Tesla is certainly on the knifes edge... they must be extremely confident
Tesla sued in wrongful death lawsuit that alleges Autopilot caused crash
201–210 of 495 posts
Re: Tesla sued in wrongful death lawsuit that alleges Autopilot caused crash
#202I would like to add over here that Tesla does indeed do sensor fusion in their cars. Their Autopilot combines radar and ultrasound with vision to decide where to drive. Commentators bringing up LIDAR are jumping the gun by assuming that this scenario isn't something that these sensors wouldn't have detected in either combination or individually to spot the anomaly. The problem over here is likely to be software due t…
This is why we use a redundant geometric approach (with LiDAR) as a fallback to NNs in our perception stack. NNs can have false negatives which is unacceptable, but dealing with false positives is more manageable and safer.
Does lidar have anything to do with the redundancy?
Re: Tesla sued in wrongful death lawsuit that alleges Autopilot caused crash
#203I would like to add over here that Tesla does indeed do sensor fusion in their cars. Their Autopilot combines radar and ultrasound with vision to decide where to drive. Commentators bringing up LIDAR are jumping the gun by assuming that this scenario isn't something that these sensors wouldn't have detected in either combination or individually to spot the anomaly. The problem over here is likely to be software due t…
The sensor technology is a big factor here. You can use LIDAR for perception without needing a neural net. On the other hand with radar + ultrasound + vision you'll need a neural net to try to detect things like obstacles, since the raw data from those sensors doesn't directly tell you whether there is an obstacle in the way.
That’s baloney and obviously wrong based on eyesight.
Re: Tesla sued in wrongful death lawsuit that alleges Autopilot caused crash
#204Earlier quoted context omitted.
In the video, that’s 50kph (30 mph). The most advanced Volvo system decelerated but still crashes pretty hard at 80kph (50mph) when approaching a stationary vehicle decoy. Note well, if the target was an inflatable rock or an inflatable jersey barrier parallel with the direction of travel, the results would be much, much worse. In those other cars the ACC is picking up a car-like object on the sensors. A non-car-like…
Again, you have failed to understand what I said. The CAR AND DRIVER link is this one: https://www.caranddriver.com/features/a24511826/safety-featu... And it shows a Tesla's performance among other cars, not a Volvo's. And anyways, that Volvo video does show AES invoking at 80km/h and 130 km/h, reducing the speed of the collision but not avoiding it. Maybe you're right about the car vs rock distinction but so far you…
I read the C&D article when it came out last year. Pretty sure it was linked from HN actually. They ran some tests with a few cars driving into inflatable vehicle decoys. As I recall they found the systems would fail as often as not.
I was speaking in terms of this crash, which was not a primary collision with another vehicle, when I said that AEB will not activate for stationary objects at speed.
The ACC will activate for a stationary vehicle at speed, but will not save you entirely at even modest speed. At 70mph even ACC can drive right into a stationary vehicle.
These systems often work by sharing the ACC sensors and lock onto and track vehicles. This is why C&D discusses specifically the case of a lead vehicle swerving to reveal a stopped vehicle, and how some systems fail to acquire the stopped vehicle in time to provide any brake input at all;
> Volvo's owner's manuals outline a target-switching problem for adaptive cruise control (ACC), the convenience feature that relies on the same sensors as AEB.
(C&D did not test drive a Volvo for that test, but they did review the Volvo system and comment on it)
It would not be fair to take the C&D testing with inflatable vehicle decoys, and use that to claim Tesla is behind by not activating AEB into a narrow profile partial forward obstruction (parallel jersey barrier).
Partial forward obstruction of a non-car object is the absolute worst case for false positives, because this roughly translates into “any signal at all on the forward sensor”.
I’ll try to find sample images of what the sensor data looks like under normal operating conditions and when looking at narrow profile forward obstructions.
This is a very tricky edge case. A bird can fly in front of you. Water can splash up from the car in front of you. An empty trash bag floating up in the air after being run over by a car in front of you.
