5 ms·

self driving systems cant well reason about untrained scenarios or the intent of other humans on the road. I think the people have grossly underestimated how d

self driving systems cant well reason about untrained scenarios or the intent of other humans on the road. I think the people have grossly underestimated how driving in an uncontrolled environment is really a general AI problem, which we're not even close to solving.

CydeWeys 8y ago
Disagree. I think that the likes of Waymo and Cruise understand exactly how hard of a problem this is. It's only companies like Uber and Tesla that are underestimating it, and doing stupid things like disabling automatic braking (Uber) or not using LIDAR (Tesla).
- jakelarkin 8y ago
  the best software engineering company in the world has spent over a decade and hundreds of $millions on this problem and they are still stuck in a private beta in the easiest road environment they could find. what does that tell you about the feasibility of the solution. There's no evidence that Cruise is in the same class as Waymo.
  - CydeWeys 8y ago
    Waymo is about to start ferrying around paying customers later this year. Nobody said the problem isn't hard, but it definitely does not seem infeasible.
decacorn1 8y ago
Decades away, at which point OEMs will have spent billions on a technology unproven and undesired. They will have bankrupted themselves because there is no viable business model other than SaaS and car companies will not transition to SaaS. They can’t even bring themselves to transition to full-electric platforms.
Animats 8y ago
This wasn't a hard case that needed advanced AI. This was a clearly visible stationary obstacle on a good freeway on a clear California day.
- jakelarkin 8y ago
  deciding how/when to make panic stop or evasive maneuvers without routinely becoming a hazard requires advanced AI determining the path of travel through lanes markings that suddenly fade out while moving in 70mph on a busy freeway requires advanced AI its pretty clear the self-driving gods are not giving these cars a grounds up understanding of physics and self-preservation. They best thing they come up with is programmed to drive X mph and to respond probabilistically to a large but finite set of training input scenarios.
  - dragontamer 8y ago
    Or they require Lidar. You know, a sensor that actually detects stationary objects at a reliable rate. Tesla is relying upon advanced AI to do things that its sensors can't do reliably. Its only Tesla who is trying to make self-driving cars off of cameras and inaccurate radar alone. Maybe Lidar really is too expensive for mass-market adoption. But its the most obvious solution to the problem right now, instead of trying to solve harder problems (ie: camera recognition of 3d objects over trained neural nets).
pipio21 8y ago
You don't need reason, you need fast reaction times slowing the car and avoiding obstacles. The same happens with humans. Reason is extremely slow. Pilots are trained to act fast training the subconscious, not the logical mind.
- jakelarkin 8y ago
  normal driving requires anticipating the behavior of other humans who may or may not entire your path of travel or anticipating unseen obstacles. it has less to do with "reaction times" than one would think. Humans routinely over-drive their vehicles capability to safely stop for an obstacle in front or adjacent to their path of travel.
  - cornholio 8y ago
    > Humans routinely over-drive their vehicles capability to safely stop for an obstacle in front or adjacent to their path of travel. Hence, the opportunity for an automated system that does not do that to be much safer, by relying on reaction times rather than strong AI. Where a human driver would use subtle cues to anticipate a slowdown of the preceding vehicle before a turn, an automated system can get by by simply slamming the brakes in less than one millisecond when it detects braking from the other vehicle.
    - jakelarkin 8y ago
      ... and that self-driving car will make its passengers car sick and get rear ended by the human driver behind it. those types of twitchy driving mechanics aren't normal and don't share the road well with normal humans. The physics of cars and reaction times would dictate that we program a self-driving car to drive like Grandma .... always maintain safe low speeds, very long following distances and braking hard for sketchy actions by other cars or random things near the road, but we know that actually makes the road more dangerous as humans drivers will aggressively cut-off and rear end this self-driving grandma. Secondly, that type of driving creates a bad public impression of self-driving cars hurting their chances of adoption. If you read some of the earlier impressions of Google/Waymo cars its clear they went down that path initially and had to change their approach.
cornholio 8y ago
> driving in an uncontrolled environment is really a general AI problem Is it? Unless you are talking about NASCAR, an adversarial game where rational agents compete for road space, driving is essentially a physics problem: maintain a speed and direction that does not exceed the vehicle's ability to correct them so as to avoid any stationary or mobile obstacle. In this case, the vehicle was accelerated to a speed (72mph) that far exceeded its safety speed against oncoming stationary barriers (0mph). I concede that discerning what is an obstacle and what is drivable road surface is a hard computer vision problem, hence the tendency to simplify it using 3D data, radar/lidar etc. But still, not quite general AI.
- jakelarkin 8y ago
  you can break it down to various expert-system subproblems - perfect cv image classification (+ low visibility ) - anticipating behavior of bikers, children, elderly, animals, near roads - responding to human non-verbal communication - responding appropriately to never before seen obstacles/scenarios without being a nuisance on the road or endangering other drivers - vehicle dynamics (+ in rain, ice, snow, emergency etc) ... I would guess to do these tasks with accuracy similar to humans approaches a problem space that's as difficult as general AI some of these seemingly require the machine to have theory-of-mind which is general AI
  - bobsil1 8y ago
    Yep. Broad world model, physics, common sense, human communication, prediction, reasoning. Waymo’s low-level stack on hi-res map rails can’t replace human drivers.

