autonomous vehicles on the road now.

> If we can, cameras can too.

A camera is not a replacement of how our eyes work, how our brain distinguishes between objects and understand from prior experience and can make independent decisions on the fly. But neither does LiDAR, but with LiDAR and good machine learning (ie, Waymo), you can drastically improve the reliability of your autonomous system.

Elon is insisting on a no-LiDAR system at the same time advertising tesla autopilot as an autonomous system and has said multiple times that All Tesla on the road is already equipped with the hardware needed for fully autonomous driving. This is dangerously false.

There is not a single autonomous car out there (at least serious ones that we know of), that has a non-LiDAR system.

Reality Distortion Field is strong with Elon, Unlike Steve Jobs and Apple products, this one will kill people.

So I do a lot of work with both vision and lidar data, and am pretty intimately familiar with the public state of the art in terms of perception systems that rely on each or both, if not planning.

I don't think you need lidar.

I'm not saying that vision-only won't take longer, because it will, but I believe it'll probably happen unless the cost of lidar falls a great deal.

Have a look at, for example, "learning to see in the dark" (the CVPR 2018 one), DensePose (the FB research one) to get a sense of how quickly difficult problems are being solved.

The cost of LiDar has been falling for a while now: https://arstechnica.com/cars/2018/01/driving-around-without-...

Waymo LiDAR is custom made and they cost a lot less than what is available in the market. https://arstechnica.com/cars/2017/01/googles-waymo-invests-i...

I don't think it's a cost, Elon has a personal bias, or maybe he didn't account for the fact that in the not so near future LiDAR cost will go down drastically.

  But the Autopilot has cameras, and therefore it has 3+1D
  data (2D frames, 1D inferred depth from stereo or
  training, 1D time from frame diffs).

When making safety-critical real-time systems, we prize simplicity as simple code has fewer bugs and can be audited more thoroughly.

I mean _really_ simple - to deploy an airbag, early systems were as simple as checking the accelerometer works at startup, then firing the airbag if the acceleration passes a threshold consistent with a crash.

If your safety system relies on multiple cameras, stereo vision, camera motion tracking, and a neural network, it ain't simple.

That's why IMHO every self-driving car needs LIDAR - you need to be able to detect the firetruck or highway divider in from of you without relying on a million lines of code.

This is not an edge case. My understanding is that the Tesla autopilot uses the camera solely for lane marker detection. The Radar is used solely for moving vehicle detection. There is no sensor for stationary object in your lane detection. That's up to the driver, and that's why Tesla emphasizes the number of seconds the driver had their hands off the wheel in each of these crashes.

Precicely this scenario you are describing is in the owners manual that comes with the car.

Nobody actually reads the manual though.

I read manuals, back-to-back. Admittingly, I skip legalese disclaimers because they are intellectual noise. However, many manuals contain golden nuggets of information revealing that most people misuse a certain product or don't use it the best possible way.

A well written manual can be a teaching lesson about the relevant stuff given by domain experts for regular users. Who'd say No to that?

Hacker News commenters seem to be a poor representation of the average population.

Heaps of people read the manual.

I feel like this is the kind of thing you should be made aware of (for real, not just in impenetrable terms of service) before you even buy the car.

I feel like certain people will constantly complain about “not being made aware” about things no matter how clearly you state it to them. It is obvious in hindsight that many owners do not read the manual, don’t read screenwar warnings and skip the first time intro. But they still complain that “they weren’t made aware!” Ultimately if Musk himself took a trip to each individual owner and screamed at the top of his lungs the terms of service in their face, we would still be hearing people saying they weren’t made aware.

Legally they were made aware, if you want to change the law do so. But please stop letting user play morons selectively when it turns out they didn’t rtfm. That is on them and should be on them.

The primary sensors are the multiple cameras so this criticism is not relevant. The radar is primarily for low visibility situations (according to marketing). Of course the use of cameras as the primary itself is controversial but they do in theory disambiguate the situations you describe regarding stationary objects, limited resolution, and degrees of vision.

In balance: https://www.theverge.com/2018/2/7/16988628/elon-musk-lidar-s...

Also there was a car between the truck and the Tesla for some time. Since they are not specific about the timing, the car in front might have just avoided the truck which leaves little time for sensors below the lights to even register that there is something else. Does Tesla also have sensors at the top to see two+ cars ahead?

My gripe with this argument is that even some humans can have a reaction time to at least start braking under these circumstances, I don't mean all drivers but as an example: I raced go-karts (just as an amateur but took some lessons and also some racing classes in cars with 300 hp) and sometimes my reflexes while racing were surprising, even for myself, narrowly avoiding collisions at 100+ km/h in tight circuits. With that said I'd have thought that sensors and computers could be much faster than even my trained reflexes.

In a continuous system like racing, there are two more things to consider: prediction and relative motion. Avoiding something you can predict, or avoiding something which is moving much slower than 100+ km/h relative to you, isn't too difficult (I do it all the time on motorbikes and scooters). But this perception can fool you into thinking you have more control than you do - should something really unexpected occur, your reaction time will surprise you by being slower.

And what about the ultrasonic sensors?

source?

https://www.wired.com/story/tesla-autopilot-why-crash-radar/

Reply on news.ycombinator.com