3 ms·

How would it show up different than a regular pot hole? What if its almost filled with water and the only thing giving away its true nature is its roundness? Th

How would it show up different than a regular pot hole? What if its almost filled with water and the only thing giving away its true nature is its roundness? The point is to explore the limits. If these aren't the limits, then they are likely very close. I'd like people to think of those cases before they jump into their car and start playing cards (like i one of the videos in this thread).

XorNot 10y ago
Basic geometry. A pothole of greater then a certain depth, at a known angle from sensor will produce a different distance profile (namely - rather then an expected arc, you'd see a sudden increase in distance, corresponding to a non-planar surface). Depth is inferred from the deviation from the plane of the road-surface around it. This is not actually a particularly interesting edge case. The water-case is more interesting, but I would counter by posing the question of how a human would know whether a water-filled pot hole is safe to navigate? A clear water pothole would not impinge the LIDAR, a murky water one might and could fool LIDAR I suppose (I can not find any solid info on this interface). Of course car RADAR and a visual camera are likely to also be fitted - both of which could identify a water-filled pothole which LIDAR might struggle with (RADAR can penetrate water and find the unknown backscatter, a camera can simply see the murky-water pothole and choose to navigate around it or stop - in both cases a human can't make a better informed decision).
- felippee 10y ago
  It depends on the context. Humans may make a better decision if there is something else providing the information. Perhaps the manhole cover sits at the curb triggering immediate attention. Perhaps there are cones, only in the wrong place because some drunk guy moved them for fun. Such is reality. The problem is not so much availability of information (a sensor packed car has a lot more information than a human behind a windshield), but making sense of that information, particularly in those 0.0001% uncommon, strange cases. Humans can connect evidence (a manhole cover sitting next to a water filled hole) because they generally know how things work (aka common sense knowledge, long standing problem in AI). Humans can infer what happened and predict what could happen. These things are not really available to AI which is these days more like see-react, not see-anticipate-react. I have a specific post on this on my blog:http://blog.piekniewski.info/2016/11/03/reactive-vs-predictive-ai/ http://blog.piekniewski.info/2016/11/03/reactive-vs-predicti... The blog has many other posts on limits of todays approach and on some ideas to fix it in the future. I agree that some of these cases may be challenging to some humans (particularly inattentive), but we want the self driving car to be (much) safer than inattentive driver, so we need to set the bar high.
- emodendroket 10y ago
  Clearly if you consider any of these scenarios in isolation we can dream up a solution, but can you do that with literally any adverse scenario? That seems like the point to me.
  - felippee 10y ago
    Yes, that is exactly my point. Solving these particular situations is like typing all the possible Winograd Schema sentences into a chat bot. Certainly possible but does not solve a larger problem. In reality there will always be a different situation. These situations sit in the statistical long tail (are too infrequent to reliably train, and too frequent to ignore) and vary enormously from case to case. These are just examples.
    - XorNot 10y ago
      Except if the first scenario is solvable (fairly trivially) then why should I assume there must exist an unsolvable scenario? An argument from the vaguery of "humans have context" with regards to driving is a poor one - humans that are driving do not do a good job of analysing context because of reaction times, and do not share a common context they react to similarly. You haven't made a compelling case, because you've yet to present a compelling example (i.e. one definitely unsolvable by reasonably usable technology on an autonomous platform). The manhole scenario is based on assumption about the operation of LIDAR which simply aren't true and you've had to modify it in a number of ways to try and make it tricky (i.e. when has there ever been a manhole completely full of water?)
      - emodendroket 10y ago
        I think you're still confused. The point isn't that any of these scenarios don't have solutions. It's that there is a practically infinite number of scenarios like them and it's impossible to plan for each and every one of them.

XorNot 10y ago

Basic geometry. A pothole of greater then a certain depth, at a known angle from sensor will produce a different distance profile (namely - rather then an expected arc, you'd see a sudden increase in distance, corresponding to a non-planar surface). Depth is inferred from the deviation from the plane of the road-surface around it. This is not actually a particularly interesting edge case. The water-case is more interesting, but I would counter by posing the question of how a human would know whether a water-filled pot hole is safe to navigate? A clear water pothole would not impinge the LIDAR, a murky water one might and could fool LIDAR I suppose (I can not find any solid info on this interface). Of course car RADAR and a visual camera are likely to also be fitted - both of which could identify a water-filled pothole which LIDAR might struggle with (RADAR can penetrate water and find the unknown backscatter, a camera can simply see the murky-water pothole and choose to navigate around it or stop - in both cases a human can't make a better informed decision).

felippee 10y ago
It depends on the context. Humans may make a better decision if there is something else providing the information. Perhaps the manhole cover sits at the curb triggering immediate attention. Perhaps there are cones, only in the wrong place because some drunk guy moved them for fun. Such is reality. The problem is not so much availability of information (a sensor packed car has a lot more information than a human behind a windshield), but making sense of that information, particularly in those 0.0001% uncommon, strange cases. Humans can connect evidence (a manhole cover sitting next to a water filled hole) because they generally know how things work (aka common sense knowledge, long standing problem in AI). Humans can infer what happened and predict what could happen. These things are not really available to AI which is these days more like see-react, not see-anticipate-react. I have a specific post on this on my blog:http://blog.piekniewski.info/2016/11/03/reactive-vs-predictive-ai/ http://blog.piekniewski.info/2016/11/03/reactive-vs-predicti... The blog has many other posts on limits of todays approach and on some ideas to fix it in the future. I agree that some of these cases may be challenging to some humans (particularly inattentive), but we want the self driving car to be (much) safer than inattentive driver, so we need to set the bar high.
emodendroket 10y ago
Clearly if you consider any of these scenarios in isolation we can dream up a solution, but can you do that with literally any adverse scenario? That seems like the point to me.
- felippee 10y ago
  Yes, that is exactly my point. Solving these particular situations is like typing all the possible Winograd Schema sentences into a chat bot. Certainly possible but does not solve a larger problem. In reality there will always be a different situation. These situations sit in the statistical long tail (are too infrequent to reliably train, and too frequent to ignore) and vary enormously from case to case. These are just examples.
  - XorNot 10y ago
    Except if the first scenario is solvable (fairly trivially) then why should I assume there must exist an unsolvable scenario? An argument from the vaguery of "humans have context" with regards to driving is a poor one - humans that are driving do not do a good job of analysing context because of reaction times, and do not share a common context they react to similarly. You haven't made a compelling case, because you've yet to present a compelling example (i.e. one definitely unsolvable by reasonably usable technology on an autonomous platform). The manhole scenario is based on assumption about the operation of LIDAR which simply aren't true and you've had to modify it in a number of ways to try and make it tricky (i.e. when has there ever been a manhole completely full of water?)
    - emodendroket 10y ago
      I think you're still confused. The point isn't that any of these scenarios don't have solutions. It's that there is a practically infinite number of scenarios like them and it's impossible to plan for each and every one of them.