I did not mean to imply ACC or even AEB will not slow down a car heading for a rear end collision with another stopped car. I was speaking about this event specifically. I apologize for not being more clear in my 1 sentence original reply.
Re: Tesla sued in wrongful death lawsuit that alleges Autopilot caused crash
#205Earlier quoted context omitted.
Everything you said is very wrong. LIDAR produces line of sight depth information. Stereo cameras produce line of sight depth information. Radar produces line of sight depth information. Ultrasound produces line of sight depth information. The only differences are how, how fast, and how detailed. The biggest difference between LIDAR and stereo vision is that LIDAR doesn't depend on illumination because it uses its ow…
Stereo cameras produce line of sight depth information That is wrong. Stereo cameras provide data which, after significant processing, can result in line-of-sight depth information. With radar (including LIDAR) and ultrasound, the depth information is essentially "free" since very little processing is required to extract the depth information (i.e., the information returned is the depth). So the biggest difference be…
Sure, but if you’ve got the processing it’s pretty easy.
Re: Tesla sued in wrongful death lawsuit that alleges Autopilot caused crash
#206Tesla made some very big claims in their recent autonomy day event- basically, they claimed that they are years ahead of competitors, while operating on "hard" mode (no lidar). And yet, a number of participants in the short autonomy day demo rides claimed that the support driver had to disengage the autopilot.
Has Tesla provided any evidence which shows evidence that they are in fact far beyond competitors?
Re: Tesla sued in wrongful death lawsuit that alleges Autopilot caused crash
#207Earlier quoted context omitted.
Aren't the algorithms for extracting depth from stereo imagery already solved and thus also essentially "free"?
No, it's like this: Radar, LIDAR, ultrasound: returning packets have X value, where X is run through a single algorithm to return depth. In the case of ultrasound and radar, the strength of the return packet is generally directly proportional to depth (algorithmically). LIDAR systems might require an additional processing step depending on the type of LIDAR system used, since many systems use shifted-arrays or rotati…
Sure it’s laggier but with dedicated silicon that’s a non-issue.
Fact is vision is more than adequate. And lidar is literally blind.
Re: Tesla sued in wrongful death lawsuit that alleges Autopilot caused crash
#208Earlier quoted context omitted.
Aren't the algorithms for extracting depth from stereo imagery already solved and thus also essentially "free"?
Depth estimation is not a solved problem. Why do you think SDC companies pay so much to buy Lidar if depth estimation was solved?
Because they bet wrong. It happens.
Re: Tesla sued in wrongful death lawsuit that alleges Autopilot caused crash
#209Earlier quoted context omitted.
These AEB features almost universally do not trigger on stationary objects when the car is moving more than 20 or 30mph.
https://youtu.be/kuxundRB1zM good aeb trigger at speed. they might not avoid the crash entirely, but try to mitigate it. Tesla is quite behind the state of the art here.
Here is the Tesla's Model S video as well:
https://www.youtube.com/watch?v=_5aFZJxuJGQ
Re: Tesla sued in wrongful death lawsuit that alleges Autopilot caused crash
#210I would like to add over here that Tesla does indeed do sensor fusion in their cars. Their Autopilot combines radar and ultrasound with vision to decide where to drive. Commentators bringing up LIDAR are jumping the gun by assuming that this scenario isn't something that these sensors wouldn't have detected in either combination or individually to spot the anomaly. The problem over here is likely to be software due t…
"What if the neural net or the system used to detect obstacles didn't see it because the precise configuration of the data fooled it?"
I think the answer is that all state must be represented as a probability over states and the controller must be set up to be a robust controller under this ambiguity of state. Then the problem becomes one of making sure that probability is well calibrated so that you get the appropriate control response. The whole software stack for autonomous driving is also a black box, that is just as big an issue. If some parts are sophisticated but are being fed into others that are naive we have no idea. Every sensor has issues, I wouldn't be so quick to worry about the neural networks in particular.