5 ms·

I don't get it, why is it so hard to detect stationary objects in front of you? wouldn't that be the "hello world" of any self-driving tech? Every time I ask t

I don't get it, why is it so hard to detect stationary objects in front of you? wouldn't that be the "hello world" of any self-driving tech? Every time I ask this question, the response is "radar as a technology has too much noise, not accurate enough to detect stationary objects". But it DOES detect stationary objects. Otherwise no one would even turn on Tesla AP at all. Why is it so hard to detect stationary objects in front of you?

childintime 8y ago
The usual answer is that the signal that reflects of many stationary objects, like trees, cars, all of which have the same dopler shift and must normally be discarded. Still, this answer doesn't quite satisfy, because it ignores the very strong signal that exists right before a crash, which crash avoidance technologies do (seem to) use.
- abalone 8y ago
  The difference is in the compare. A collision avoidance system is there to help correct driver errors, so even if it only works 50% of the time it is still a massive reduction in the accident rate. Whereas with autopilot the expectation is that it should be recognizing objects as good or better than drivers, so a 50% failure rate would be considered massively buggy. Note: Tesla also has a collision avoidance system, which is actually on all the time and a distinct feature from autopilot (although it uses the same underlying sensors).
mentat 8y ago
Subaru Eye Sight does this reliably and actually auto brakes correctly. Going from that to APv2 has been a real step back. There's a curve on 92W that it consistently try to accelerate you into an embankment. I think Elon just doesn't understand the real world reality of how bad AP is.
- Someone1234 8y ago
  And Tesla's newer vehicles have the same sensors as Eye Sight, they just choose not to use them for Automatic Emergency Braking (AEB). Tesla's fans always defend the failure of their AEB by pointing out flaws in radar, while entirely ignoring that Tesla's vehicles have contained front facing optical cameras for years now. The Tesla shouldn't have hit the divider in the previous accident, the optical (visual light) camera should have seen it, even if radar didn't. Why doesn't Tesla's system work as well as Sabaru's Eye Sight?
- djsumdog 8y ago
  It really bothers me manufacturers are pushing incremental updates to car safety systems. Does anyone besides me think this is a terrible idea?
  - Gustomaximus 8y ago
    Aren't cars doing that today with their physical saftey systems? How else would they do it than incremental?
    - InternetOfStuff 8y ago
      > How else would they do it than incremental? How about... not? I'm not being disingenious here. If there's autonomus tech, then I find it scary that it might change its behaviour. You see, I've learned how (say) Autopilot responds in certain situations. I know when to trust it, and when not to. If this behaviour changes (perhaps even if it changes for the better!), that's a very dangerous thing. Some of my experience is now invalidated, but I don't know which part. But it mostly works as before, which gives me a false sense of security. Changing the driving behaviour of a vehicle, especially in totally un-obvious ways, is super dangerous, and such changes must have very strong justifications. Just think of that poor guy whose Tesla drove into the divider. Oh look, it behaves just like before -- except for its lane following behaviour.
      - Gustomaximus 8y ago
        I dont understand how you would expect not to do incremental improvement. Build it perfectly the first time? That's not realistic. Pretty much everything humans do is incremental improvement. I understand your point that we dont want to introduce weaknesses in 'upgrades' but this should be a discussion about how to create robust QA rather that stopping incremental improvement.
        InternetOfStuff 8y ago
        > I dont understand how you would expect not to do incremental improvement I wonder if we're talking past each other? Sure, incrementally improve during engineering. But don't incrementally change the safety-critical behaviour of a car once it's in the customer's hands. > Build it perfectly the first time? That's not realistic. Absolutely agreed. But for subjects such as autopilots, predictability may very well trump improvement. I'm assuming that the initial iteration is already reasonably safe and useful (if it isn't it shouldn't have been shipped, right?). If the behaviour changes unpredictably (how will you predict autopilot changes caused by an update?), this may well be less safe than keeping the existing system, whoose quirks the dirver has noww learned. > Pretty much everything humans do is incremental