CydeWeys 8y ago

Disagree. I think that the likes of Waymo and Cruise understand exactly how hard of a problem this is. It's only companies like Uber and Tesla that are underestimating it, and doing stupid things like disabling automatic braking (Uber) or not using LIDAR (Tesla).

jakelarkin 8y ago
the best software engineering company in the world has spent over a decade and hundreds of $millions on this problem and they are still stuck in a private beta in the easiest road environment they could find. what does that tell you about the feasibility of the solution. There's no evidence that Cruise is in the same class as Waymo.
- CydeWeys 8y ago
  Waymo is about to start ferrying around paying customers later this year. Nobody said the problem isn't hard, but it definitely does not seem infeasible.

decacorn1 8y ago

Decades away, at which point OEMs will have spent billions on a technology unproven and undesired. They will have bankrupted themselves because there is no viable business model other than SaaS and car companies will not transition to SaaS. They can’t even bring themselves to transition to full-electric platforms.

Animats 8y ago

This wasn't a hard case that needed advanced AI. This was a clearly visible stationary obstacle on a good freeway on a clear California day.

jakelarkin 8y ago
deciding how/when to make panic stop or evasive maneuvers without routinely becoming a hazard requires advanced AI determining the path of travel through lanes markings that suddenly fade out while moving in 70mph on a busy freeway requires advanced AI its pretty clear the self-driving gods are not giving these cars a grounds up understanding of physics and self-preservation. They best thing they come up with is programmed to drive X mph and to respond probabilistically to a large but finite set of training input scenarios.
- dragontamer 8y ago
  Or they require Lidar. You know, a sensor that actually detects stationary objects at a reliable rate. Tesla is relying upon advanced AI to do things that its sensors can't do reliably. Its only Tesla who is trying to make self-driving cars off of cameras and inaccurate radar alone. Maybe Lidar really is too expensive for mass-market adoption. But its the most obvious solution to the problem right now, instead of trying to solve harder problems (ie: camera recognition of 3d objects over trained neural nets).

pipio21 8y ago

You don't need reason, you need fast reaction times slowing the car and avoiding obstacles. The same happens with humans. Reason is extremely slow. Pilots are trained to act fast training the subconscious, not the logical mind.

jakelarkin 8y ago
normal driving requires anticipating the behavior of other humans who may or may not entire your path of travel or anticipating unseen obstacles. it has less to do with "reaction times" than one would think. Humans routinely over-drive their vehicles capability to safely stop for an obstacle in front or adjacent to their path of travel.
- cornholio 8y ago
  > Humans routinely over-drive their vehicles capability to safely stop for an obstacle in front or adjacent to their path of travel. Hence, the opportunity for an automated system that does not do that to be much safer, by relying on reaction times rather than strong AI. Where a human driver would use subtle cues to anticipate a slowdown of the preceding vehicle before a turn, an automated system can get by by simply slamming the brakes in less than one millisecond when it detects braking from the other vehicle.
  - jakelarkin 8y ago
    ... and that self-driving car will make its passengers car sick and get rear ended by the human driver behind it. those types of twitchy driving mechanics aren't normal and don't share the road well with normal humans. The physics of cars and reaction times would dictate that we program a self-driving car to drive like Grandma .... always maintain safe low speeds, very long following distances and braking hard for sketchy actions by other cars or random things near the road, but we know that actually makes the road more dangerous as humans drivers will aggressively cut-off and rear end this self-driving grandma. Secondly, that type of driving creates a bad public impression of self-driving cars hurting their chances of adoption. If you read some of the earlier impressions of Google/Waymo cars its clear they went down that path initially and had to change their approach.

cornholio 8y ago

> driving in an uncontrolled environment is really a general AI problem Is it? Unless you are talking about NASCAR, an adversarial game where rational agents compete for road space, driving is essentially a physics problem: maintain a speed and direction that does not exceed the vehicle's ability to correct them so as to avoid any stationary or mobile obstacle. In this case, the vehicle was accelerated to a speed (72mph) that far exceeded its safety speed against oncoming stationary barriers (0mph). I concede that discerning what is an obstacle and what is drivable road surface is a hard computer vision problem, hence the tendency to simplify it using 3D data, radar/lidar etc. But still, not quite general AI.

jakelarkin 8y ago
you can break it down to various expert-system subproblems - perfect cv image classification (+ low visibility ) - anticipating behavior of bikers, children, elderly, animals, near roads - responding to human non-verbal communication - responding appropriately to never before seen obstacles/scenarios without being a nuisance on the road or endangering other drivers - vehicle dynamics (+ in rain, ice, snow, emergency etc) ... I would guess to do these tasks with accuracy similar to humans approaches a problem space that's as difficult as general AI some of these seemingly require the machine to have theory-of-mind which is general AI
- bobsil1 8y ago
  Yep. Broad world model, physics, common sense, human communication, prediction, reasoning. Waymo’s low-level stack on hi-res map rails can’t replace human drivers.