improvement. > I understand your point that we dont want to introduce weaknesses in 'upgrades' The nasty thing is that even a change to objectively superior behaviour may be problematic, because the driver will expect the car to behave one way, when it now behaves a different way (regardless of merit). Not even a changelog would help in such cases. I feel like the autopilot ought to inform the driver in each situation where it used to decide one way, and now decides another way. That might be a reasonably safe way of allowing incremental change, but is likely too annoying for most people. > but this should be a discussion about how to create robust QA rather that stopping incremental improvement. Robust QA is certainly a must. (side note: I used to work in safety-critical automotive projects. I may have a more intimate understanding of the issues)
megy 8y ago
> I don't get it, why is it so hard to detect stationary objects in front of you? This could only be said by someone who has never worked on the problem. I don't get it, why is programming so hard, you are just telling the computer exactly what to do?
- maxxxxx 8y ago
  It seems a legitimate question. I have never worked on this problem either but I would have thought that detecting a stationary object right in front of the car should be one of the basic and simpler functions of a self driving system. Obviously this doesn't seem to be the case but maybe you can enlighten us why this is hard?
  - bobsil1 8y ago
    False positives from road signs, overhead signs, radar-reflective trash. Commonly used radars have limited vertical discrimination.
    - lagadu 8y ago
      Why would the radar even be pointed at areas several meters above the car? Maybe radars can't be pointed at a narrow space, if that's the case why not use lidar instead?
      - bobsil1 8y ago
        Most radars aren't coherent like laser, beams spread out. Also radar measures speed more accurately than location. Lidar only used by R&D + taxi services, price not reasonable for car buyers yet.
- dingaling 8y ago
  Instead of being condescending this would have been a good time to explain the challenges A static object has a negative Doppler shift rate linearly in proportion to vehicle speed and a y-axis position greater than zero but less than four metres. What makes that difficult to detect or distinguish?
serioussecurity 8y ago
Tesla are clowns. It's possible to do this right, but they're shipping a product they know isn't suited to the task. (Lidar really does matter, but also extremely rigorous testing.)
ehsankia 8y ago
You do realize that when driving, 90% of the things around you are stationary, right? Also, it may not apply to this specific case, but try comparing a plastic bag in the streets vs a big rock. Telling those two apart isn't trivial, but only one of those you can drive through.
- 6nf 8y ago
  I'd like the car to stop for plastic bags too please. Anything in the road directly in my path. I mean that plastic bag could contain a rock or something, driving through it is just a bad idea.
  - rickycook 8y ago
    what about a large leaf? it’s dangerous to stop in the middle of the road unexpectedly too
    - confiscate 8y ago
      ya but, other cars that DON'T have autopilot/self-driving tech (e.g. Subaru Eye Sight as mentioned above) already have reliable emergency braking in case of obstacles. Aren't they already able to tell the difference between a plastic bag and a big rock? If they can, then it's already a solved problem, and Tesla should be able to do it reliably too, no?
- lagadu 8y ago
  If you routinely drive through plastic bags on the street, prepare for some disappointment when it turns out that the bag had a rock inside it. The car should 100% also avoid bags or any obstacles.
LoSboccacc 8y ago
two things to consider: as speed increase, braking distance increase squared as distance increase, the angular size of an object decreases inversely. a radar requires an increase in power both to see at a longer distance and to see a smaller object, because the signature reduces with 2*distance^4 ( this is an oversimplification to help introduce the topic, you can see the details here http://www.radartutorial.eu/01.basics/The%20Radar%20Range%20Equation.en.html http://www.radartutorial.eu/01.basics/The%20Radar%20Range%20... ) a static object is the worst case scenario: you can't use its doppler signature because it's not moving against the background and the radar becomes less efficient the moment you need it the most (at high speed) as an aside, given the problem statement it should be possible to find a maximum speed that allows for timely detection and stopping. chances are it's still slow. I wonder why the radar/lidar isn't helped by an ir coincidence rangefinder.
- confiscate 8y ago
  that's a very good explanation

childintime 8y ago

The usual answer is that the signal that reflects of many stationary objects, like trees, cars, all of which have the same dopler shift and must normally be discarded. Still, this answer doesn't quite satisfy, because it ignores the very strong signal that exists right before a crash, which crash avoidance technologies do (seem to) use.

abalone 8y ago
The difference is in the compare. A collision avoidance system is there to help correct driver errors, so even if it only works 50% of the time it is still a massive reduction in the accident rate. Whereas with autopilot the expectation is that it should be recognizing objects as good or better than drivers, so a 50% failure rate would be considered massively buggy. Note: Tesla also has a collision avoidance system, which is actually on all the time and a distinct feature from autopilot (although it uses the same underlying sensors).

mentat 8y ago

Subaru Eye Sight does this reliably and actually auto brakes correctly. Going from that to APv2 has been a real step back. There's a curve on 92W that it consistently try to accelerate you into an embankment. I think Elon just doesn't understand the real world reality of how bad AP is.

Someone1234 8y ago
And Tesla's newer vehicles have the same sensors as Eye Sight, they just choose not to use them for Automatic Emergency Braking (AEB). Tesla's fans always defend the failure of their AEB by pointing out flaws in radar, while entirely ignoring that Tesla's vehicles have contained front facing optical cameras for years now. The Tesla shouldn't have hit the divider in the previous accident, the optical (visual light) camera should have seen it, even if radar didn't. Why doesn't Tesla's system work as well as Sabaru's Eye Sight?
djsumdog 8y ago
It really bothers me manufacturers are pushing incremental updates to car safety systems. Does anyone besides me think this is a terrible idea?
- Gustomaximus 8y ago
  Aren't cars doing that today with their physical saftey systems? How else would they do it than incremental?
  - InternetOfStuff 8y ago
    > How else would they do it than incremental? How about... not? I'm not being disingenious here. If there's autonomus tech, then I find it scary that it might change its behaviour. You see, I've learned how (say) Autopilot responds in certain situations. I know when to trust it, and when not to. If this behaviour changes (perhaps even if it changes for the better!), that's a very dangerous thing. Some of my experience is now invalidated, but I don't know which part. But it mostly works as before, which gives me a false sense of security. Changing the driving behaviour of a vehicle, especially in totally un-obvious ways, is super dangerous, and such changes must have very strong justifications. Just think of that poor guy whose Tesla drove into the divider. Oh look, it behaves just like before -- except for its lane following behaviour.
    - Gustomaximus 8y ago
      I dont understand how you would expect not to do incremental improvement. Build it perfectly the first time? That's not realistic. Pretty much everything humans do is incremental improvement. I understand your point that we dont want to introduce weaknesses in 'upgrades' but this should be a discussion about how to create robust QA rather that stopping incremental improvement.
      - InternetOfStuff 8y ago
        > I dont understand how you would expect not to do incremental improvement I wonder if we're talking past each other? Sure, incrementally improve during engineering. But don't incrementally change the safety-critical behaviour of a car once it's in the customer's hands. > Build it perfectly the first time? That's not realistic. Absolutely agreed. But for subjects such as autopilots, predictability may very well trump improvement. I'm assuming that the initial iteration is already reasonably safe and useful (if it isn't it shouldn't have been shipped, right?). If the behaviour changes unpredictably (how will you predict autopilot changes caused by an update?), this may well be less safe than keeping the existing system, whoose quirks the dirver has noww learned. > Pretty much everything humans do is incremental improvement. > I understand your point that we dont want to introduce weaknesses in 'upgrades' The nasty thing is that even a change to objectively superior behaviour may be problematic, because the driver will expect the car to behave one way, when it now behaves a different way (regardless of merit). Not even a changelog would help in such cases. I feel like the autopilot ought to inform the driver in each situation where it used to decide one way, and now decides another way. That might be a reasonably safe way of allowing incremental change, but is likely too annoying for most people. > but this should be a discussion about how to create robust QA rather that stopping incremental improvement. Robust QA is certainly a must. (side note: I used to work in safety-critical automotive projects. I may have a more intimate understanding of the issues)

megy 8y ago

> I don't get it, why is it so hard to detect stationary objects in front of you? This could only be said by someone who has never worked on the problem. I don't get it, why is programming so hard, you are just telling the computer exactly what to do?

maxxxxx 8y ago
It seems a legitimate question. I have never worked on this problem either but I would have thought that detecting a stationary object right in front of the car should be one of the basic and simpler functions of a self driving system. Obviously this doesn't seem to be the case but maybe you can enlighten us why this is hard?
- bobsil1 8y ago
  False positives from road signs, overhead signs, radar-reflective trash. Commonly used radars have limited vertical discrimination.
  - lagadu 8y ago
    Why would the radar even be pointed at areas several meters above the car? Maybe radars can't be pointed at a narrow space, if that's the case why not use lidar instead?
    - bobsil1 8y ago
      Most radars aren't coherent like laser, beams spread out. Also radar measures speed more accurately than location. Lidar only used by R&D + taxi services, price not reasonable for car buyers yet.
dingaling 8y ago
Instead of being condescending this would have been a good time to explain the challenges A static object has a negative Doppler shift rate linearly in proportion to vehicle speed and a y-axis position greater than zero but less than four metres. What makes that difficult to detect or distinguish?

serioussecurity 8y ago

Tesla are clowns. It's possible to do this right, but they're shipping a product they know isn't suited to the task. (Lidar really does matter, but also extremely rigorous testing.)

ehsankia 8y ago

You do realize that when driving, 90% of the things around you are stationary, right? Also, it may not apply to this specific case, but try comparing a plastic bag in the streets vs a big rock. Telling those two apart isn't trivial, but only one of those you can drive through.

6nf 8y ago
I'd like the car to stop for plastic bags too please. Anything in the road directly in my path. I mean that plastic bag could contain a rock or something, driving through it is just a bad idea.
- rickycook 8y ago
  what about a large leaf? it’s dangerous to stop in the middle of the road unexpectedly too
  - confiscate 8y ago
    ya but, other cars that DON'T have autopilot/self-driving tech (e.g. Subaru Eye Sight as mentioned above) already have reliable emergency braking in case of obstacles. Aren't they already able to tell the difference between a plastic bag and a big rock? If they can, then it's already a solved problem, and Tesla should be able to do it reliably too, no?
lagadu 8y ago
If you routinely drive through plastic bags on the street, prepare for some disappointment when it turns out that the bag had a rock inside it. The car should 100% also avoid bags or any obstacles.

LoSboccacc 8y ago

two things to consider: as speed increase, braking distance increase squared as distance increase, the angular size of an object decreases inversely. a radar requires an increase in power both to see at a longer distance and to see a smaller object, because the signature reduces with 2*distance^4 ( this is an oversimplification to help introduce the topic, you can see the details here http://www.radartutorial.eu/01.basics/The%20Radar%20Range%20Equation.en.html http://www.radartutorial.eu/01.basics/The%20Radar%20Range%20... ) a static object is the worst case scenario: you can't use its doppler signature because it's not moving against the background and the radar becomes less efficient the moment you need it the most (at high speed) as an aside, given the problem statement it should be possible to find a maximum speed that allows for timely detection and stopping. chances are it's still slow. I wonder why the radar/lidar isn't helped by an ir coincidence rangefinder.

confiscate 8y ago
that's a very good explanation